Method for extracting semantic information of complex equipment

Abstract

The invention relates to the technical field of industrial interconnection, in particular to a method for extracting semantic information of complex equipment. Comprising the following steps of 1, constructing a one-dimensional semantic information extraction model of the complex equipment; according to the method, a Times Block model is adopted, the Times Block converts data of complex equipment from a time domain to a frequency domain through fast Fourier transform, wavelet transform is added before Fourier transform is carried out on the Times Block model, and one-dimensional data semantic information is extracted; step 2, constructing a time sequence graph neural network model for extracting semantic information among sensors of the complex equipment; and step 3, constructing a multidimensional semantic information extraction model of the complex equipment on the basis of the step 1 and the step 2. The one-dimensional data semantic information extraction is realized by constructing the wavelet transform-based Times Block model, and experimental comparative analysis shows that the classification accuracy of the wavelet transform-integrated Times Block model is higher than that of other methods, so that the feature information contained in the data can be effectively captured.

Description

Technical Field

[0001] This invention relates to the field of industrial internet technology, specifically to a method for extracting semantic information from complex equipment. Background Technology

[0002] Various industries are facing a transformation from traditional communication technologies to intelligent, information-based, and digital communication methods. This transformation is crucial for improving transmission efficiency and meeting the demands of future massive data transmission. Currently, most industries primarily use traditional communication methods, which are proving inadequate in the face of the challenges of the big data era. The main problems include slow transmission speeds, low levels of automation, insufficient security, and an inability to adapt to the continuously evolving needs of future communications.

[0003] Semantic communication is a communication method that captures deep semantic information from raw data, compresses this information, and transmits it using semantic-level information. The biggest difference between semantic communication and traditional communication is that semantic communication is task-oriented, prioritizing understanding before transmission. Compared to traditional communication, this method significantly improves transmission efficiency and reliability. Therefore, researching a method for extracting semantic information from complex equipment is of great significance, as it is crucial for combining complex equipment with semantic communication, enabling complex equipment to take a significant step towards intelligentization.

[0004] The most crucial step in this process is extracting semantic information from the raw data, as the quality of this semantic information determines the efficiency of data transmission in a semantic communication system. The development of artificial intelligence technology in recent years has made it possible to extract high-quality semantic information from data. Deep learning, as the most important part of artificial intelligence, possesses excellent feature extraction capabilities, enabling it to learn and apply features from data.

[0005] The existing technology has the following drawbacks: 1. In the industrial sector, there is limited research on using deep learning to learn data features for feature extraction because industrial data is subject to noise and interference during transmission, which can easily lead to data loss and other problems, and the reliability of communication technologies is relatively low.

[0006] 2. In the process of extracting semantic information, traditional methods can only extract semantic information from single data and cannot obtain the correlation information between data from different sensors.

[0007] Therefore, how to efficiently and accurately extract semantic information from the raw data of complex equipment to solve the above problems is a problem that needs to be solved by those skilled in the art. Therefore, this application proposes a method for extracting semantic information from complex equipment. Summary of the Invention

[0008] In order to overcome the shortcomings of the existing technology and solve the technical problems existing in the background technology, the present invention proposes a method for extracting semantic information of complex equipment.

[0009] This invention is achieved through the following technical solution: A method for extracting semantic information from complex equipment includes the following steps: Step 1: Construct a one-dimensional semantic information extraction model for complex equipment; The Times Block model is adopted. Times Block uses Fast Fourier Transform to transform the data of complex equipment from the time domain to the frequency domain. Wavelet transform is added before the Fourier transform in the Times Block model to extract semantic information of one-dimensional data. Step 2: Construct a time-series graph neural network model for extracting semantic information between various sensors of complex equipment: Step 3: Based on Step 1 and Step 2, construct a multi-dimensional semantic information extraction model for complex equipment.

[0010] Preferably, the wavelet basis function of the wavelet transform is the Symlet wavelet, and the wavelet transform formula is discretized during the calculation and solution process.

[0011] Preferably, in step one, after obtaining the one-dimensional denoised frequency domain signal, periodic clipping and transpose are performed to obtain a two-dimensional tensor. After transformation, the Inception module is used to process it, and then the learned two-dimensional tensor is converted back into the one-dimensional information space.

