Industrial semantic communication method and system based on adaptive deep compressed sensing

Abstract

The invention relates to an industrial semantic communication method and system based on adaptive deep compressed sensing, and the method comprises the steps: carrying out the preprocessing of a data set, and dividing the preprocessed data set into a training set and a test set; the data set comprises machine tool cutter wear test spindle vibration data with different semantics; constructing an industrial semantic communication model; training the industrial semantic communication model to obtain a trained industrial semantic communication model; performing industrial semantic communication through the trained industrial semantic communication model; according to the invention, more efficient signal transmission and reconstruction are realized.

Description

Technical Field

[0001] The present invention relates to an industrial semantic communication method and system based on adaptive deep compressed sensing, belonging to the technical field of industrial semantic communication. Background Art

[0002] With the rise of the Fourth Industrial Revolution, the amount of data generated in the industrial sector is growing exponentially, and traditional communication methods are no longer able to effectively address this challenge. In large-scale industrial scenarios, the acquisition of massive amounts of real-time sensor data poses significant data transmission and processing challenges for resource-limited Industrial Internet of Things (IIoT) systems. Existing communication systems primarily focus on efficient symbol transmission and recovery, but these systems face significant transmission bottlenecks when faced with massive data streams. Therefore, to meet the demands of the Industrial Internet, there is an urgent need to develop more efficient transmission technologies to improve the communication performance of industrial networks.

[0003] With the rapid development and deep integration of sixth-generation mobile communications (6G) and artificial intelligence (AI) technologies, semantic communication, as an emerging communications paradigm, is demonstrating significant advantages. While the evolution of communications technology from 2G to 5G primarily focused on bit transmission, 6G intelligent interconnection places greater emphasis on information recognition, analysis, understanding, and reasoning. One of the core advantages of semantic communication technology lies in its tolerance for symbol errors during communication, enabling it to overcome the limitations of Shannon's law to some extent, thereby achieving superior communication performance. To address bandwidth bottlenecks, semantic communication proposes a new communications paradigm: a shift from "transmit first, understand later" to "understand first, transmit later." This approach focuses on the semantic layer of information, aiming to ensure effective message delivery and improve transmission efficiency.

[0004] In traditional wireless communication systems, source coding and channel coding are usually designed independently, making it difficult to optimize them together to achieve maximum communication capacity, and prone to the "cliff effect". Joint Source Channel Coding (JSCC) integrates source coding and channel coding to fully utilize channel characteristics and source information, reduce transmission error rates, and improve communication quality and system capacity. Most semantic communication systems are based on the JSCC architecture, which exhibits superior performance compared to traditional separate coding schemes. In today's information age, semantic communication is widely applicable to many fields, such as semantic-level transmission of multimodal data such as text, images, voice, and video.

[0005] When dealing with sparse or compressible signals, compressed sensing (CS) technology reduces the amount of sampled data by performing a sparse representation or compressible transformation on the signal, thereby saving storage space while still obtaining sufficient information. The advantage of deep compressed sensing theory lies in its ability to effectively capture the sparsity or compressibility of signals, enabling efficient data acquisition and transmission. This approach offers significant advantages in large-scale data processing, providing a more efficient and cost-effective sensing and data processing solution for information processing, communication systems, and other application areas. Therefore, the introduction of deep compressed sensing technology into JSCC systems is of great significance. By combining deep compressed sensing with JSCC, efficient signal compression and transmission can be achieved while improving the reliability of communication systems.

[0006] In industrial applications, sensor data is often continuous and real-time, requiring a complex data analysis environment to extract semantic information. However, industrial datasets suitable for training deep neural networks are often scarce. Therefore, there is an urgent need to develop semantic communication systems specifically for industrial data processing to support task-based transmission. Furthermore, due to the dynamic nature of wireless channels, the system must be able to adaptively adjust the signal-to-noise ratio (SNR). Therefore, an industrial semantic communication system based on adaptive deep compressed sensing technology should be designed to achieve efficient transmission of industrial data and improve the reliability and efficiency of data transmission.

