Intelligent processing method and system for structured data to text based on natural language

Abstract

The application provides a natural language-based intelligent processing method and system for structured data to text, and relates to the technical field of data processing.The method comprises the following steps: optimizing and adjusting a text chunk analysis process according to analysis optimization parameters, segmenting a sentence into non-overlapping phrases with syntactic role labels to obtain text after chunk analysis; performing syntactic and semantic structure analysis on the text after chunk analysis, establishing semantic correlation between internal structures of the sentence through component analysis, dependency analysis and semantic dependency graph analysis, and obtaining structured semantic information; performing multi-sentence logical correlation analysis on the structured semantic information at the chapter level to obtain semantic information of the overall chapter; and obtaining a target natural language text based on the semantic information of the overall chapter by using a pre-trained large language model.The application improves the accuracy and fluency of the conversion of structured data to text.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an intelligent processing method and system for structured data to text based on natural language. Background Technology

[0002] In industrial equipment operation and maintenance management, it is often necessary to automatically convert the structured data of equipment operation into fluent and coherent status report text, such as generating analysis summaries that clearly reflect the health status of equipment operation for the equipment management system. In the existing technology, the method based on pre-trained large language model has become the mainstream. This type of method usually inputs the structured data of equipment operation (such as tabular data such as temperature of different components, running time, energy consumption value, number of fault warnings, etc.) directly or after simple serialization into the model, and the model generates equipment status report text end-to-end.

[0003] However, such methods may have certain limitations in practical applications. The main problem is that, when faced with discrete, structured data in equipment operation, large language models may sometimes struggle to fully understand the deep and complex semantic and logical relationships between these data elements. For example, the relationship between excessively high temperatures of heat dissipation components and the resulting increase in energy consumption and fault warnings. Models often focus more on the surface information of the data itself, such as emphasizing specific values ​​of temperature and energy consumption, and find it difficult to systematically build and maintain the semantic structure from data phrases to analysis sentences, and then to the complete report chapters in the generated status report. This may lead to flaws in the logical coherence, information accuracy, and semantic integrity of the generated text, thereby affecting the overall quality and reliability of the generated text to some extent. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide an intelligent processing method and system for structured data to text based on natural language, which improves the accuracy and fluency of structured data to text conversion.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] Firstly, an intelligent method for processing structured data into text based on natural language, the method comprising:

[0007] Step 1: Build a neural network model and train it; use the trained neural network model to perform lexical analysis and feature extraction on the natural language text to be processed, and obtain a set of lexical feature vectors.

[0008] Step 2: Construct a multidimensional feature space based on the set of lexical feature vectors; divide the multidimensional feature space into feature subspaces, analyze the statistical characteristics of the feature subspaces, and generate analysis and optimization parameters;

[0009] Step 3: Optimize and adjust the text chunking analysis process according to the analysis and optimization parameters, and segment the sentence into non-overlapping phrases with syntactic function tags to obtain the chunked text;

[0010] Step 4: Perform syntactic and semantic structure parsing on the text after chunking analysis. Through component analysis, dependency analysis and semantic dependency graph analysis, establish semantic relationships between internal sentence structures to obtain structured semantic information.

[0011] Step 5: Perform multi-sentence logical association analysis on the structured semantic information at the discourse level to obtain the semantic information of the entire discourse;

[0012] Step 6: Based on the semantic information of the entire text, obtain the target natural language text using a pre-trained large language model.

[0013] Secondly, intelligent processing systems for structured data to text based on natural language include:

[0014] The extraction module is used to build and train a neural network model; the trained neural network model is then used to perform lexical analysis and feature extraction on the natural language text to be processed, resulting in a set of lexical feature vectors.

[0015] The partitioning module is used to construct a multidimensional feature space based on the set of lexical feature vectors; based on the multidimensional feature space, it divides the feature subspaces, analyzes the statistical characteristics of the feature subspaces, and generates analysis and optimization parameters;

[0016] The optimization module is used to optimize and adjust the text chunking analysis process according to the analysis optimization parameters, segmenting sentences into non-overlapping phrases with syntactic function tags to obtain the chunked text;

[0017] The parsing module is used to perform syntactic and semantic structure parsing on the text after chunking analysis. Through component analysis, dependency analysis and semantic dependency graph analysis, it establishes semantic relationships between the internal structures of sentences and obtains structured semantic information.

[0018] The analysis module is used to perform multi-sentence logical association analysis on structured semantic information at the discourse level to obtain the semantic information of the entire discourse.

[0019] The generation module is used to obtain the target natural language text based on the semantic information of the entire text and using a pre-trained large language model.

[0020] Thirdly, a computing device includes:

[0021] One or more processors;

[0022] A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0023] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0024] The above-described solution of the present invention has at least the following beneficial effects:

[0025] By training a neural network model to perform lexical analysis and feature extraction, the model's ability to learn the semantic rules of natural language can be leveraged to accurately capture the deep semantic features and contextual information of words. The generated set of lexical feature vectors can effectively avoid the limitations of manual feature design and reduce subsequent analysis biases caused by insufficient feature representation. Based on the lexical feature vectors, a multi-dimensional feature space is constructed and divided into subspaces. The statistical characteristics of the subspaces are combined to generate analysis optimization parameters. By optimizing the parameters, the processing focus and rules are clarified, which can reduce ineffective analysis processes, improve the targeting of text processing, reduce redundant calculations, and thus improve overall processing efficiency. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the intelligent processing method for structured data to text based on natural language provided in an embodiment of the present invention.

[0027] Figure 2 This is a schematic diagram of an intelligent data-to-text system based on natural language structure provided in an embodiment of the present invention. Detailed Implementation

[0028] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0029] like Figure 1 As shown, embodiments of the present invention propose an intelligent processing method for structured data to text based on natural language, the method comprising the following steps:

[0030] Step 1: Build a neural network model and train it; use the trained neural network model to perform lexical analysis and feature extraction on the natural language text to be processed, and obtain a set of lexical feature vectors.

[0031] Step 2: Construct a multidimensional feature space based on the set of lexical feature vectors; divide the multidimensional feature space into feature subspaces, analyze the statistical characteristics of the feature subspaces, and generate analysis and optimization parameters;

[0032] Step 3: Optimize and adjust the text chunking analysis process according to the analysis and optimization parameters, and segment the sentence into non-overlapping phrases with syntactic function tags to obtain the chunked text;

[0033] Step 4: Perform syntactic and semantic structure parsing on the text after chunking analysis. Through component analysis, dependency analysis and semantic dependency graph analysis, establish semantic relationships between internal sentence structures to obtain structured semantic information.

[0034] Step 5: Perform multi-sentence logical association analysis on the structured semantic information at the discourse level to obtain the semantic information of the entire discourse;

[0035] Step 6: Based on the semantic information of the entire text, obtain the target natural language text using a pre-trained large language model.

[0036] In this embodiment of the invention, lexical analysis and feature extraction are achieved by training a neural network model. This model can learn the semantic rules of natural language, accurately capture the deep semantic features and contextual information of words, and generate a set of word feature vectors. This effectively avoids the limitations of manual feature design and reduces subsequent analysis biases caused by insufficient feature representation. Based on the word feature vectors, a multi-dimensional feature space is constructed and divided into subspaces. Analysis optimization parameters are generated by combining the statistical characteristics of the subspaces. By optimizing the parameters, the processing focus and rules are clarified, which can reduce invalid analysis processes, improve the targeting of text processing, reduce redundant calculations, and thus improve the overall processing efficiency.

[0037] In a preferred embodiment of the present invention, step 1 involves constructing and training a neural network model; using the trained neural network model to perform lexical analysis and feature extraction on the natural language text to be processed, obtaining a set of lexical feature vectors, which may include:

[0038] Step 101: An initial neural network model is obtained by constructing a neural network model containing an input layer, multiple hidden layers, and an output layer. Natural language text samples and lexical annotation information are input into the initial neural network model, and the model parameters are optimized using the backpropagation algorithm to obtain a trained neural network model. Specifically, the core function of the input layer is to receive text sequences. The number of its nodes must be consistent with the preset maximum text length, which is set to 512 characters based on the common length of industrial equipment status reports to ensure complete coverage of the text content of a single equipment report. Each node corresponds to a character position in the text and is used to receive the initial vector representation of the word at that position. The initial vector is obtained based on pre-trained word vectors specific to the industrial equipment field. The training corpus of these word vectors covers equipment operation manuals, fault repair records, and monthly status reports from the past three years, with a total corpus size of 1 million entries. This ensures that the initial vectors accurately reflect the specific semantics of industrial terms such as motors, cooling fans, temperature, and vibration frequency, avoiding confusion with general domain semantics.

[0039] The hidden layer consists of three functionally defined sublayers. The first sublayer employs a bidirectional long short-term memory network, which can simultaneously capture the contextual relationships between words in the text. For example, when analyzing a warning triggered by excessive temperature, it is necessary to connect the temperature to the causal relationship of the preceding cooling fan malfunction, as well as the excessive temperature to the subsequent warning. Therefore, this layer has 512 neurons to ensure sufficient storage of contextual semantic information. The second sublayer is a convolutional neural network layer, which extracts local phrase features from the text. Considering that industrial equipment texts often contain three-word phrases combining components and parameters, such as cooling fan speed and bearing vibration frequency, 128 convolutional kernels of size 3 are used. Each convolutional kernel focuses on capturing one type of phrase feature, improving the targeting of local semantic extraction. The third sublayer is a fully connected layer, which uses the ReLU activation function to enhance the non-linear expressive power of features. With 256 neurons, it integrates the contextual features output by the bidirectional long short-term memory network with the local phrase features output by the convolutional neural network layer to form a more comprehensive intermediate feature vector.

[0040] The output layer is designed as a fully connected layer, and the number of its neurons must perfectly match the number of lexical annotation categories. The lexical annotation categories are specifically divided into five main categories: equipment component categories (including 20 common industrial components such as motors, gearboxes, bearings, and lubricating oil pipelines), parameter categories (including 15 core monitoring parameters such as temperature, energy consumption, vibration frequency, and pressure), numerical categories (including monitoring values ​​with units, such as 85℃, 280kWh, and 0.3MPa), state description categories (including 8 parameter states such as normal, abnormal, high, and low), and action categories (including 10 equipment operation actions such as raising, lowering, triggering, and stopping), totaling 63 subcategories. Therefore, the number of neurons in the output layer is set to 63. The output layer uses a softmax function to convert the intermediate feature vector output from the hidden layer into the probability of each word belonging to each category. The probability value ranges from 0 to 1, and the sum of the probabilities of all categories is 1, which facilitates subsequent determination of the word's annotation category.