[0012] Preferably, the specific steps of step two are as follows: Graph construction; Construction of a semantic extraction model among sensors in complex equipment; Sequence graph attention network is used to obtain semantic information and relationships between various parts of complex equipment, and node features are transformed into vector space. Then, graph attention mechanism is used to extract feature vectors between nodes.

[0013] Preferably, the central node uses a softmax operation to obtain the attention coefficient of each node from the feature vectors of its neighboring nodes. After obtaining the attention coefficient, it performs a weighted summation to obtain the relevant features of the neighboring nodes, resulting in aggregated features. The attention mechanism then obtains the temporal relationship between these features.

[0014] Preferably, step two further includes computation using a multi-head attention mechanism. The multi-head attention mechanism contains multiple attention heads, each of which learns different semantic information. This information is then integrated and computed, and the results are concatenated together using a matrix. The above computation process constitutes a graph attention layer. The graph attention layer performs self-attention computation based on the node feature set of the data input, and then distributes attention from the node to adjacent nodes through a Mask Attention operation. The attention value of a node to its adjacent nodes is obtained through a feedforward network function.

[0015] Preferably, it includes fusing a semantic information extraction model of individual sensor information of complex equipment with a semantic information extraction model of the relationship between sensors of complex equipment.

[0016] Preferably, in step three, for the data of each node, it first needs to go through the Times Block module, which obtains the semantic information of each node. After capturing the semantic information of each node, the attention mechanism is used to obtain the spatial and temporal association of the nodes. After the graph attention network is calculated, a graph with spatiotemporal correlation is obtained. The semantic information of the nodes in the graph includes the spatiotemporal association with its neighboring nodes, and the information of other nodes in the graph is obtained based on it.

[0017] The beneficial effects of this invention are: 1. This invention extracts semantic information from one-dimensional data by constructing a Times Block model based on wavelet transform. Experimental comparative analysis shows that the classification accuracy of the Times Block model incorporating wavelet transform is higher than that of other methods, and it can effectively capture the feature information contained in the data.

[0018] 2. This invention abstracts the various sensors of complex equipment into a graph, with each sensor as a node. It learns the relationships between different parts of the complex equipment through a graph attention network, thereby obtaining the overall data features of the complex equipment. This solves the problem that traditional methods can only extract semantic information from single data and cannot obtain the correlation information between data from different sensors.

[0019] 3. This invention constructs a multi-dimensional semantic information extraction model for complex equipment by integrating the previous two methods, thereby realizing the extraction of multi-dimensional semantic information of complex equipment. Through experimental comparison and analysis, it can be seen that the multi-dimensional semantic information extraction model for complex equipment can obtain high-quality data features, which is beneficial to the subsequent establishment of semantic knowledge base and semantic communication model. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the semantic information extraction method of the present invention; Figure 2This is a comparison chart of the effects of single FFT and FWT+FFT in this invention; Figure 3 This is a comparison diagram of wavelet transform in this invention with other denoising methods; Figure 4 This is a schematic diagram illustrating the GAT network feature extraction principle of the present invention; Figure 5 This is a schematic diagram illustrating the principle of multidimensional semantic information extraction for complex equipment in this invention. Detailed Implementation

[0021] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Experimental methods in the following embodiments, unless otherwise specified, are generally performed under conventional conditions or as recommended by the manufacturer.

[0022] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of skill in the art. The reagents and raw materials used in this invention are readily available through conventional means, and unless otherwise specified, they shall be used in accordance with conventional methods in the art or as per the product instructions. Furthermore, any methods and materials similar to or equivalent to those described herein may be applied to the methods of this invention. The invention will now be further described with reference to the accompanying drawings and specific embodiments. The preferred embodiments and materials described herein are for illustrative purposes only.