[0007] In the Industrial Internet of Things (IIoT), traditional communication systems face significant challenges in coping with the surge in industrial data. While traditional communication methods focus on symbol transmission and recovery, they face significant transmission bottlenecks when processing large amounts of real-time sensor data, making them unable to meet the demands of the IIoT. Furthermore, existing communication systems lack the ability to process data semantics, making it difficult to achieve deep understanding and reasoning about information. This limits further improvements in communication performance and, in turn, impacts the efficiency of industrial data transmission, processing, and analysis.

[0008] Future Industrial IoT communication systems must not only focus on data transmission efficiency and accuracy but also on the semantic level of data—deeply exploring the underlying meaning and value of the data. An efficient industrial communication system must be able to deeply understand the semantics of industrial data to enable higher-level information utilization, such as intelligent decision-making and predictive maintenance. Therefore, one direction for future communication system development is the transition to semantic communication technology. By deeper understanding and analyzing data, the efficiency and accuracy of data transmission and processing can be improved. The application of this technology will effectively overcome the limitations of existing communication systems and promote the intelligent and sustainable development of the Industrial IoT. Summary of the Invention

[0009] In view of the shortcomings of the existing technology, the present invention provides an industrial semantic communication method based on adaptive deep compressed sensing;

[0010] The present invention also provides an industrial semantic communication system based on adaptive deep compressed sensing.

[0011] The present invention aims to improve the transmission efficiency and reliability of industrial tool wear data.

[0012] Explanation of terms:

[0013] 1. Industrial Semantic Communications (ISC): This technology combines semantic information with communication technologies to improve the efficiency and intelligence of information transmission in the industrial sector. Unlike traditional communication methods, ISC not only transmits raw data but also focuses on the semantics and context of the data, thereby reducing redundant data and improving communication efficiency. It can support real-time decision-making, reduce latency, and optimize bandwidth usage, and is particularly suitable for scenarios such as smart manufacturing, the Industrial Internet of Things (IIoT), and robotic control. By transmitting critical information rather than redundant data, ISC improves device collaboration and automation, promoting the intelligent and efficient operation of industrial systems.

[0014] 2. Deep Compressive Sensing (DCS): is a method that combines deep learning and compressed sensing technology, aiming to efficiently sample and accurately reconstruct sparse signals. Compressed sensing achieves complete reconstruction of the signal through a small amount of sampling, while deep learning can optimize the sampling and reconstruction process by utilizing the nonlinear representation capabilities of neural networks. Deep compressed sensing automatically learns the sparse characteristics and reconstruction rules of the signal by training deep neural networks, greatly improving sampling efficiency and reconstruction accuracy. This technology is widely used in medical imaging (such as MRI, CT), remote sensing image processing, and wireless communications. It can maintain high-quality signal reconstruction under low sampling rates, significantly reducing acquisition time and storage requirements, while improving the intelligence level and computing efficiency of the system.

[0015] 3. Joint Source Channel Coding (JSCC): This combines source coding and channel coding to improve the efficiency and reliability of data transmission. Source coding compresses data, channel coding corrects errors during transmission, and JSCC optimizes both during transmission. Through this joint design, JSCC reduces redundancy, enhances interference resistance, and improves transmission efficiency in low signal-to-noise ratio environments. It is widely used in wireless communications, deep space communications, and video transmission, helping to improve bandwidth utilization and reduce bit error rates.

[0016] 4. SNR (Signal-to-Noise Ratio): This is a metric that measures the ratio of signal strength to noise strength and is commonly used to assess the quality of communication systems. Higher SNR values ​​indicate a stronger signal relative to noise, resulting in better transmission quality and lower bit error rates. SNR is typically expressed in decibels (dB) and is calculated as the logarithm of the ratio of signal power to noise power. A high SNR helps improve system performance and ensure accurate data transmission. A low SNR can lead to signal distortion and bit errors, compromising communication quality.