[0041] After completing the model structure construction, the model training phase begins. Natural language text samples are prepared, extracted from the historical database of the enterprise's industrial equipment operation and maintenance system. A total of 100,000 valid samples are selected, covering 10 equipment types including motors, gearboxes, cooling systems, and hydraulic devices, and including four operating scenarios: normal operation, minor faults, moderate faults, and emergency shutdown, to ensure the diversity and representativeness of the samples. At the same time, corresponding lexical annotation information is prepared. The annotation work is completed by 5 experts with more than 5 years of industrial operation and maintenance experience. The experts annotate the category of each word in each sample according to a unified annotation manual. The annotation process is carried out in 5 batches. After each batch is annotated, experts cross-validate to ensure that the annotation consistency rate reaches more than 95% and to ensure the accuracy of the annotation information.

[0042] The text samples are converted into a sequence format that the input layer can accept. Specifically, each word is replaced with a corresponding domain-specific pre-trained word vector to form a vector sequence. This vector sequence, along with the corresponding lexical annotation information (represented in one-hot encoding form, such as the encoding of a motor in the equipment component category as a vector with the first bit set to 1 and the rest to 0), is input into the initial neural network model. After the model outputs the predicted probability of each word belonging to one of 63 categories, the loss value between the predicted result and the true annotation is calculated using the cross-entropy loss function. The calculation process is as follows: for each word in the sample, the predicted probability corresponding to the true annotation category of the word is first found, the natural logarithm of the predicted probability is taken, and then the logarithm result is multiplied by -1. Then, the calculated results of all words in the sample are summed to obtain the total loss value of the entire sample, which reflects the degree of deviation between the model prediction and the actual situation.

[0043] The model parameters are optimized using the backpropagation algorithm, following a fixed process. Starting from the output layer, the partial derivatives of the total loss value with respect to all parameters in the model (including weights from the input layer to the hidden layer, between sub-layers of the hidden layer, from the hidden layer to the output layer, and the biases of each layer) are calculated. The magnitude of these partial derivatives directly reflects the degree to which changes in the corresponding parameter affect the total loss value; the larger the absolute value of the partial derivative, the more significant the impact of that parameter on the model's prediction results. The parameters are adjusted according to the direction of the partial derivatives, with the adjustment range controlled by the learning rate. The initial learning rate is set to 0.001, and new parameters... The numerical calculation method is to subtract the learning rate multiplied by the partial derivative of the parameter from the original parameter value. If the partial derivative is positive, it means that the parameter needs to be reduced to reduce the loss; if the partial derivative is negative, it means that the parameter needs to be increased to reduce the loss. Model training is carried out in batches, with 32 samples input in each batch. After one parameter adjustment, the next batch of training begins. During this period, the total loss value of each batch is continuously monitored. If the total loss value of 20 consecutive batches decreases by less than 0.0001, and the loss value on the test set no longer decreases, it means that the model has converged. At this time, training is stopped, and the trained neural network model is obtained.

[0044] Step 102: Receive structured data from external sources and convert it into natural language text to obtain the natural language text to be processed. Input the natural language text to be processed into the input layer of the trained neural network model to obtain the original word sequence. Specifically, this includes receiving external structured data, which mainly comes from the enterprise's industrial control system and IoT monitoring platform. The industrial control system pushes real-time equipment operation data every 5 minutes in a table format, containing seven columns: equipment number, component name, parameter name, monitoring value, unit, collection time, and status identifier. The IoT monitoring platform generates the previous day's equipment data at 2:00 AM every day. The system generates operational statistics data in JSON format. The JSON structure includes three primary fields: device information, operational parameter statistics, and warning records. Device information includes subfields for device number, device name, and installation location. Operational parameter statistics includes subfields for component, parameter, daily average, maximum, and minimum values. Warning records include subfields for warning time, warning type, and number of warnings. Data is received through a dedicated data interface using HTTPS to ensure secure data transmission. The interface also performs integrity checks on the received data. If any column or field is missing in the table data or JSON data, a retransmission request is immediately sent to the data source until complete data is obtained.

[0045] The received structured data is converted into natural language text. The conversion process strictly follows the expression conventions of industrial equipment status reports, combining data elements in a fixed logical order of component + parameter + value + status. This ensures consistent expression and conforms to the reading habits of maintenance personnel. For example, a record in the table with equipment number M001, component name "bearing," parameter name "vibration frequency," monitoring value "65Hz," unit "Hz," acquisition time "2025.09.30 10:00," and status "abnormal" is converted to: equipment M001 bearing vibration frequency 65Hz, acquisition time "September 30, 2025 10:00," status "abnormal." The equipment number is M002, the equipment name is cooling fan, the operating parameter statistics include the component is cooling fan, the parameter is speed, the daily average value is 1200 r / min, and the warning record includes the warning time is 2025.09.30 08:00, the warning type is over-temperature warning, and the number of warnings is 2. This is converted to the information that the daily average speed of cooling fan M002 is 1200 r / min and that there were 2 over-temperature warnings on September 30, 2025 at 8:00. During the conversion process, the data is checked field by field to ensure that no key elements such as equipment number, component name, parameter value, and status identifier are omitted, while avoiding redundant descriptions, so as to form a coherent and complete natural language text to be processed.

[0046] The natural language text to be processed is input into the input layer of the trained neural network model. The input layer performs word segmentation on the text. The word segmentation logic is based on an industry-specific dictionary, which includes 2,000 new industry-specific terms such as motor temperature and current overload warning, to ensure the accuracy of word segmentation. Two core rules are followed during word segmentation: First, compound terms containing components and parameters are split into component words and parameter words, such as gearbox temperature being split into gearbox and temperature, and bearing vibration frequency being split into bearing and vibration frequency; second, all combinations containing numerical values ​​and units are retained as individual words without splitting, such as 85℃, 280kWh, and 0.3MPa being retained as independent words. After word segmentation, the segmented words are arranged in the original order of the text to form an ordered sequence of original words.

[0047] Step 103: The original word sequence is processed by lexical analysis through the hidden layer of the neural network model to obtain the lexical analysis result; the lexical analysis result is then processed by feature extraction through the output layer of the neural network model to obtain a set of lexical feature vectors. Specifically, the original word sequence is input into the hidden layer of the trained neural network model. Each word in the original word sequence enters the first sub-layer of the hidden layer, namely the bidirectional long short-term memory network, in sequence. The bidirectional long short-term memory network captures the contextual semantics of each word. During processing, it starts from the position of the word in the sequence, associating the semantics of all words before the word forward and the semantics of all words after the word backward. For example, for the vibration frequency in the original word sequence, it associates the semantics of bearing forward and the semantics of 65Hz backward. Through the collaborative computation of 512 neurons, this association information is transformed into the contextual feature vector of the word. Each vector contains 512 values, corresponding to the output results of the 512 neurons.

[0048] The convolutional neural network layer uses 128 convolutional kernels of size 3. Each kernel slides sequentially across the word sequence, covering the context feature vectors of three consecutive words at a time. For each sliding window, the kernel calculates the sum of the products of the three vectors within the window and the kernel weights, plus the kernel bias, to obtain a local feature value. This is used to extract local features associated with combinations of three words such as "part," "parameter," and "value." For example, when processing the three consecutive words "bearing," "vibration frequency," and "65Hz," the kernel calculates the sum of the products of the context feature vectors of these three words and their corresponding weights, plus the bias, to obtain a local feature reflecting the overall semantics of this combination. The 128 convolutional kernels each output 128 local features, which together constitute the local feature vector of the window. The local feature vector and its corresponding upper and lower bounds... The textual feature vectors are fed into the third sub-layer of the hidden layer, namely the fully connected layer. The 256 neurons of the fully connected layer are connected to each value in the input feature vector. Each neuron calculates the sum of the products of all values ​​in the input feature vector and their corresponding weights, adds the neuron's bias value, and then processes it through the ReLU activation function (i.e., if the calculation result is negative, the output is 0; if it is positive, the original result is retained) to obtain the neuron's output value. The output values ​​of the 256 neurons together constitute the intermediate feature vector. This vector integrates contextual association and local phrase features to complete the lexical analysis of each word, clarifying the category attribute of each word in the industrial scenario. For example, bearings belong to the equipment component category, vibration frequency belongs to the parameter category, and 65Hz belongs to the numerical category, forming the lexical analysis result.

[0049] The intermediate feature vector output from the hidden layer is input into the output layer of the neural network model. Each of the 63 neurons in the output layer is connected to one of the 256 values ​​in the intermediate feature vector. Each neuron calculates the sum of the products of all values ​​in the intermediate feature vector and their corresponding weights, and adds this sum to the neuron's bias value to obtain 63 raw output values. These 63 raw output values ​​are then processed using a softmax function. Specifically, each raw output value is used as the result of the natural exponential function calculation, and each result is divided by the sum of the natural exponential function calculations of all raw output values ​​to obtain the probability that each word belongs to one of the 63 categories. The category with the highest probability value is the final lexical annotation result for that word. During the generation of probability values, the output layer retains the correlation information between the intermediate feature vector and the output layer weights and biases, and fuses the intermediate feature vector with the probability value of the corresponding category to form the lexical feature vector of each word. The fusion method is to concatenate the 256 values ​​of the intermediate feature vector with the probability value of the category to which the word belongs in order to form a vector containing 257 values, thereby preserving the semantic features, contextual associations and category attributes of the word. The lexical feature vectors of all words are arranged in the order of the original word sequence to obtain the lexical feature vector set.

[0050] This embodiment, by constructing a neural network model and training it with domain samples, enables the model to accurately identify the keyword categories in device-related text, thereby improving the targeting and accuracy of lexical analysis.

[0051] In a preferred embodiment of the present invention, step 2, constructing a multidimensional feature space based on the set of lexical feature vectors; dividing the multidimensional feature space into feature subspaces, analyzing the statistical characteristics of the feature subspaces, and generating analysis and optimization parameters, may include:

[0052] Step 201: Integrate all word feature vectors in the word feature vector set into a feature matrix to obtain a multidimensional feature space. A clustering algorithm is then used in the multidimensional feature space to group semantically similar word feature vectors into the same region, resulting in multiple feature subspaces. Specifically, each word feature vector obtained in step 103 is 257-dimensional (256 intermediate features + 1 category probability). Based on the order of the original word sequence, all word feature vectors are arranged sequentially in row vector form to form a feature matrix. The number of rows in the feature matrix equals the total number of word feature vectors, and the number of columns equals the vector dimension (257 columns). The vector space corresponding to this feature matrix is ​​the multidimensional feature space. For example, if the original word sequence contains 10 words (M001, equipment, bearing, of, vibration frequency, for, 65Hz, state, is, abnormal), then the feature matrix is ​​10 rows and 257 columns, with each row corresponding to a 257-dimensional word feature vector, fully preserving the semantic and category information of words in the industrial equipment text.