[0023] Example 1: A method for extracting semantic information from complex equipment, such as Figure 1 As shown, it includes the following steps: Step 1: Construct a one-dimensional semantic information extraction model for complex equipment; In the original method, Times Block uses Fast Fourier Transform to transform the data of complex equipment from the time domain to the frequency domain, and sorts it in descending order of frequency, taking the top K frequency bands as features, denoted as: (1) In the formula, For the calculation of the amplitude after Fourier transform, This indicates that the average value of the calculated results is taken. In the formula for: (2) In equation (2) for: (3) in For periodicity; The above three formulas can be summarized as follows: (4) However, the working environment of the equipment and the vibration of the equipment itself cause a lot of noise to be mixed in the transmitted signal. Since there are no cases of the above algorithm being applied in industrial scenarios, it is impossible to judge its effect in industrial scenarios. On the other hand, the Fourier transform used in the algorithm is not specifically designed for noise reduction. Its role in the algorithm is to convert the time-domain signal into the frequency-domain signal. Therefore, the Fourier transform is insufficient for processing signals with a lot of noise, especially for data generated in harsh working environments, such as equipment data, where its noise reduction capability is even more inadequate.

[0024] To address the aforementioned issues, the Wavelet Transform (WT) algorithm is introduced. Compared to other noise reduction methods, the Wavelet Transform offers advantages such as time-frequency locality, multi-scale analysis, adaptability, and fast computation. Furthermore, it can remove noise while accurately describing the time-frequency domain characteristics of the signal, depending on the selected parameters, making it suitable for complex equipment operating scenarios.

[0025] Wavelet transform is a time-frequency domain analysis method with a fixed window area. It obtains different wavelet coefficients by translating and scaling the mother wavelet.

[0026] In the space formed by square integrable functions in,signal It should meet the following requirements: (5) The wavelet basis function is defined as: (6) In the formula satisfy ; Scale factor; This represents the translation amount.

[0027] The wavelet transform for continuous signals is: (7) The inverse wavelet transform is: (8) in .

[0028] In actual communication tasks, the data processed is often digital signals rather than continuous signals, and directly calculating and solving the above formulas is very computationally expensive. Therefore, it is usually discretized to improve the calculation speed and also to facilitate the storage of data after transmission.

[0029] The formula for the discretized wavelet transform is as follows: (9) Its inverse transformation is: (10) The wavelet transforms described above depend on the choice of wavelet basis functions, which are the foundation of the wavelet transform. The Daubechies wavelet, also known as "dbN" where N represents the order, is a class of orthogonal wavelets with excellent time-frequency locality and performs well in signal compression and denoising. The Symlet wavelet is similar to the Daubechies wavelet, but is smoother at higher orders. This characteristic allows it to preserve some high-frequency signals without excessive compression.

[0030] Based on the analysis, considering the working scenarios and data composition of complex equipment, Symlet wavelets are more suitable for application scenarios of complex equipment.

[0031] Adding a wavelet transform before the Fourier transform can remove noise from the signal, making the signal clearer after the Fourier transform and laying the foundation for subsequent convolutional feature extraction operations. Let: (11) Then, after the discrete transformation, equation (11) becomes: (12) The signal after discrete wavelet transform is denoted as... Then equation (1) becomes: (13) In the formula, Record .

[0032] like Figure 2 As shown, for the same data, there is still some noise in the signal under a single Fourier transform; however, by adding wavelet transform before Fourier transform for signal processing, and then performing Fourier transform, it can be seen that the signal under this processing has less noise and is smoother.

[0033] like Figure 3 The graph shows a comparison of the denoising performance of wavelet transform with two commonly used denoising methods, median filtering and mean filtering, under the same signal conditions. As can be seen from the graph, under the same noise conditions, the wavelet transform signal is closer to the original signal after noise removal. Therefore, it can be concluded that the denoising capability of wavelet transform is significantly higher than that of the other two methods.

[0034] Obtain a one-dimensional denoised frequency domain signal Then, it is periodically pruned and transposed to obtain a two-dimensional tensor. The specific process is as follows:

[0035] (14) In the formula, It is the first The transformed two-dimensional tensor is processed using the Inception module, and then the learned two-dimensional tensor is... Transform back into one-dimensional information space , in which The length is The padding sequence is truncated to the original length T.

[0036] Step 2: Construct a time-series graph neural network model for extracting semantic information between various sensors of complex equipment: Specifically, the following steps are included: Graph construction; set up Indicates having An irregular time series dataset of labeled samples.

[0037] Each sample It has corresponding tags Irregular multivariate time series, where irregularity refers to data collected at different times rather than at the same time, for example, in The equipment's rotation speed was collected in real time. Rotation speed and temperature were collected at all times. Rotational speed and tension were collected at all times. Although each sensor collected data at the same interval, the information transmitted by the equipment as a whole was relatively irregular.