[0017] In order to achieve the above object, the technical solution of the present invention is:

[0018] An industrial semantic communication method based on adaptive deep compressed sensing, comprising:

[0019] The data set is preprocessed and divided into a training set and a test set; the data set includes spindle vibration data of machine tool tool wear tests with different semantics;

[0020] Build an industrial semantic communication model;

[0021] Training the industrial semantic communication model to obtain a trained industrial semantic communication model;

[0022] Conduct industrial semantic communication through trained industrial semantic communication models;

[0023] The industrial semantic communication model includes: industrial semantic recognition module, SNR adaptive deep compressed sensing encoder module, transmission module, SNR adaptive deep compressed sensing decoder module;

[0024] The industrial semantic recognition module extracts industrial semantic information from industrial tool wear data. The industrial semantic information is divided into two categories: important semantic information and non-important semantic information. The important semantic information includes the three key stages of initial wear, stable wear, and accelerated wear; the non-important semantic information includes the idling state.

[0025] The SNR-adaptive deep compressed sensing encoder module compresses key semantic information and extracts industrial semantic features;

[0026] In the transmission module, the extracted industrial semantic features are transmitted to the SNR adaptive deep compressed sensing decoder module through the noisy physical channel industrial semantics;

[0027] The SNR-adaptive deep compressed sensing decoder module decodes and reconstructs the received industrial semantic features to recover the fine semantic information in the original data.

[0028] According to a preferred embodiment of the present invention, the industrial semantic information in the industrial tool wear data is extracted; comprising:

[0029] The industrial semantic recognition module includes a residual module, a multi-head self-attention module, a fully connected layer, and an output layer;

[0030] The input data of the industrial semantic recognition module is the preprocessed vibration signal collected by the three-axis accelerometer, including vibration signals in the X, Y, and Z directions. The residual module includes an initial convolution layer, a multi-level convolution layer, and a residual block. The initial convolution layer and the multi-level convolution layer both include a convolution layer, a batch normalization layer, and a ReLU activation function.

[0031] In the residual module, two convolution operations are performed through the initial convolution layer and the multi-level convolution layer, and the ReLU activation function is applied after each convolution operation;

[0032] After the initial convolutional layer processing, low-level features are extracted, and nonlinearity is introduced through the ReLU activation function; then, high-level features are extracted layer by layer through multi-level convolutional layers, and the feature extraction capability is enhanced through the residual block;

[0033] The features extracted by the industrial semantic recognition module are mapped to the multi-head self-attention module. Key, Query, and Value features are generated through convolution. The dot product of Query and Key is used to calculate the similarity score between features and normalized to generate attention weights. The attention weights are used to perform weighted summation on the Value features to generate an output feature map containing global context information. Multiple self-attention modules work in parallel, and the outputs of each attention head are finally fused into an overall feature map containing more comprehensive global semantic information, resulting in the multi-head self-attention output.

[0034] The processed multi-head self-attention output, i.e., the high-dimensional feature vector, is subjected to a convolution operation to further extract semantic information related to the task. A learnable parameter Gamma is used for weighted fusion to form the final feature representation. The fused feature map is input into the fully connected layer for semantic recognition and classification, corresponding to different semantic states: non-important semantic information and important semantic information.

[0035] Important semantic information is fed into the SNR-adaptive deep compressed sensing encoder module.

[0036] Further preferably, the preprocessing process of the vibration signal collected by the triaxial accelerometer includes: filtering and denoising, trend removal, normalization, outlier removal, resampling, signal smoothing and data alignment.