[0053] K-means clustering algorithm is used to divide the feature subspace, and the number of clusters K is determined. Combined with the lexical annotation categories of industrial equipment text (equipment component, parameter, numerical, state description, and action), the value of K is set to 5 to ensure that the clustering results match the semantic classification of the vocabulary in the industrial scenario. Five cluster centers are initialized by randomly selecting five dispersed vectors from the lexical feature vector set as initial centers. The selected vectors must correspond to the five lexical annotation categories respectively; for example, the vector for bearings is selected as the initial center for the equipment component category, the vector for vibration frequency as the initial center for the parameter category, and the vector for 65Hz as the initial center for the numerical category. ; Calculate the Euclidean distance from each word feature vector to the 5 cluster centers; The calculation process is as follows: For a single word feature vector, calculate the difference between its value in 257 dimensions and the value in the corresponding dimension of a certain cluster center, square each difference, and add the squared differences in 257 dimensions to obtain the Euclidean distance from the word feature vector to the cluster center; Repeat this process to obtain the 5 Euclidean distances from the vector to the 5 cluster centers; Assign the word feature vector to the category of the cluster center with the smallest Euclidean distance. For example, if the vector of motor has the smallest distance to the center of equipment parts, it will be assigned to the equipment parts category.

[0054] After completing the initial classification of all vectors, update the center of each cluster; for all word feature vectors in each cluster, calculate the sum of their values ​​in each dimension, divide the sum by the number of word feature vectors in that cluster to obtain the average value of that dimension; arrange the average values ​​of the 257 dimensions in order to form new cluster centers; repeat the process of calculating distance, classifying, and updating centers until the change in the value of each dimension of all cluster centers is less than 0.001 in two consecutive iterations. At this point, the clustering result is stable, and five feature subspaces are obtained, corresponding to the device component subspace, parameter subspace, numerical subspace, state description subspace, and action subspace, respectively.

[0055] Step 202 involves statistical analysis of each feature subspace, calculating the vector distribution characteristics in each feature subspace, and obtaining the statistical properties of each feature subspace. Specifically, this includes: calculating four types of vector distribution characteristics for each feature subspace sequentially, where the first type is the subspace mean vector. For all word feature vectors in this subspace, the first dimension value of each vector is extracted, and these values ​​are summed to obtain the sum of the first dimension. The same method is used to calculate the sum of each of the 257 dimensions. The sum of each dimension is divided by the number of word feature vectors in this subspace to obtain the average value of that dimension. The average values ​​of the 257 dimensions are arranged in order to form the mean vector of this feature subspace. For example, the mean vector of the equipment component subspace reflects the common semantic features of equipment component words such as bearings and motors in each dimension. If the semantic features of the vectors in this subspace are... The closer the mean vectors are, the more representative they are for that type of vocabulary. The second type is subspace variance. For each vocabulary feature vector in the subspace, the difference between its first dimension value and the first dimension value of the mean vector is calculated, and this difference is squared. The same method is used to calculate the squared differences of each of the 257 dimensions of the vector. The squared differences of all dimensions of the vector are summed to obtain the squared error of the vector. The squared errors of all vocabulary feature vectors in the subspace are summed to obtain the total squared error. The total squared error is divided by the number of vocabulary feature vectors in the subspace to obtain the variance of the feature subspace. The smaller the variance, the more concentrated the distribution of vocabulary feature vectors in the subspace, and the higher the semantic consistency. For example, the variance of the numerical subspace (containing vectors such as 65Hz and 85℃) is usually smaller than that of the equipment component subspace because the semantic structure of numerical vocabulary is more uniform.

[0056] The third category is the maximum and minimum distances within a subspace. For any two distinct word feature vectors within this subspace, calculate the Euclidean distance between them (the calculation method is the same as in step 201). Iterate through all pairwise vector combinations and record the maximum and minimum values ​​among all Euclidean distances, which are respectively used as the maximum and minimum intra-class distances for this feature subspace. For example, in a state description subspace (containing vectors for normal, abnormal, high, and low values), the Euclidean distance between normal and abnormal vectors may be the maximum intra-class distance, while the Euclidean distance between high and low vectors may be the minimum intra-class distance. The fourth category is the average distance within a subspace. For all pairwise vectors within this subspace... The Euclidean distances are summed to obtain the total distance. The total number of pairwise vector combinations within the subspace is calculated (the total number of combinations equals the number of vectors in the subspace multiplied by the number of vectors minus one, then divided by two). The total distance is divided by the total number of combinations to obtain the intra-class average distance of the feature subspace. The intra-class average distance reflects the overall semantic similarity of vectors within the subspace. For example, the intra-class average distance of the action subspace (including vectors for raising, lowering, triggering, etc.) can serve as an important basis for subsequent judgments on whether action-related words belong to the same phrase. The mean vector, variance, intra-class maximum distance, intra-class minimum distance, and intra-class average distance of each feature subspace are integrated to form the statistical characteristics of each feature subspace.

[0057] Step 203: Based on the statistical characteristics of each feature subspace, determine the boundary judgment conditions and phrase segmentation rules for chunk analysis, and obtain the analysis optimization parameters. Specifically, this includes: for every two adjacent word feature vectors, first determine the type of feature subspace they belong to; if the two vectors belong to different types of feature subspaces, such as the former belonging to the equipment component subspace and the latter to the parameter subspace, then they are initially determined to be potential phrase boundaries; if the two vectors belong to the same type of feature subspace, such as both belonging to the parameter subspace, then calculate the Euclidean distance between the two vectors and compare it with the intra-class average distance of that feature subspace; if the Euclidean distance is greater than the intra-class average distance... If the distance is 1.2 times the average distance within a class, it is considered a phrase boundary; if it is less than or equal to 1.2 times the average distance within a class, it is considered a non-boundary, and the threshold is adjusted in conjunction with the variance of the feature subspace; if the subspace variance is less than the preset value, the threshold is adjusted to 1.1 times the average distance within a class to avoid over-splitting semantically similar words; if the variance is greater than the preset value, the threshold is kept at 1.2 times to ensure accurate splitting of words with large semantic differences. For example, for the vector vibration frequency and temperature of two adjacent parameter subspaces, after calculating their Euclidean distance, if the distance is greater than 1.2 times the average distance within a class of parameter subspaces, it is considered a phrase boundary, indicating that the two do not belong to the same device parameter phrase.

[0058] Based on the textual expression logic of industrial equipment (component + parameter + value + state + action), and combined with the semantic association of each feature subspace, phrase combination priorities are set. The first priority is the combination of the equipment component subspace vector and the parameter subspace vector, such as the combination of bearing (equipment component) and vibration frequency (parameter) to form the bearing vibration frequency. The second priority is the combination of the first priority combination and the value subspace vector, such as the combination of bearing vibration frequency and 65Hz (value) to form the bearing vibration frequency 65Hz. The third priority is the combination of the second priority combination and the state description subspace vector, such as the combination of bearing vibration frequency 65Hz and abnormal (state) description vector. The first priority is the combination of state description) into a bearing vibration frequency of 65Hz abnormality; the second priority is the combination of any combination + action subspace vector combination, such as temperature (parameter) and rise (action) combined into temperature rise; at the same time, if a vector of a certain feature subspace appears alone (e.g., there are no adjacent vectors to combine), it is divided into an independent phrase, such as when the equipment (equipment component subspace vector) has no adjacent vectors to combine, it is treated as a separate equipment phrase; the above boundary judgment conditions (including subspace type judgment rules, distance threshold and variance adjustment strategy) and phrase division rules (including combination priority and independent phrase judgment rules) are integrated to form analysis and optimization parameters.

[0059] This embodiment quantifies the semantic consistency and difference of each subspace by calculating distribution features such as mean vector and variance, avoiding subjective judgment of statistical characteristics; the boundary judgment conditions and phrase segmentation rules generated based on statistical characteristics reduce phrase splitting or combination errors caused by semantic confusion.

[0060] In a preferred embodiment of the present invention, step 3, optimizing the text chunking analysis process according to the analysis optimization parameters, and segmenting sentences into non-overlapping phrases with syntactic function tags to obtain the chunked text, may include:

[0061] Step 301: Generate text segmentation rules based on the boundary judgment conditions in the analysis and optimization parameters to obtain optimized text segmentation rules; perform initial segmentation of the natural language text to be processed according to the optimized text segmentation rules to obtain a sequence of sentence units. Specifically, this includes: combining the three types of boundary judgment conditions in the analysis and optimization parameters—subspace type judgment, distance threshold, and variance adjustment strategy—and converting them into directly executable text segmentation rules, which contain four core rules. The first rule is the word boundary rule between different subspaces. When two adjacent words belong to the equipment component subspace and parameter subspace, parameter subspace and numerical subspace, or numerical subspace and state description subspace, respectively, they are not judged as segmentation boundaries. This conforms to the expression logic of industrial text: component plus parameter plus numerical plus state. When two adjacent words belong to the equipment component subspace and numerical subspace, parameter subspace and action subspace, or state description subspace and equipment component subspace, they are judged as segmentation boundaries. The second rule is the word distance rule within the same subspace. When two adjacent words belong to the equipment component subspace and numerical subspace, parameter subspace and action subspace, or state description subspace and equipment component subspace, they are judged as segmentation boundaries. When words belong to the same subspace, the Euclidean distance between the feature vectors of the two words is calculated, and this distance is compared with the result of multiplying the average distance within the subspace by 1.2. If the distance is greater than the result, it is determined to be a segmentation boundary; if the distance is less than or equal to the result, it is not determined to be a segmentation boundary. If the variance of the subspace is less than a preset value, which reflects the high semantic consistency of the subspace vectors, the multiplier is adjusted from 1.2 to 1.1 before comparison. The third rule is the action word independence rule. When the words in the action subspace, such as "rise" or "fall", are adjacent to words that do not belong to the parameter subspace or the state description subspace, they are determined to be segmentation boundaries before and after the action words. This is because in industrial text, actions need to be associated with parameters. For example, a temperature rise is meaningless if it is an action on its own. The fourth rule is the equipment identifier priority rule. For word combinations containing equipment numbers, such as M001, regardless of the subspace to which the adjacent words belong, they are determined to be segmentation boundaries before and after the equipment number word combination. This is because equipment identifiers need to be independent units to facilitate association with subsequent chapters.