[0038] In the sample Includes A sensor for non-uniform measurement, denoted as .

[0039] For the sample Sensors in Representing a single observation as , Indicates in At any moment, the sensor The value is .

[0040] sensor or The set of at least one timestamp in the L-hop neighbors is denoted as L is the message passing layer.

[0041] In the dataset In, each sample It contains multiple complex equipment sensors. In the constructed map In the diagram, nodes represent sensors, and edges represent the relationships between sensors. (The diagram shows the sensors...) To the sensor The weight of a directed edge is denoted as Nodes in the diagram Initiate message passing at the location, through calculation sensors By understanding the relationship between sensors and their neighboring sensors, information about neighboring nodes can be obtained. (In the sensor...) Record sample The value is denoted as Then in the diagram Spreading from the middle, along the edge from To neighboring sensors, sensor embeddings are generated by aggregating the observations from each sensor. Finally, based on the sensor embeddings, each... Embedded Samples .

[0042] Constructed graph In this context, it can be referred to as a sample. Dependency graph, nodes For inferred sensors, edge For samples The dependencies between sensors.

[0043] From the sensor arrive The edge is denoted as a triplet ,in This indicates the strength of the relationship between two sensors. Describes the sensor and The relationship between them, when When data is collected, it will follow the edge Towards Send a message. If If so, then the two sensors are uncorrelated. Assume... The impact of is different right The influence of this means that the graph can be viewed as a directed graph, i.e. All graphs are initialized as fully connected graphs (for any...). , and , The formula for updating edge weights is: (15) in Indicates that the message comes from Passed to A collection of all timestamps; set during graph structure initialization. .

[0044] Construction of a semantic extraction model among sensors in complex equipment; For the constructed graph, Temporal Graph Attention Networks (TGAT) first transforms the node features into a vector space, and then uses Graph Attention Networks (GAT) to extract the feature vectors between nodes. The principle diagram of GAT is shown below. Figure 4 As shown.

[0045] Because sensors in different parts of complex equipment have varying effects on other sensors, different weights need to be assigned to these sensor nodes when aggregating them. Graph attention networks can capture the spatial correlation between nodes. In the graph, the central node... For adjacent nodes and The attention coefficient of each node is obtained by using the softmax operation on the feature vector. The calculation process is as follows: (16) in, (17) Indicates belonging to the first The feature vector of each node; and These are relevant parameters that can be obtained through learning; among them This represents the feature dimension. After obtaining the attention coefficients, they are weighted and summed to obtain the relevant features of neighboring nodes, resulting in the aggregated features. The nodes are then updated. The characteristics are denoted as: (18) Mish is the activation function.

[0046] After obtaining the node After identifying the characteristics of its neighboring nodes at a certain moment, it is necessary to obtain their temporal relationship.

[0047] Compared to traditional RNNs and other models, attention mechanisms, due to their unique computational method, can effectively capture temporal correlations. (In the graph...) Each node in the sequence generates a timestamped sequence, denoted as... .

[0048] Each sequence is processed using an attention mechanism to extract its temporal features. The calculation formula for the attention mechanism is as follows: (19) in , and The query vector, key vector, and value vector in the attention mechanism are calculated using the following formulas:

[0049]

[0050] (20) This is the corresponding parameter matrix. To capture a wider range of relevance features, a multi-head attention mechanism is used for computation. This mechanism involves multiple attention heads, each learning different semantic information. This information is then integrated and computed, and the results are concatenated into a matrix. The specific computation process is as follows: (twenty one) in: (twenty two) The above calculation process constitutes the graph attention layer, in the graph In the graph, the attention layer performs self-attention calculation based on the node feature set of the input data, and then uses the Mask Attention operation to shift attention from the nodes. Assigned to adjacent nodes The above is obtained through the feedforward network function. right Attention value.

[0051] Step 3: Based on Step 1 and Step 2, construct a multi-dimensional semantic information extraction model for complex equipment.

[0052] The semantic information acquired from a single sensor is used as the node input to the graph neural network model. Since the TimesBlock model can acquire the essential information of the data, this information is highly representative of the nodes in the graph neural network and provides a data foundation for calculating the relationships between nodes in the graph.