[0037] Preferably, according to the present invention, important semantic information is compressed and industrial semantic features are extracted; including:

[0038] The SNR adaptive deep compressed sensing encoder module includes a convolution module, a residual convolution block and an SNR adaptive module; the convolution module includes a one-dimensional convolution layer and a normalization layer;

[0039] Important semantic features are input into the convolution module, and the deep compressed sensing module compresses the important semantic features. This includes: first, the important semantic features pass through a one-dimensional convolution layer and a normalization layer; then, the semantic features are further enhanced through a residual convolution block;

[0040] In the SNR adaptive deep compressed sensing encoder, the important semantic information x k After mapping transformation, it is converted into industrial semantic features represented as y k , as shown below:

[0041] y k =f θ (x k ,γ),

[0042] Among them, x k is the important semantic information extracted by the industrial semantic recognition module, γ is the channel SNR, and f θ (·) represents the mapping function from original data to compressed data, θ is the training parameter of the SNR adaptive deep compressed sensing encoder, and y k is the encoded industrial semantic feature;

[0043] Subsequently, the encoded industrial semantic features are processed by the residual convolution block to generate new semantic features, which are then passed to the SNR adaptive module for processing. In the SNR adaptive module, global features F are generated through global pooling, and the global features F are connected with the current channel SNR = γ to form a new comprehensive feature F′. The scaling factor generated by the nonlinear deep neural network is then multiplied with the input features to obtain the channel environment adaptive features. The channel environment adaptive features are passed to the transmission module after continuous learning and adjustment.

[0044] Further preferably, it is assumed that the important semantic feature x output by the industrial semantic recognition module is expressed as:

[0045]

[0046] Where h×w represents the size of the feature, and c represents the number of channels. The important semantic feature x is compressed and extracted by the SNR-adaptive deep compressed sensing encoder module, and the compressed three-dimensional semantic feature data y is expressed as:

[0047]

[0048] According to the preferred embodiment of the present invention, in the SNR adaptive module, the semantic features output by the upper module are represented as As the input of the SNR adaptation module, represents the output features of the SNR adaptation module, where h×w is the size of each feature and c represents the feature size;

[0049] The global pooling method is used to pool the semantic features m G Perform downsampling of spatial dimensions;

[0050] First, the SNR adaptation module is used to adjust the feature m G Perform global average pooling processing; the details are as follows:

[0051] The global average pooling method M(·) is used to combine the input semantic features m G Processing is performed to extract the global feature F; for the input feature m G The element m in ij , the calculation formula is as follows:

[0052]

[0053] Among them, h×w represents the size of the feature, m ij is feature m G The elements in , F represents the global information feature;

[0054] The global information feature F is concatenated with the signal-to-noise ratio information γ to form a comprehensive feature representation F′, as shown below:

[0055]

[0056] Next, F′ is input into a nonlinear deep neural network to generate a scaling factor α, which is calculated as follows:

[0057]

[0058] in, Represent the PReLU layer and Sigmoid layer respectively, w1, w2, b1, b2 are the parameters of the linear layer;

[0059] Perform the scaling factor α and the input semantic feature m of the SNR adaptation module G Element-by-element multiplication of to realize the calculation of channel environment adaptive feature m′, based on input feature m G Re-adapt feature m′:

[0060] m′=α⊙m G ,

[0061] Finally, the comprehensive feature F′ is output after being processed by MLP and multiplied element by element with the adaptive feature m′ to obtain the final output feature m A,The process is expressed as:

[0062] m A =MLP(F′)⊙m′.

[0063] According to the preferred embodiment of the present invention, the encoded industrial semantic feature y k Transmitted through noisy physical channel industry semantics to the SNR adaptive deep compressed sensing decoder module; including:

[0064] The extracted industrial semantic features y k The signal is transmitted to the SNR adaptive deep compressed sensing decoder module through the additive white Gaussian noise channel, i.e., the noisy physical channel, as shown below:

[0065]

[0066] Among them, h∈C is the channel coefficient, n is the channel noise, and it obeys the complex Gaussian distribution n~CN(0,σ 2 ), is the noisy semantic feature received after passing through the noisy physical channel;

[0067] SNR is defined as follows:

[0068]

[0069] Among them, P m represents the average power of semantic features, P n Represents the noise power.