[0062] The natural language text to be processed, such as "Vibration frequency of bearing in device M001 is 65Hz, state is abnormal, motor temperature rise," is scanned word by word according to the above rules: First, identify that device M001 contains the device number, and mark the boundaries before and after it according to the fourth rule to segment out the first unit, device M001; subsequently, the bearing belongs to the equipment component subspace and the vibration frequency belongs to the parameter subspace, which are non-boundary cases according to the first rule, so no boundary is marked; the vibration frequency belongs to the parameter subspace and 65Hz belongs to the numerical subspace, which are also non-boundary cases, so no boundary is marked; 65Hz belongs to the numerical subspace and the state belongs to the state description subspace, which are also non-boundary cases, so no boundary is marked; the state belongs to the state description subspace and the abnormal belongs to the state description subspace, which are also non-boundary cases, so no boundary is marked; Subspaces belonging to the same subspace are compared by calculating the Euclidean distance between their vectors. If the distance is less than the average distance of the state description subspace multiplied by 1.1, the boundary is not marked, resulting in the unit bearing vibration frequency of 65Hz being an abnormal state. Subsequent abnormalities belonging to the state description subspace and the motor belonging to the equipment component subspace are boundary cases that conform to the first rule, and the boundary is marked. The motor belonging to the equipment component subspace and the temperature belonging to the parameter subspace are non-boundary cases, and the temperature belonging to the parameter subspace and the rise belonging to the action subspace are boundary cases that conform to the first rule, and the boundary is marked, resulting in the units motor temperature and rise. Finally, the sentence unit sequence is obtained, namely, equipment M001, bearing vibration frequency of 65Hz being an abnormal state, motor temperature and rise.

[0063] Step 302 involves performing boundary matching processing between the sentence unit sequence and the vocabulary feature vector set to obtain a preliminary phrase boundary marking sequence. This sequence is then validated using the phrase segmentation rules in the analysis and optimization parameters to obtain a validated phrase boundary sequence. Specifically, this includes: extracting the subspace type corresponding to each word from the vocabulary feature vector set (e.g., M001 corresponds to the equipment identifier subspace, bearings correspond to the equipment component subspace, vibration frequency corresponds to the parameter subspace, etc.), and organizing these subspaces into a vocabulary subspace lookup table according to the original word sequence order; splitting each unit in the sentence unit sequence into individual words, determining the subspace type of each word by referring to the vocabulary subspace table, and then re-checking whether each adjacent word should be marked with a boundary according to the segmentation rules of Step 301, marking it as either a boundary or not, thus forming the preliminary phrase boundaries. The sequence is marked; taking the vibration frequency of a unit bearing as 65Hz and the state as abnormal as an example, the words are split into bearing, of, vibration frequency, as, 65Hz, state, is, abnormal, and the corresponding subspace types are, in order, equipment component, function word, parameter, function word, numerical value, state description, function word, state description; adjacent words are checked, including bearing and of which belong to equipment component and function word, non-boundary; of and vibration frequency belong to function word and parameter, non-boundary; vibration frequency and as belong to parameter and function word, non-boundary; as and 65Hz belong to function word and numerical value, non-boundary; 65Hz and state belong to numerical value and state description, non-boundary; state and is belong to state description and function word, non-boundary; is and abnormal belong to function word and state description, non-boundary. All adjacent positions are initially marked as non-boundary, and the unit as a whole is a preliminary phrase boundary segment.

[0064] Combining the phrase segmentation rules in the optimization parameters, which include combination priority and independent phrase determination, the preliminary boundary marker sequence is checked and adjusted segment by segment. The first verification logic is combination priority verification. If the preliminary boundary segment contains a combination of equipment component, parameter, and value, belonging to the first and second priority, it needs to be confirmed that there are no incorrectly marked boundaries within this combination. If the preliminary boundary segment only contains parameter and action, belonging to the fourth priority, it needs to be confirmed that the action word is not bound to irrelevant words. Taking the unit motor temperature as an example, the word is split into motor and temperature, and the subspace type is equipment component and parameter, which meets the first priority combination. It is initially marked as non-boundary and maintained after verification. Non-boundary; taking unit elevation as an example, the split words are only elevation belonging to the action subspace, and there are no parameter subspace words before and after it, which meets the independent phrase determination rules. The words before and after it are initially marked as boundaries, and the boundaries are maintained after verification; if there is an initial boundary segment containing equipment components plus values ​​that do not meet any priority, such as bearing 65Hz, then the boundary needs to be added between bearing and 65Hz, splitting it into two segments: bearing and 65Hz; the final adjusted boundary marking sequence is the boundary before and after of equipment M001, the boundary before and after the bearing vibration frequency of 65Hz is abnormal, the boundary before and after of motor temperature, and the boundary before and after elevation, which is the verified phrase boundary sequence.

[0065] Step 303: Segment the sentence unit sequence according to the verified phrase boundary sequence to obtain non-overlapping phrase units; perform syntactic function matching between the non-overlapping phrase units and the vocabulary feature vector set to obtain phrases with syntactic function labels, and reorganize the text according to the original word order to obtain the text after chunking analysis. Specifically, this includes: cutting at the boundary positions of the sentence unit sequence according to the verified phrase boundary sequence to ensure that each cut phrase unit does not overlap and can cover the entire text; taking the sentence unit sequence M001, bearing vibration frequency is 65Hz, state is abnormal, motor temperature, rise as an example, after cutting according to the boundary, four non-overlapping phrase units are obtained: M001, bearing vibration frequency is 65Hz, state is abnormal, motor temperature, rise.

[0066] The subspace type of the core vocabulary within each phrase unit is extracted from the lexical feature vector set. Core vocabulary refers to non-function words. For example, the phrase unit "the vibration frequency of the bearing is 65Hz and its state is abnormal" has the core vocabulary "bearing," "vibration frequency," "65Hz," and "abnormal." Based on the subspace type and the syntactic logic of industrial text, five types of syntactic action tags and matching rules are defined. The first type is the equipment identification phrase tag, and the matching rule is that the phrase unit contains the equipment number, such as M001, or the equipment name, such as "cooling fan," and the core vocabulary belongs to the equipment identification-related subspace. For example, equipment M001 matches this tag. The second type is the equipment component phrase tag, and the matching rule is that the core vocabulary of the phrase unit belongs to the equipment. The first category is component subspaces, which do not contain core words from other subspaces, or only contain function words. For example, if a bearing is a single phrase unit, it matches this tag. The second category is parameter and numerical phrase tags, where the core words of the phrase unit contain words from both the parameter and numerical subspaces. For example, a vibration frequency of 65Hz matches this tag. The third category is parameter and state phrase tags, where the core words of the phrase unit contain words from both the parameter and state description subspaces, or words from both the numerical and state description subspaces. For example, 65Hz is an abnormal state and matches this tag. The fifth category is action phrase tags, where the core words of the phrase unit belong to the action subspace. For example, "lifting" matches this tag.

[0067] For each phrase unit, match them one by one according to the above rules. The device numbered M001 contains the device number and matches the device identification phrase tag; the state where the vibration frequency of the bearing is 65 Hz is abnormal includes the device component bearing, the parameter vibration frequency, the value 65 Hz, and the state description abnormal. After integration, it is matched as the device component plus parameter plus state phrase tag. Since this phrase unit contains multiple types of core vocabulary, a combined tag is required; motor temperature includes the device component motor and the parameter temperature, and matches the device component plus parameter phrase tag; rise belongs to the action subspace and matches the action phrase tag; According to the original word order of the natural language text to be processed, arrange the phrases with syntactic role tags in sequence, and mark the tags in parentheses behind the corresponding phrases to form the text after chunk analysis.

[0068] In this embodiment, through the matching of syntactic role tags, the function of each phrase in the industrial equipment text is clarified, such as device identification and parameter description. The restructured text not only conforms to the original word order but also clearly presents the semantic structure, providing a structured phrase basis for subsequent syntactic and semantic parsing, and improving the accuracy and logic of industrial equipment status report text processing.

[0069] In a preferred embodiment of the present invention, in step 4, perform syntactic and semantic structure parsing on the text after chunk analysis. Through constituent analysis, dependency analysis, and semantic dependency graph analysis, establish the semantic association relationship between the internal structures of the sentence to obtain structured semantic information, which may include:

[0070] In step 401, perform constituent analysis on the text after chunk analysis to identify the subject, predicate, and object grammatical components in the sentence, and obtain the phrase structure tree; perform dependency analysis on the phrase structure tree to obtain the dependency relationship set, specifically including: the processing object is the text after chunk analysis. For example, the device numbered M001 corresponds to the device identification phrase, the vibration frequency of the bearing is 65 Hz corresponds to the device component plus parameter plus value phrase, the state is abnormal corresponds to the parameter state phrase, the motor temperature corresponds to the device component plus parameter phrase, and rise corresponds to the action phrase; when analyzing, first locate the main sentence backbone, and the core is to identify the core predicate; in industrial equipment text, "is" is usually used to connect parameters and monitoring values, "is" is usually used to connect parameters and operating states, and verbs such as "rise" and "fall" are used to describe the change trend of parameters. These three types of verbs will be preferentially determined as core predicates; taking the example text as an example, the core predicates are "is", "is", and "rise" respectively. Accordingly, the original sentence is split into three clauses. The first clause focuses on "is", that is, the vibration frequency of the bearing of the device numbered M001 is 65 Hz; the second clause focuses on "is", that is, the state is abnormal; the third clause focuses on "rise", that is, the motor temperature rises.