[0053] The principle diagram of the multi-dimensional semantic information extraction of complex equipment constructed in this invention is as follows: Figure 5 As shown.

[0054] like Figure 5 As shown, Time to Time Node Its neighboring nodes and Due to time differences in data transmission, the data from each node first needs to pass through the Times Block module, which acquires the semantic information of each node. Thus, for each node, there is one and only one set of semantic information belonging to that node. This semantic information includes not only the information from the sensor-collected data but also the node's temporal information. After capturing the semantic information of each node, for the graph... To obtain semantic information, it is necessary to calculate the relationships between nodes. This semantic information is then processed through a graph attention network. As mentioned above, the semantic information carried by each node has representational meaning, so it is relatively easy to obtain the spatial and temporal relationships between nodes using the attention mechanism.

[0055] After computation using the graph attention network, we can obtain the following: Figure 5 The diagram on the right shows a spatiotemporal correlation. In the diagram Middle node The semantic information includes its neighboring nodes. and The spatiotemporal connections allow us to obtain information about other nodes in the graph.

[0056] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A method for extracting semantic information from complex equipment, characterized in that, Includes the following steps: Step 1: Construct a one-dimensional semantic information extraction model for complex equipment; The Times Block model is adopted. Times Block uses Fast Fourier Transform to transform the data of complex equipment from the time domain to the frequency domain. Wavelet transform is added before the Fourier transform in the Times Block model to extract semantic information of one-dimensional data. Step 2: Construct a time-series graph neural network model for extracting semantic information between various sensors of complex equipment: Step 3: Based on Step 1 and Step 2, construct a multi-dimensional semantic information extraction model for complex equipment.

2. The method for extracting semantic information from complex equipment according to claim 1, characterized in that, The wavelet basis function of the wavelet transform is the Symlet wavelet, and the wavelet transform formula is discretized during the calculation and solution process.

3. The method for extracting semantic information from complex equipment according to claim 2, characterized in that, In step one, after obtaining the one-dimensional denoised frequency domain signal, periodic clipping and transpose are performed to obtain a two-dimensional tensor. After transformation, the Inception module is used to process it, and then the learned two-dimensional tensor is converted back into the one-dimensional information space.

4. The method for extracting semantic information from complex equipment according to claim 1, characterized in that, The specific steps for step two are as follows: Graph construction; Construction of a semantic extraction model among sensors in complex equipment; Sequence graph attention network is used to obtain semantic information and relationships between various parts of complex equipment, and node features are transformed into vector space. Then, graph attention mechanism is used to extract feature vectors between nodes.

5. The method for extracting semantic information from complex equipment according to claim 4, characterized in that, The central node uses a softmax operation to obtain the attention coefficient of each node based on the feature vectors of its neighboring nodes. After obtaining the attention coefficient, it performs a weighted summation to obtain the relevant features of the neighboring nodes, resulting in aggregated features. The attention mechanism then obtains the temporal relationship between these features.

6. The method for extracting semantic information from complex equipment according to claim 4, characterized in that, Step two also includes computation using a multi-head attention mechanism. The multi-head attention mechanism contains multiple attention heads, each of which learns different semantic information. This information is then integrated and computed, and the results are concatenated together using a matrix. The above computation process constitutes a graph attention layer. The graph attention layer performs self-attention computation based on the node feature set of the data input, and then distributes attention from the node to neighboring nodes through a mask attention operation. The attention value of a node to its neighboring nodes is obtained through a feedforward network function.

7. The method for extracting semantic information from complex equipment according to claim 6, characterized in that, This includes fusing semantic information extraction models of individual sensor information in complex equipment with semantic information extraction models of relationships between sensors in complex equipment.

8. The method for extracting semantic information from complex equipment according to claim 6, characterized in that, In step three, the data for each node first needs to go through the Times Block module to obtain the semantic information of each node. After capturing the semantic information of each node, the attention mechanism is used to obtain the spatial and temporal associations of the nodes. After the graph attention network is calculated, a graph with spatiotemporal correlations is obtained. The semantic information of the nodes in the graph includes the spatiotemporal associations with their neighboring nodes. Based on this, the information of other nodes in the graph is obtained.

Images