[0070] Furthermore, the received industrial semantic features are decoded and reconstructed to restore the fine semantic information in the original data; including:

[0071] The SNR adaptive deep compressed sensing decoder module includes a residual transposed convolution block, an SNR adaptive module, and a depth reconstruction block; the residual transposed convolution block and the SNR adaptive module are decoded, and the depth decompression module is used to perform deep reconstruction to obtain the final reconstructed data. The specific process is:

[0072] First, the received After the residual transposed convolution block, it is used to achieve feature upsampling and recovery, and the resolution of the feature map is restored to a higher level through the transposed convolution operation;

[0073] Then, the SNR adaptive module is entered. The SNR adaptive module introduces the attention mechanism to capture the feature m obtained by the SNR adaptive module. A is sent on a physical channel;

[0074] Finally, deep reconstruction is performed through the deep decompression module to obtain the final reconstructed data.

[0075] Further preferably, the final reconstructed data is expressed by the following formula:

[0076]

[0077] Where γ is the channel SNR, Implement the mapping of the received compressed data to the restored original data for the decoding function, are the training parameters of the SNR adaptive deep compressed sensing decoder module;

[0078] The relationship between the original relevant semantic information and the reconstructed semantic information is transformed into an optimization problem, which is expressed as:

[0079]

[0080] s.ty k =f θ (x k ,γ)

[0081] Among them, x k is important semantic information, γ is the channel SNR, f θ (·) is the encoding function that realizes the mapping from original data to compressed data, θ is the training parameter of the SNR adaptive deep compressed sensing encoder module, Implement the mapping of the received compressed data to the restored original data for the decoding function, are the training parameters of the SNR adaptive deep compressed sensing decoder module.

[0082] Preferably, according to the present invention, the industrial semantic communication model is trained to obtain a trained industrial semantic communication model; comprising:

[0083] In the industrial semantic communication model, the optimization goal is to minimize the loss function. For each training sample i, the input of the industrial semantic communication model is z i , the loss function L is expressed as:

[0084]

[0085] Where N is the number of samples in the training set, z i is the input data, x i is the processed output;

[0086] Train an optimal strategy and use the gradient descent algorithm to update the system parameters θ, Until the loss function converges, it can be expressed as:

[0087]

[0088] s.tyk =f θ (x k ,γ),

[0089] Input the test set to verify the industrial semantic communication model; obtain the trained industrial semantic communication model.

[0090] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of an industrial semantic communication method based on adaptive deep compressed sensing when executing the computer program.

[0091] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of an industrial semantic communication method based on adaptive deep compressed sensing.

[0092] An industrial semantic communication system based on adaptive deep compressed sensing, comprising:

[0093] The data acquisition and preprocessing module is configured to: preprocess the data set and divide the preprocessed data set into a training set and a test set;

[0094] The model building and training module is configured to: build an industrial semantic communication model; and train the industrial semantic communication model to obtain a trained industrial semantic communication model;

[0095] The industrial semantic communication module is configured to: perform industrial semantic communication through a trained industrial semantic communication model; including: extracting industrial semantic information from industrial tool wear data; industrial semantic information is divided into two categories: important semantic information and non-important semantic information; important semantic information includes three key stages: initial wear, stable wear and accelerated wear; non-important semantic information includes idling state; compress key semantic information and extract industrial semantic features; decode and reconstruct the received industrial semantic features to restore the fine semantic information in the original data.

[0096] The beneficial effects of the present invention are:

[0097] 1. This invention embeds a semantic recognition method at the transmitting end. By analyzing the semantic information of the data to be transmitted, it classifies it into important and unimportant semantic features based on their semantic importance. This classification method can provide optimization support for the subsequent encoding process, thereby enabling resource allocation priority adjustment, giving higher transmission priority to important semantic features, and improving the efficiency of the communication system and the accuracy of semantic information transmission.

[0098] 2. The present invention is based on training encoders and decoders for adaptive deep compressed sensing; the present invention effectively combines the advantages of compressed sensing and deep neural networks and integrates them into the joint source-channel coding (JSCC) architecture to achieve more efficient signal transmission and reconstruction. BRIEF DESCRIPTION OF THE DRAWINGS

[0099] In order to more clearly illustrate the technical solutions in the present disclosure or related technologies, the following briefly introduces the illustrations required for use in the implementation examples or related technical descriptions. Obviously, these illustrations are only applicable to the implementation examples of the present disclosure.