[0071] Construct a phrase structure tree for each clause one by one. Take the vibration frequency of the bearing of device M001 in the first clause as an example, which is 65Hz. First, divide the basic phrase types. That is, the whole clause is composed of a noun phrase and a verb phrase. Among them, the verb phrase is "is 65Hz" because the core verb of this clause is "is", and 65Hz is the object dominated by this verb. The noun phrase is "the vibration frequency of the bearing of device M001" because this part is the object described by the core verb "is", that is, the subject of the sentence. Stratify and subdivide the noun phrase. That is, in "the vibration frequency of the bearing of device M001", the vibration frequency is the central word to be modified, and "the bearing of device M001" is the modifier, that is, the attributive. Therefore, this noun phrase is split into the attributive "the bearing of device M001" plus the central word "vibration frequency". Further split "the bearing of device M001". Among them, the bearing is the central word, and "device M001" is the attributive used to limit the device to which the bearing belongs. So it is also split into the attributive "device M001" plus the central word "bearing". Subdivide the verb phrase. In "is 65Hz", "is" is the core verb, and 65Hz is the object of the verb, belonging to the noun phrase type. Therefore, it is split into the core verb "is" plus the object "65Hz". Split all clauses according to this logic, and mark the grammatical components of each phrase, such as attributive, central word, predicate, and object, in the hierarchical structure to form a phrase structure tree. The top layer of the tree is the clause, the next layer is the phrase types such as noun phrase and verb phrase, and the next layer is the specific grammatical components and corresponding words, ensuring that the grammatical functions of each word are clearly reflected in the tree.

[0072] Based on the phrase structure tree, construct the dependency relationship using the dependency grammar theory. The core is to determine the grammatical dependency of each word on the core word. First, determine the core word for each clause. The core word is the core predicate. The core word of the first clause is "is", the core word of the second clause is "is", and the core word of the third clause is "increases". Take the first clause as an example to analyze the relationship between each word and the core word "is". The vibration frequency is the subject described by "is", that is, the value of who or what is 65Hz. The two form a subject-predicate dependency relationship. 65Hz is the object dominated by "is", that is, the specific value connected by "is", forming a verb-object dependency relationship. The bearing is used to modify the vibration frequency, that is, which type of vibration frequency, forming a modifier-head dependency relationship. "Device M001" is used to modify the bearing, that is, the bearing of which device, forming a modifier-head dependency relationship.

[0073] The core word of the second clause, "state," is "is," and "state" is the subject described by "is," forming a subject-predicate relationship. "Abnormality" is the object governed by "is," forming a verb-object relationship. Simultaneously, "state" specifically refers to the vibration frequency in the previous clause, describing the state of vibration frequency, requiring additional appositive dependency to clarify the referent of "state." The core word of the third clause, "motor temperature rises," is "rises," and "motor temperature rises" is the subject described by "rises," i.e., what parameter is rising, forming a subject-predicate relationship. "Motor" modifies "temperature," i.e., what type of temperature, forming a noun-head dependency relationship. All dependency relationships are organized according to the format of dependency word, core word, and relationship type. For example, vibration frequency, "is," subject-predicate dependency relationship; 65Hz, "is," verb-object dependency relationship; bearing, vibration frequency, noun-head dependency relationship; M001 equipment, bearing, noun-head dependency relationship, ultimately forming a set of dependency relationships.

[0074] Step 402: Perform semantic role annotation on the phrase structure tree and dependency relation set to obtain semantic role annotation results; perform semantic association processing on the semantic role annotation results and the vocabulary feature vector set to construct semantic dependencies between words and obtain a semantic relation network. Specifically, this includes: combining the grammatical components of the phrase structure tree and the dependency types of the dependency relation set, and customizing a semantic role system based on the industrial equipment scenario, with a total of seven roles: Equipment Identifier role corresponding to words containing equipment number and model, such as equipment M001; Component role corresponding to physical components of the equipment, such as bearings and motors; Parameter role corresponding to monitoring indicators, such as vibration frequency and temperature; Numerical role corresponding to the monitored values ​​of parameters, such as 65Hz and 85℃; State role corresponding to the operating status of parameters, such as abnormal and normal; Action role corresponding to parameter changes, such as increase and decrease; and Predicate role corresponding to verbs connecting core elements, such as "is".

[0075] Each clause is labeled with its semantic role. In the first clause, "The vibration frequency of the bearing in device M001 is 65Hz," device M001 contains the device number and is labeled as the device identifier; the bearing is a device component and is labeled as the component; the vibration frequency is a monitoring parameter and is labeled as the parameter; "to" is the verb connecting the parameter and the value and is labeled as the predicate; 65Hz is the monitored value of the parameter and is labeled as the value. In the second clause, "The state is abnormal," the state refers to the vibration frequency in the previous clause, which is the parameter and is labeled as the parameter association; "to" is the verb connecting the parameter and the state and is labeled as the predicate; "abnormal" is the state of the parameter and is labeled as the state. In the third clause, "The motor temperature rises," the motor is a device component and is labeled as the component; the temperature is a monitoring parameter and is labeled as the parameter; "rises" is the action of the parameter change and is labeled as the action. When annotating, it is necessary to strictly match the grammatical components and dependency relationships. For example, the vocabulary of the component role should be the modifier or headword in the modifier-head dependency relationship, such as bearing is the modifier of vibration frequency. The vocabulary of the numerical role should be the object in the verb-object dependency relationship, such as 65Hz is the object of . This ensures that the role annotation is consistent with the grammatical function and forms the semantic role annotation result.

[0076] From the vocabulary feature vector set, a 257-dimensional feature vector corresponding to each word with a labeled semantic role is extracted, containing 256 intermediate features plus 1-dimensional category probability. The vector similarity of word pairs with association potential, such as components and parameters, or parameters and values, is calculated using Euclidean distance. Specifically, the feature vectors of two words are taken, the difference in each dimension is calculated, the squares of all differences are summed, and the square root of the sum is taken to obtain the Euclidean distance. Simultaneously, the average intra-class distances calculated in step 202 are used; for example, the average distance for the equipment / component subspace is 0.8, for the parameter subspace 0.7, and for the value subspace 0.6. If the Euclidean distance between two words is less than the mean of the average distances in their corresponding subspaces (e.g., the average distance between components and parameters in the subspace is 0.75), then the two are considered to have a close semantic association. For example, the Euclidean distance between a bearing (component role) and vibration frequency (parameter role) is calculated. If the sum of the squared differences between the values ​​of each dimension of the bearing vector and the corresponding dimension of the vibration frequency vector is 0.36, and the square root distance is 0.6, which is less than 0.75, a semantic dependency relationship between component parameters is determined. The distance between the vibration frequency (parameter role) and 65Hz (numerical role) is calculated. If the sum of the squared differences is 0.25, the distance is 0.5, which is less than the average distance of 0.65 between the parameter and numerical subspaces, a semantic dependency relationship between parameter values ​​is determined. The distance between the vibration frequency (parameter role) and the abnormality (state role) is calculated. If the distance is 0.6, which is less than the average distance of 0.7 between the parameter and state subspaces, a semantic dependency relationship between parameters and states is determined. The distance between the motor temperature (component plus parameter role) and the rise (action role) is calculated. If the distance is 0.55, which is less than the average distance of 0.6 between the parameter and action subspaces, a semantic dependency relationship between parameters and actions is determined.

[0077] All semantic dependencies are organized into vocabulary 1, semantic relation type, vocabulary 2, and association strength. Association strength is represented by the reciprocal of Euclidean distance. For example, the strength of a distance of 0.6 is 1 divided by 0.6, which is approximately 1.67. A semantic relation network is constructed with vocabulary as nodes, node attributes containing vocabulary content and semantic role, and semantic dependencies as edges, edge attributes containing relation type and association strength.

[0078] Step 403 involves associating and fusing the semantic relation network and dependency relation set to establish a correspondence between grammatical structures and semantic relations, resulting in a complete semantic association. Specifically, this includes traversing each dependency relation in the dependency relation set and finding the corresponding semantic dependency relation in the semantic relation network through lexical pair matching. That is, the lexical pair consisting of the dependency word and the core word in the dependency relation must match the lexical pair consisting of word 1 and word 2 in the semantic relation. For example, in the dependency relation, bearing and vibration frequency form a definite-central relation, and the lexical pair is bearing and vibration frequency. The semantic relation network then finds the relation containing this lexical pair, i.e., bearing and vibration frequency forming a component parameter relation, establishing a definite-central relation. The mapping of semantic correspondence between the component parameters in the grammar is established. In the dependency relation, the vibration frequency and the lexical pair form a subject-predicate relation, and the vibration frequency and the lexical pair form a parameter predicate relation. The mapping of semantic correspondence between the subject-predicate grammar and the parameter predicate semantic is established. In the dependency relation, 65Hz and the lexical pair form a verb-object relation, and the lexical pair form a numerical predicate relation. The mapping of semantic correspondence between the verb-object grammar and the numerical predicate semantic is established. In the dependency relation, the motor temperature and the rise form a subject-predicate relation, and the lexical pair form a parameter-action relation. The mapping of semantic correspondence between the subject-predicate grammar and the parameter-action semantic is established.

[0079] For each corresponding mapping, check according to the syntax type and semantic type matching rules. If the dependency relationship is noun-head (modifier and modified), the corresponding semantic relationship must belong to the core class of modifier, such as component parameter, device identifier, component; if the dependency relationship is subject-verb (subject and action), the corresponding semantic relationship must belong to the core element predicate or action class, such as parameter predicate, parameter action; if the dependency relationship is verb-object (action and object), the corresponding semantic relationship must belong to the object predicate class, such as numerical predicate, state predicate. If a mismatch is found, backtracking and adjustment are required. For example, in a dependency relationship, exception and yes form a verb-object relationship. If the semantic relationship of the word pair is exception and yes is incorrectly labeled as exception and yes forming a parameter predicate relationship, there will be no match because exception is a state role. The semantic role labeling needs to be rechecked to confirm the role of exception as a state role, and the semantic relationship needs to be corrected to exception and yes forming a state predicate relationship. Then, the vector similarity is recalculated to confirm the association until a match is found.

[0080] After integrating the matched relationships, supplement the semantic context information. The parameter numerical semantic relationship needs to be associated with the corresponding device identifier and component, such as the bearing vibration frequency of device M001 being 65Hz. The context information comes from the device identifier phrase and component phrase after block analysis. The parameter action semantic relationship needs to be associated with the timing of the action, such as the timing of the motor temperature rise being the acquisition time 2025.09.30 14:00 in the structured data, where the timing information comes from the structured data received in step 102. After integration, a complete semantic relationship is formed. Each relationship includes grammatical dependency type, semantic relationship type, associated vocabulary, and associated context, ensuring that the grammatical and semantic relationships are traceable.

[0081] Step 404 integrates the phrase structure tree, dependency relation set, and complete semantic association to obtain structured semantic information. Specifically, this includes: a four-layer hierarchical structure, with each layer assigned a unique identifier for easy location. The first layer is sentence-level information, identified as S1, containing the number of clauses (e.g., 3) and the clause order: S1, C1S1, C2S1, C3. The second layer is clause-level information, where each clause corresponds to a sub-frame (e.g., S1, C1 corresponding to the first clause), containing the phrase structure tree of that clause, the dependency relation subset within the clause containing only the clause's dependency relations, and the complete semantic association subset within the clause containing only the clause's semantic association. The third layer is phrase-level information, where each phrase corresponds to a sub-frame (e.g., S1, C1, NP1 corresponding to the first clause). The first noun phrase of a clause includes the phrase type (noun phrase, verb phrase), vocabulary within the phrase (e.g., the vibration frequency of the bearing of device M001), the corresponding grammatical components (modifier and headword), and the semantic roles within the phrase (device identifier role, component role, and parameter role). The fourth layer is lexical information, where each word corresponds to a subframe (e.g., S1, C1, NP1, W1). S1 corresponds to the first word of the first noun phrase of the first clause, containing lexical content (e.g., device M001), the subspace type (device identifier subspace), grammatical component label (modifier), dependency relationship (device M001 and bearing form a modifier-head relationship), semantic role label (device identifier role), and semantic association relationship (device M001 and bearing form a device identifier-component relationship).

[0082] Fill in the information according to the framework. Taking clauses S1 and C1, namely the vibration frequency of bearing M001, as 65Hz, as an example, the clause-level information includes: The phrase structure tree is filled with the hierarchical splitting results of the noun phrase "vibration frequency of bearing M001" and the verb phrase "65Hz"; the dependency subset is filled with "vibration frequency" forming a subject-predicate relationship, "65Hz" forming a verb-object relationship, "bearing" and "vibration frequency" forming a noun-head relationship, and "M001" and "bearing" forming a noun-head relationship; the complete semantic association subset is filled with "M001" and "bearing" forming an equipment identification component relationship, "bearing" and "vibration frequency" forming a component parameter relationship, and "vibration frequency" and "65Hz" forming a parameter relationship. Numerical relationships; In phrase-level information, S1, C1, NP1, which is the vibration frequency of bearing M001, should be filled in with the phrase type noun phrase, vocabulary M001 bearing vibration frequency, grammatical components as attributive M001 bearing plus headword vibration frequency, semantic role set equipment identifier role plus component role plus parameter role; In lexical-level information, S1, C1, NP1, W2, which is bearing, should be filled in with the vocabulary content bearing, subspace type equipment component subspace, grammatical components label attributive, dependency relationship bearing and vibration frequency form attributive-head relationship, semantic role label component role, semantic association relationship bearing and vibration frequency form component parameter relationship.

[0083] Verify consistency across all levels. Phrasal-level grammatical components must match the clause-level phrase structure tree. For example, the modifiers in S1, C1, and NP1 must match the modifying function of bearing M001 in the phrase structure tree. Lexical-level dependency relations must match the clause-level dependency subsets. For example, the dependency of "bearing" must be within the dependency subset of S1 and C1. Lexical-level semantic association relations must match the clause-level semantic association subsets. For example, the semantic association of "bearing" must be within the semantic subset of S1 and C1. Supplement cross-clause association information. If the vibrational frequencies of the preceding clauses S1 and C1 have an appositive dependency relationship with the states of the following clauses S1 and C2, supplement the appositive dependency vibrations in the lexical-level information of the vibrational frequencies (S1, C1, NP1, W3) and the lexical-level information of the states (S1, C2, NP1, W1), respectively. Frequency and state form an appositive relationship, state and vibration frequency form an appositive relationship, and corresponding semantic association relationships are formed. Vibration frequency and anomaly form a parameter state relationship. The hierarchical information is converted into a nested dictionary format. The example fragment is sentence-level information containing the identifier S1, the number of clauses 3, and the clause order S1, C1, S1, C2, S1, C3. In the clause-level information, S1 and C1 contain phrase structure trees. The noun phrase is the vibration frequency of the bearing of device M001, which is an attributive plus a headword. The verb phrase is 65Hz, which is a core verb plus an object. The dependency subset contains vibration frequency and forms a subject-predicate relationship. 65Hz and forms a verb-object relationship. The semantic association subset contains device M001 and the bearing, forming a device identifier component relationship. This ensures that each information node has a clear identifier and association path, forming complete structured semantic information.

[0084] This embodiment achieves precise alignment of syntax and semantics through word pair matching, and then corrects annotation errors through consistency checks, ensuring the synchronization of grammatical structure and semantic association, and improving the reliability of semantic relationships.

[0085] In a preferred embodiment of the present invention, step 5, performing multi-sentence logical association analysis on the structured semantic information at the discourse level to obtain the semantic information of the entire discourse, may include:

[0086] Step 501: Identify the logical relationships between multiple sentences in the structured semantic information to obtain a set of logical relationships between sentences, and match them with the phrase structure tree to obtain a discourse-level grammatical structure. Analyze the discourse-level grammatical structure and the dependency relationship set to obtain discourse-level dependency relationships, specifically including: the core semantic elements of each clause contained in the structured semantic information, including the component role vocabulary, parameter role vocabulary, state or action or numerical role vocabulary of the clause, and the collection time sequence information corresponding to these elements, which comes from the industrial equipment structured data received in step 102, and also includes the semantic association strength between the vocabulary of each clause, which comes from the reciprocal of the Euclidean distance calculated in step 402; during identification, first extract the core semantic unit of each clause. The core semantic unit is composed of component role vocabulary plus parameter role vocabulary plus corresponding state or action or numerical role vocabulary. For example, the core semantic unit of one clause is cooling fan plus speed plus below rated value, the core semantic unit of another clause is CPU plus temperature plus 85℃, and the core semantic unit of yet another clause is CPU plus over-temperature warning plus 2 times.

[0087] To determine the type of logical relationship between clauses, common logical relationships in industrial equipment texts are divided into three categories. The method for determining causal relationships is as follows: First, compare the acquisition time of the core semantic units of the two clauses. If the acquisition time of clause A is earlier than that of clause B, then check the semantic association strength of the core semantic units of the two clauses. If the strength is greater than a preset threshold (this threshold is set based on experience in industrial equipment fault association to ensure that the association conforms to the equipment's operating rules), and the core action or state of clause A can induce changes in the core state or value of clause B (e.g., a speed below the rated value can induce a temperature increase), then it is determined to be a causal relationship and labeled as clause A as cause and clause B as effect. The method for determining parallel relationships is as follows: The method for determining a parallel relationship is as follows: if the core semantic units of two clauses contain the same component role words, and the parameter role words are different but both are monitoring items of that component, such as motor vibration frequency and motor energy consumption, and the collection time sequence is similar with a time difference of less than 5 minutes, then these conditions are met and the clauses are judged to be parallel, labeled as clause A parallel and clause B parallel. The method for determining a progressive relationship is as follows: if the core semantic units of two clauses contain the same component and parameter role words, and the state role words of the latter clause indicate a stronger degree than those of the former clause, such as high temperature and abnormal temperature, and the collection time sequence of the latter clause is later, then these conditions are met and the clauses are judged to be progressive, labeled as clause A progressive before clause B progressive after clause.

[0088] All identified logical relationships are organized according to the format of clause identifier + logical relationship type + clause identifier to form a set of logical relationships between sentences. For example, clause 1 is cause and clause 2 is effect, clause 2 is cause and clause 3 is effect, clause 4 is parallel and clause 5 is parallel. A phrase structure tree association matching is performed to obtain a discourse-level grammatical structure. The phrase structure tree of each clause is extracted, focusing on noun phrases and verb phrases. Noun phrases correspond to components plus parameters of the core semantic unit, and verb phrases correspond to actions plus states of the core semantic unit. Based on the set of logical relationships between sentences, the phrase structure trees of clauses with logical relationships are associated as follows: if the verb phrase "speed is lower than rated value" in clause 1 and the noun phrase "CPU temperature" in clause 2 are causally related, then "verb phrase 1 cause and noun phrase 2 effect" is marked in the discourse-level grammatical structure; if the noun phrase "motor vibration frequency" in clause 4 and the noun phrase "motor energy consumption" in clause 5 are parallel, then "noun phrase 4 is parallel" and "noun phrase 5 is parallel" is marked. Simultaneously, all phrase structure trees are arranged according to the original clause order to form a discourse-level grammatical structure that contains both logical relationships and retains the original word order.

[0089] The dependency relationships within the original clauses, such as attributive-head, subject-predicate, and verb-object relationships, need to be supplemented with inter-clause dependency relationships: For clauses with causal relationships, the core verb (i.e., the core word of the verb phrase) in the first clause and the core noun (i.e., the core word of the noun phrase) in the second clause form a causal dependency, marked as "core verb in the first clause causal dependency, core noun in the second clause"; For clauses with coordinate relationships, the core nouns (i.e., the core words of the noun phrases) in both clauses form a coordinate dependency, marked as "core noun in clause 4, coordinate dependency in clause 5"; The clauses corresponding to the progressive relationship form a progressive dependency between the core verbs of the two clauses, which are also the core words of the verb phrases. This is marked as "Center verb of clause A progressively depends on core verb of clause B". Supplementary appositive dependency, which comes from step 401, for example, the temperature of clause 2 refers to the CPU temperature of clause 1, and is marked as "Temperature of clause 2 refers to CPU temperature of clause 1". Integrate the dependency relationships within and between clauses, and organize them according to the dependency word containing the clause identifier, the dependency type, and the core word containing the clause identifier to form a discourse-level dependency relationship.

[0090] Step 502 involves fusing the discourse-level dependency relations with the semantic relation network to obtain cross-sentence semantic associations; integrating these cross-sentence semantic associations with the complete semantic associations to establish a discourse-level semantic framework, which is then validated to obtain the overall discourse semantic information. Specifically, this includes fusing the discourse-level dependency relations with the semantic relation network. The semantic relation network contains semantic dependencies such as component parameters and parameter values ​​within clauses, while the discourse-level dependency relations contain causal and coordinate dependencies between clauses. During fusion, the lexical nodes corresponding to dependency words and core words in the discourse-level dependency relations are first found in the semantic relation network, and then cross-sentence semantic associations are supplemented according to the type of dependency between clauses: causal dependencies correspond to cross-sentence semantic associations... For sentence semantic association, if the preceding clause node corresponds to a component with added parameters and actions, and the following clause node corresponds to a component with added parameters and values, then the preceding clause component's parameters, actions, causes, and following clause component's parameters, values, and effects are added. The association strength is taken as the association strength of the core words of the two clauses in step 402. For cross-sentence semantic association corresponding to parallel dependencies, if the two clause nodes correspond to the same component with different parameters, then the component clause 4 parameters and the parallel component clause 5 parameters are added. The association strength is taken as the association strength of the core words of the two clauses. For cross-sentence semantic association corresponding to progressive dependencies, if the two clause nodes correspond to the same parameter with different states, then the parameter clause A state and the progressive parameter clause B state are added. The association strength is taken as the association strength of the core words of the two clauses.