[0100] Figure 1 A schematic flow chart of an industrial semantic communication method based on adaptive deep compressed sensing provided by an embodiment of the present invention;

[0101] Figure 2 This is an overall system diagram of an industrial semantic communication system based on adaptive deep compressed sensing in the present invention;

[0102] Figure 3 Schematic diagram of a semantic recognition method for industrial scenarios in an industrial semantic communication system based on adaptive deep compressed sensing in the present invention;

[0103] Figure 4 Schematic diagram of an SNR adaptive deep compressed sensing encoder module of an industrial semantic communication system based on adaptive deep compressed sensing in the present invention;

[0104] Figure 5 A performance comparison chart of the present invention and other semantic communication systems; DETAILED DESCRIPTION

[0105] The following is a more detailed description of the specific implementation of the present invention in combination with the accompanying drawings and detailed embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. The embodiments or illustrations will be used to illustrate the present invention but will not limit its scope. Unless otherwise defined, the technical and scientific terms involved in the text will be interpreted according to the meaning commonly understood by ordinary technicians.

[0106] Example 1

[0107] An industrial semantic communication method based on adaptive deep compressed sensing, comprising:

[0108] The data set is preprocessed and divided into a training set and a test set; the data set includes spindle vibration data of machine tool tool wear tests with different semantics;

[0109] Build an industrial semantic communication model;

[0110] The industrial semantic communication model is trained to obtain a trained industrial semantic communication model. In order to train the industrial semantic recognition network, the cross entropy function is used as the objective function and the ADAM optimizer is used for optimization. During the training process, the gradient descent algorithm is used to update the system parameters θ, This process continues until the loss function converges. When the loss function converges, training stops and the corresponding parameter set is saved for later use. The trained industrial semantic communication model is verified using the test set, and the optimal parameter set is determined by analyzing the test results.

[0111] Conduct industrial semantic communication through trained industrial semantic communication models;

[0112] like Figure 2 As shown, the industrial semantic communication model includes: industrial semantic recognition module, SNR adaptive deep compressed sensing encoder module, transmission module, SNR adaptive deep compressed sensing decoder module;

[0113] The industrial semantic recognition module accurately extracts industrial semantic information from industrial tool wear data. This information is divided into two categories: important and unimportant. Important semantic information includes three key stages: initial wear, stable wear, and accelerated wear; these are crucial for monitoring and preventing tool wear. Unimportant semantic information includes idle state, which is not of substantial value for tool wear monitoring.

[0114] The SNR-adaptive deep compressed sensing encoder module deeply compresses key semantic information and extracts industrial semantic features;

[0115] In the transmission module, the extracted industrial semantic features are transmitted to the SNR adaptive deep compressed sensing decoder module through the noisy physical channel industrial semantics;

[0116] The SNR-adaptive deep compressed sensing decoder module effectively decodes and reconstructs the received industrial semantic features to recover the fine semantic information in the original data.

[0117] Example 2

[0118] According to the industrial semantic communication method based on adaptive deep compressed sensing described in Example 1, Figure 1 As shown, the difference is:

[0119] Combine Figure 3The processed data is then accurately input into the industrial semantic recognition module for in-depth analysis. The pre-processed industrial data is then input into the industrial semantic recognition module for semantic information recognition. The industrial semantic information extracted from the industrial tool wear data includes:

[0120] The industrial semantic recognition module includes a residual module, a multi-head self-attention module, a fully connected layer and an output layer. By combining the residual module and the multi-head self-attention module, the system can effectively identify and extract key semantic features in the industrial field, thereby helping to understand and analyze industrial tool wear data and improve the accuracy and efficiency of data processing.