[0091] For example, in the semantic relation network, clause 1 (lower than) and clause 2 (temperature) (causal dependency), find that "lower than" corresponds to "cooling fan speed lower than rated value" and "temperature" corresponds to "CPU temperature 85℃". The cross-sentence semantic association is then added as "cooling fan speed lower than rated value causal CPU temperature 85℃", with an association strength of 1.6. Similarly, clause 4 (vibration frequency) and clause 5 (energy consumption) (parallel dependency), are added as "motor vibration frequency parallel motor energy consumption", with an association strength of 1.2. These associations are then organized to form cross-sentence semantic associations. Integrating these cross-sentence semantic associations with the complete semantic association relationships... The complete semantic association comes from step 403, which constructs a chapter-level semantic framework according to the hierarchy of device, component, and parameter. The top layer of the framework is the device identifier, such as device M001; the next layer is the set of components, such as cooling fan, CPU, and motor; the next layer is the set of parameters corresponding to each component, such as the speed of the cooling fan, the temperature and over-temperature warning of the CPU, and the vibration frequency and energy consumption of the motor; the bottom layer is the semantic information of the parameters, including the parameter's value, state, or action, the acquisition time sequence, semantic association within sentences, and semantic association across sentences.

[0092] For example, the semantic information for the cooling fan speed under device M001 includes "below rated value," "collection time 2025.09.30 10:00," "within clause 1, cooling fan speed plus below rated value," and "cross-sentence, corresponding CPU temperature 85℃." The semantic information for the CPU temperature includes the value 85℃, "collection time 2025.09.30 10:05," "within clause 2, CPU temperature plus 85℃," and "cross-sentence, corresponding cooling fan speed below rated value and two over-temperature warnings." The discourse-level semantic framework is validated, which is divided into logical consistency validation and information integrity validation. Logical consistency validation checks whether the cross-sentence semantic associations conform to the operation of industrial equipment. The pattern is as follows: for example, a low cooling fan speed leading to a high CPU temperature and then an over-temperature warning, which conforms to the pattern of poor heat dissipation causing temperature rise, and is judged as consistent. If an over-temperature warning occurs and the CPU temperature is high, it does not conform to the pattern, and step 501 needs to be backtracked to re-identify the logical relationship. The method of verifying information integrity is to check whether the framework contains the core semantic elements of all clauses. These elements include equipment, components, parameters, values ​​or states, and timing. If motor energy consumption parameter information is missing, step 404 needs to be backtracked to supplement the structured semantic information. After the verification is passed, the chapter-level semantic framework is organized into a complete set of information including equipment, components, parameters, semantic associations, logical relationships, and collection timing, to obtain the semantic information of the entire chapter.

[0093] This embodiment integrates discourse-level dependencies and semantic networks to supplement cross-sentence semantic associations, integrates information according to the parameter level of equipment components, and verifies the logic and completeness. The resulting overall discourse semantic information can fully present the association logic and operating status of the parameters of each component of the equipment, providing semantic support for the subsequent generation of coherent and accurate industrial equipment status reports.

[0094] In a preferred embodiment of the present invention, step 6, obtaining the target natural language text using a pre-trained large language model based on the semantic information of the entire text, may include:

[0095] Step 601: The semantic information of the entire text is converted into a serialized semantic data; the serialized semantic data is input into a pre-trained large language model for text generation to obtain the initial generated text, which specifically includes: the semantic information of the entire text is a hierarchical structure of device, component, and parameter, including device identifier, component set, parameter information corresponding to each component, and parameter information including parameter name, value or status or action, acquisition time sequence, and cross-sentence relationship. During the conversion, following the presentation logic of industrial equipment status reports, hierarchical information is broken down into ordered, structured descriptive items to ensure that each core piece of information has a clear correspondence: Equipment identification information is extracted to clarify the equipment number and name, such as equipment identification M001 industrial motor; component sets are extracted and sorted according to their functional priority within the equipment, with core components given priority, such as the component list: cooling fan, CPU, bearing, motor body; for each component, all corresponding parameter information is extracted and organized in the format of parameter name plus attribute. Attributes include the parameter's value, status, or action, acquisition time, and cross-sentence association type, such as component 1 cooling fan; parameter 1-1 speed; parameter 1-1 attribute action below rated value, acquisition time 2025.09.3010: 00. Cross-sentence associations correspond to CPU temperature changes; Parameter 1-2 operating noise; Parameter 1-2 attribute status normal, collection time 2025.09.30 10:00; Supplement the overall logical association overview, summarize all cross-sentence associations, such as the logical association that the cooling fan speed is lower than the rated value (cause) corresponds to the CPU temperature rising to 85℃ (effect); the CPU temperature of 85℃ (cause) corresponds to the over-temperature warning being triggered twice (effect); the vibration frequency of the motor body and energy consumption (parallel); Arrange the above-organized information in the order of equipment identification to component list to component parameter details to logical association overview, forming a clear description sequence, with each description item separated by a uniform semicolon, ensuring no information omissions and a regular structure, to obtain serialized semantic data.

[0096] Model selection and infrastructure construction included choosing GPT-3.5 as the base model due to its strong ability to organize long text logic and understand semantics, making it suitable for generating texts like industrial equipment reports that require precise transmission of parameter relationships. The model structure retained the original Transformer encoder-decoder architecture, but the vocabulary of the decoder output layer was adjusted by adding industrial equipment-specific terms such as vibration frequency, over-temperature warning, and rated values, expanding the vocabulary to 50,000 words to ensure accurate output of domain-specific terminology. Training data preparation involved selecting 50,000 valid status reports from the historical database of the enterprise's industrial equipment operation and maintenance system. These reports covered 10 common industrial equipment categories, including motors, cooling systems, hydraulic devices, and gearboxes, and included four scenarios: normal operation, minor faults, moderate faults, and emergency shutdown. Each report was labeled with structured tags for equipment parameters, relationships, and conclusions to ensure domain-appropriateness of the training data. The model training process consisted of two phases. The first phase was domain-appropriate training, where the training data was broken down into parameter descriptions, relationship analysis, and scenario conclusions, and input into the model to learn the expression logic of industrial texts. The learning rate was set to 0.0001, with 64 data points input per batch, and training for 10 epochs. The cross-entropy loss function was used to calculate the loss value between the model's predicted text and the actual report. Backpropagation was used to adjust the weights of each layer of the model, such as the attention weights of the Transformer layer and the biases of the fully connected layers, enabling the model to understand industry-specific correlations such as a cooling fan malfunction corresponding to an increase in CPU temperature. The second stage was instruction fine-tuning training. Specific instructions for generating equipment reports were designed, such as generating status reports containing collection time, parameter values, and cross-component impacts based on equipment component parameters and their relationships. Serialized semantic data format inputs were paired with corresponding instructions to form training samples of instructions plus inputs plus outputs, trained for 5 epochs to further optimize the model's ability to parse serialized data and control the output format. After training, the serialized semantic data is used as the model input. The generation parameters are set, with a maximum generation length of 1000 characters to cover the common length of a complete status report for a single device. The temperature is set to 0.7 to balance the accuracy and naturalness of the generated text, avoiding excessive rigidity or logical confusion. The repetition penalty coefficient is 1.2 to reduce the repetition of words such as temperature and parameters. The Top-p sampling threshold is 0.9 to ensure that the generated text focuses on core information. Initial text generation: The model generates text according to the paragraph structure from the overall overview of the device to the description of the parameters of each component, to the cross-component correlation analysis, and finally to the overall status summary. First, the device identification and monitoring time range are introduced, then the parameter values, status and collection time of each component are described one by one. Next, the causal and parallel relationships between parameters are analyzed, and finally the overall operating status of the device is summarized to obtain the initial generated text.

[0097] Step 602: Perform semantic and logical consistency verification on the initial generated text to obtain the verified text; optimize the verified text by adjusting vocabulary selection and sentence structure to obtain the optimized text; perform a final semantic quality assessment on the optimized text to obtain the target natural language text, specifically including: based on the discourse-level semantic framework obtained in step 502, organize five levels of comparison items: equipment, component, parameter, association, and time sequence. For example, motor M001 includes cooling fan, speed, below rated value, 2025.09.3010:00, and corresponding CPU temperature of 85℃, forming a verification comparison library; scan the initial generated text sentence by sentence to extract key information, including equipment number, component name, parameter value, association type, and collection time, and match it with the corresponding items in the verification comparison library. Parameter matching checks whether the parameter values ​​and states mentioned in the text are consistent with the comparison library. For example, if the CPU temperature in the comparison library is 85℃, and the text states that the CPU temperature is 90℃, then the parameter value is determined to be mismatched; association matching checks... The text is checked to see if the causal and parallel relationships described in the text are consistent with the reference library. For example, if the reference library states that a low cooling fan speed is the cause of a high CPU temperature, and the text states that a high CPU temperature causes a low cooling fan speed, then the relationship is considered reversed. Time sequence matching checks if the collection time order of parameters in the text is consistent with the reference library. For example, if the reference library states that cooling fan speed data was collected at 10:00 and CPU temperature data at 10:05, and the text states that CPU temperature was collected at 10:00 and cooling fan speed at 10:05, then the time sequence is considered disordered. The number of successfully matched reference items is counted. If the matching pass rate reaches a preset standard, which is set based on the accuracy requirements of industrial equipment reports, then the initially generated text is considered logically consistent and is directly used as the verified text. If there are mismatches, such as abnormal motor vibration frequency causing an over-temperature warning, while the reference library states that a high CPU temperature causes an over-temperature warning, then step 502 is traced back to check if the relationship of the chapter-level semantic framework is correct. After correction, the initial text is regenerated and verified again until the logic is consistent.