[0121] The industrial semantic recognition module completes the entire process from input data to industrial semantic state recognition through convolutional neural networks and a self-attention mechanism. The input data for the industrial semantic recognition module is the vibration signal collected by a preprocessed three-axis accelerometer, including vibration signals in the X, Y, and Z directions. Using a highly sensitive and high-precision three-axis accelerometer, it comprehensively captures the tool wear vibration signals of the CNC machine tool spindle in the three key directions of X, Y, and Z at an extremely high sampling frequency. The residual module includes an initial convolutional layer, a multi-stage convolutional layer, and a residual block. Both the initial convolutional layer and the multi-stage convolutional layer include a convolutional layer, a batch normalization layer, and a ReLU activation function.

[0122] In the residual module, two convolution operations are performed through the initial convolution layer and the multi-level convolution layer, and the ReLU activation function is applied after each convolution operation; the input is directly added to the output of the multi-level convolution layer through the skip connection to form a skip connection; this can alleviate the gradient vanishing problem and improve the training stability of the model.

[0123] After the initial convolutional layer, low-level features are extracted. The initial convolutional layer consists of a convolution kernel and a ReLU activation function, which can efficiently extract the basic features of the signal and lay a solid foundation for subsequent analysis. The ReLU activation function also introduces nonlinearity. Subsequently, multi-level convolutional layers are used to extract high-level features layer by layer, and residual blocks are used to enhance feature extraction capabilities.

[0124] The features extracted by the industrial semantic recognition module are mapped to the multi-head self-attention module. Key, Query, and Value features are generated through convolution. The dot product of Query and Key is used to calculate the similarity score between features and normalized to generate attention weights. The attention weights are used to perform weighted summation on the Value features to generate an output feature map containing global context information. Multiple self-attention modules work in parallel, and the outputs of each attention head are finally fused into an overall feature map containing more comprehensive global semantic information, resulting in the multi-head self-attention output.

[0125] The processed multi-head self-attention output, i.e., the high-dimensional feature vector, is subjected to a convolution operation to further extract semantic information relevant to the task. A learnable parameter, Gamma, is used to weightedly fuse the two pieces of information to form the final feature representation. The fused feature map is then input into the fully connected layer for semantic recognition and classification, corresponding to different semantic states: non-important semantic information and important semantic information. This module combines the local feature extraction capabilities of the convolutional network with the global modeling capabilities of the attention mechanism, and can efficiently handle complex industrial semantic recognition tasks and accurately classify different industrial states. Important semantic information includes three key stages: initial wear, stable wear, and accelerated wear, which are crucial for monitoring and preventing tool wear. Non-important semantic information, on the other hand, mainly refers to the idling state, which has no substantial reference value for monitoring tool wear.

[0126] Considering the particularity of non-important semantic information in the transmission task, this information only needs to be sent to the edge server / cloud server in the form of extremely brief semantic tags without the need for compression and data transmission through the codec model.

[0127] Conversely, important semantic information needs to be input into the SNR-adaptive deep compressed sensing encoder module for compression and extraction of industrial semantic features. This encoder intelligently extracts industrial semantic features and minimizes data transmission while ensuring information quality.

[0128] An industrial semantic knowledge base is used to improve the accuracy and intelligence of tool wear detection. After data preprocessing and feature extraction, the vibration signal is input into the industrial semantic recognition module, which, combined with the knowledge base, performs in-depth analysis. The knowledge base stores tool wear patterns, operating conditions, and fault characteristics, enhancing the recognition of important semantic information such as initial wear, stable wear, and accelerated wear. The knowledge base then assists the system in semantic classification and optimizes data transmission strategies, enabling rapid processing of non-critical information while highly compressing and efficiently transmitting critical information to the edge or cloud. Ultimately, this knowledge base improves the accuracy of fault diagnosis, providing strong support for intelligent manufacturing and predictive maintenance.

[0129] The preprocessing process for vibration signals collected by triaxial accelerometers includes filtering and denoising, trend removal, normalization, outlier removal, resampling, signal smoothing, and data alignment to improve signal quality. Furthermore, time-domain, frequency-domain, and time-frequency-domain feature extraction techniques are combined to enhance data representation and provide reliable support for subsequent fault diagnosis.