[0098] Text optimization involves two main aspects: vocabulary selection and sentence structure adjustment. Vocabulary selection involves establishing a standard industrial equipment terminology library, including 2000 standard terms covering areas such as vibration frequency, over-temperature warning, rated values, and energy consumption values. After scanning and verifying the text, non-standard expressions are replaced. For example, "CPU temperature too high" is changed to "CPU temperature abnormal 85℃," "fan spinning slowly" is changed to "cooling fan speed below rated value," and "alarm twice" is changed to "triggered over-temperature warning twice," ensuring professional and consistent terminology. Sentence structure adjustment, tailored to the reading habits of maintenance personnel, breaks down complex compound sentences into concise short sentences to avoid information confusion caused by lengthy sentences. For example, "M..." The cooling fan speed of device 001 was lower than the rated value, which caused the CPU temperature to rise to 85℃ at 10:05 on September 30, 2025, triggering two over-temperature warnings. This was broken down into the cooling fan speed of device M001 being lower than the rated value; as a result, the CPU temperature rose to 85℃ at 10:05 on September 30, 2025, triggering two over-temperature warnings. Simultaneously, the word order was adjusted, placing the parameter value and the collection time at the beginning. For example, instead of saying the CPU temperature reached 85℃ at 10:05, it was changed to "The CPU temperature was 85℃ at 10:05 on September 30, 2025," highlighting the core parameter information and resulting in the optimized text.

[0099] Three core evaluation dimensions are established, each with clearly defined evaluation criteria. Information completeness assesses whether the text covers all core information within the text-level semantic framework, including equipment identification, all components and corresponding parameters, parameter values, states, or actions, data collection time, and cross-sentence relationships. If a core parameter, such as motor vibration frequency, is missing, that dimension is deemed unqualified. Logical correctness assesses whether the parameter relationships and temporal sequence in the text are completely consistent with the verification and comparison library, without issues such as causal reversal, parameter mismatch, or temporal disorder. If any logical error exists, that dimension is deemed unqualified. Terminology accuracy assesses whether the industrial terminology in the text fully conforms to the terminology standard library, without non-standard expressions or terminology misuse, such as changing "over-temperature warning" to "high-temperature alarm." If terminology errors exist, that dimension is deemed unqualified. The optimized text is scored independently for each of these criteria using a 100-point scale. If the average score of all three dimensions reaches a preset passing score (set based on the requirements of maintenance reports), the evaluation is considered passed, and the optimized text becomes the target natural language text.

[0100] This embodiment effectively avoids problems affecting accuracy, such as parameter mismatch and causal reversal in the initial text, through semantic logic verification; vocabulary and sentence structure optimization improve the readability and professionalism of the text; and the final quality assessment controls the text quality from multiple dimensions to ensure that the target natural language text can completely and accurately convey the device's operating status and parameter correlation.

[0101] like Figure 2As shown, embodiments of the present invention also provide an intelligent processing system for structured data to text based on natural language, including:

[0102] The extraction module is used to build and train a neural network model; the trained neural network model is then used to perform lexical analysis and feature extraction on the natural language text to be processed, resulting in a set of lexical feature vectors.

[0103] The partitioning module is used to construct a multidimensional feature space based on the set of lexical feature vectors; based on the multidimensional feature space, it divides the feature subspaces, analyzes the statistical characteristics of the feature subspaces, and generates analysis and optimization parameters;

[0104] The optimization module is used to optimize and adjust the text chunking analysis process according to the analysis optimization parameters, segmenting sentences into non-overlapping phrases with syntactic function tags to obtain the chunked text;

[0105] The parsing module is used to perform syntactic and semantic structure parsing on the text after chunking analysis. Through component analysis, dependency analysis and semantic dependency graph analysis, it establishes semantic relationships between the internal structures of sentences and obtains structured semantic information.

[0106] The analysis module is used to perform multi-sentence logical association analysis on structured semantic information at the discourse level to obtain the semantic information of the entire discourse.

[0107] The generation module is used to obtain the target natural language text based on the semantic information of the entire text and using a pre-trained large language model.

[0108] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0109] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0110] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0111] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An intelligent processing method for structured data to text based on natural language, characterized in that: The method includes: Step 1: Build a neural network model and train it; use the trained neural network model to perform lexical analysis and feature extraction on the natural language text to be processed, and obtain a set of lexical feature vectors. Step 2: Construct a multi-dimensional feature space based on the set of lexical feature vectors; divide the multi-dimensional feature space into feature subspaces, analyze the statistical characteristics of the feature subspaces, and generate analysis optimization parameters, including: integrating all lexical feature vectors in the lexical feature vector set into a feature matrix to obtain the multi-dimensional feature space; using a clustering algorithm on the multi-dimensional feature space to divide semantically similar lexical feature vectors into the same region to obtain multiple feature subspaces; performing statistical analysis on each feature subspace, calculating the vector distribution characteristics in each feature subspace, and obtaining the statistical characteristics of each feature subspace; determining the boundary judgment conditions and phrase division rules for chunking analysis based on the statistical characteristics of each feature subspace, and obtaining analysis optimization parameters. Step 3: Optimize and adjust the text chunking analysis process according to the analysis and optimization parameters, segment the sentence into non-overlapping phrases with syntactic function tags, and obtain the chunked text; among which, the syntactic function tags include: device identification phrase tags, device component phrase tags, parameter value phrase tags, and action phrase tags; Step 4 involves parsing the syntactic and semantic structures of the chunked text. Through component analysis, dependency analysis, and semantic dependency graph analysis, semantic relationships between sentence structures are established to obtain structured semantic information. This includes: performing component analysis on the chunked text to identify subject-verb-object grammatical components and obtain a phrase structure tree; performing dependency analysis on the phrase structure tree to obtain a dependency relation set; performing semantic role labeling on the phrase structure tree and dependency relation set to obtain semantic role labeling results; performing semantic association processing between the semantic role labeling results and the lexical feature vector set to construct semantic dependencies between words and obtain a semantic relation network; performing association and fusion processing on the semantic relation network and dependency relation set to establish a correspondence between grammatical structures and semantic relations, obtaining complete semantic relationships; and integrating the phrase structure tree, dependency relation set, and complete semantic relationships to obtain structured semantic information. The semantic roles include: device identification role, component role, and numerical role. Step 5: Perform multi-sentence logical association analysis on the structured semantic information at the discourse level to obtain the semantic information of the entire discourse; Step 6: Based on the semantic information of the entire text, obtain the target natural language text using a pre-trained large language model.

2. The intelligent processing method for structured data to text based on natural language according to claim 1, characterized in that, Construct and train a neural network model; use the trained neural network model to perform lexical analysis and feature extraction on the natural language text to be processed, obtaining a set of lexical feature vectors, including: An initial neural network model is obtained by constructing a neural network model containing an input layer, multiple hidden layers, and an output layer; natural language text samples and lexical annotation information are input into the initial neural network model, and the model parameters are optimized using the backpropagation algorithm to obtain a trained neural network model. The system receives structured data from external sources and converts it into natural language text to obtain the natural language text to be processed. The natural language text to be processed is then input into the input layer of the trained neural network model to obtain the original word sequence. The original word sequence is processed by lexical analysis through the hidden layer of the neural network model to obtain the lexical analysis result; the lexical analysis result is then processed by feature extraction through the output layer of the neural network model to obtain a set of lexical feature vectors.

3. The intelligent processing method for structured data to text based on natural language according to claim 2, characterized in that, The text chunking analysis process is optimized and adjusted based on the analysis and optimization parameters. Sentences are segmented into non-overlapping phrases with syntactic function labels, resulting in the chunked text, including: Based on the boundary judgment conditions in the analysis and optimization parameters, text segmentation rules are generated to obtain optimized text segmentation rules; the natural language text to be processed is initially segmented according to the optimized text segmentation rules to obtain a sequence of sentence units; The sentence unit sequence is matched with the vocabulary feature vector set to obtain a preliminary phrase boundary marker sequence. The phrase segmentation rules in the analysis and optimization parameters are then used for boundary verification to obtain the verified phrase boundary sequence. The sentence unit sequence is segmented based on the verified phrase boundary sequence to obtain non-overlapping phrase units. The non-overlapping phrase units are then matched with the lexical feature vector set for syntactic function to obtain phrases with syntactic function labels. Finally, the text is reorganized according to the original word order to obtain the text after chunking analysis.

4. The intelligent processing method for structured data to text based on natural language according to claim 3, characterized in that, At the discourse level, multi-sentence logical association analysis is performed on structured semantic information to obtain the semantic information of the entire discourse, including: The logical relationships between sentences are identified using structured semantic information to obtain a set of logical relationships between sentences. This set is then matched with a phrase structure tree to obtain a discourse-level grammatical structure that contains both logical relationships and retains the original word order. The discourse-level grammatical structure is then analyzed with a set of dependency relations to obtain discourse-level dependency relations. By fusing text-level dependency relations with semantic relation networks, cross-sentence semantic associations are obtained. Cross-sentence semantic associations are then integrated with complete semantic associations to establish a text-level semantic framework. Logical consistency and information integrity verification are performed. After successful verification, the text-level semantic framework is organized into a complete set of information including devices, components, parameters, semantic associations, logical relations, and acquisition timing, thus obtaining the semantic information of the entire text.

5. The intelligent processing method for structured data to text based on natural language according to claim 4, characterized in that, Based on the semantic information of the entire text, the target natural language text is obtained using a pre-trained large language model, including: The semantic information of the entire text is converted into a serialized semantic data; the serialized semantic data is then input into a pre-trained large language model for text generation to obtain the initial generated text. The initial generated text is subjected to semantic and logical consistency verification to obtain the verified text; the verified text is then optimized by adjusting the word selection and sentence structure to obtain the optimized text; finally, the optimized text is subjected to a final semantic quality assessment to obtain the target natural language text.

6. An intelligent data-to-text processing system based on natural language, the system implementing the method as described in any one of claims 1 to 5, characterized in that, include: The extraction module is used to build and train a neural network model; the trained neural network model is then used to perform lexical analysis and feature extraction on the natural language text to be processed, resulting in a set of lexical feature vectors. The partitioning module is used to construct a multidimensional feature space based on the set of lexical feature vectors. Based on the multidimensional feature space, feature subspaces are obtained, and the statistical properties of the feature subspaces are analyzed to generate analysis and optimization parameters. The optimization module is used to optimize and adjust the text chunking analysis process according to the analysis optimization parameters, segmenting sentences into non-overlapping phrases with syntactic function tags to obtain the chunked text; The parsing module is used to perform syntactic and semantic structure parsing on the text after chunking analysis. Through component analysis, dependency analysis and semantic dependency graph analysis, it establishes semantic relationships between the internal structures of sentences and obtains structured semantic information. The analysis module is used to perform multi-sentence logical association analysis on structured semantic information at the discourse level to obtain the semantic information of the entire discourse. The generation module is used to obtain the target natural language text based on the semantic information of the entire text and using a pre-trained large language model.

7. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 5.

Images