[0130] Compress important semantic information and extract industrial semantic features; including:

[0131] like Figure 4 As shown in the figure, the SNR adaptive deep compressed sensing encoder module includes a convolution module, a residual convolution block, and an SNR adaptive module; it achieves efficient compression and extraction of industrial semantic features to achieve excellent performance and adaptability in various noise environments. The convolution module includes a one-dimensional convolution layer and a normalization layer;

[0132] Important semantic features are input into the convolution module, where the deep compressed sensing module efficiently compresses them. First, these important semantic features pass through a one-dimensional convolutional layer and a normalization layer. This step optimizes feature extraction and ensures effective information compression. Next, a residual convolution block further enhances the semantic features, improving the overall capability of joint source-channel coding.

[0133] In the SNR adaptive deep compressed sensing encoder, the important semantic information x k After the mapping transformation of the encoder, it is converted into industrial semantic features represented as y k , as shown below:

[0134] y k =f θ (x k ,γ),

[0135] Among them, x k is the important semantic information extracted by the industrial semantic recognition module, γ is the channel SNR, and f θ (·) represents the mapping function from original data to compressed data, θ is the training parameter of the SNR adaptive deep compressed sensing encoder, and y k is the encoded industrial semantic feature;

[0136] The encoded industrial semantic features are then processed through a residual convolution block to generate new semantic features, which are then passed to the SNR adaptation module for further processing. At this stage, the attention-based SNR adaptation module applies greater attention and refinement to the input features, ensuring that important semantic information is accurately captured and represented. Within the SNR adaptation module, important semantic information is passed through the residual convolution block to generate intermediate semantic features. This intermediate semantic feature is then passed to the attention-based SNR adaptation module, where global pooling is performed to generate a global feature F. This global feature F is then concatenated with the current channel SNR = γ to form new contextual information F′. A nonlinear deep neural network is then used to generate a scaling factor, which is multiplied by the input features to generate a channel-environment-adaptive feature. This feature is continuously learned and adjusted before being passed to the transmission module, providing a richer and more accurate information foundation for subsequent data processing and semantic analysis. This adaptive design enables the entire process to more accurately capture and mine the deep characteristics of key semantic information, significantly improving the efficiency of data processing and semantic analysis.

[0137] Assume that the important semantic feature x output by the industrial semantic recognition module is expressed as:

[0138]

[0139] Where h×w represents the size of the feature, and c represents the number of channels. To improve data transmission efficiency and reduce redundant information, the important semantic features x are compressed and extracted by the SNR (signal-to-noise ratio) adaptive deep compressed sensing encoder module. The compressed three-dimensional semantic feature data y is expressed as:

[0140]

[0141] During the compression process, convolutional layers are introduced to process important semantic features to optimize feature extraction and improve information compression efficiency. Convolutional layers capture local correlations and patterns between features, making the extracted features more compact and expressive. This process ensures that key information is effectively preserved while reducing data dimensionality, thereby improving the overall performance of the communication system.

[0142] Semantic features are further enhanced by introducing residual convolution blocks. Residual connections can help alleviate the vanishing or exploding gradient problems that occur when training deep neural networks, while also making the network easier to optimize. This enhancement helps improve the overall capabilities of joint source-channel coding, making the compressed features more stable and reliable during transmission.

[0143] It should be noted that these preliminarily processed features are passed to the SNR adaptive module based on the attention mechanism. In the SNR adaptive module, the focus on different features can be dynamically adjusted according to the channel state, and refined processing can be performed to ensure that important semantic information is accurately captured and efficiently represented. By introducing the attention mechanism, the SNR adaptive module adaptively adjusts the feature weights according to the changes in the signal-to-noise ratio (SNR), enabling the system to optimize the expression of features under different channel conditions;

[0144] In the SNR adaptive module, the semantic features output by the residual convolution block are expressed as As the input of the SNR adaptation module; represents the output features of the SNR adaptation module, where h×w is the size of each feature and c represents the feature size;

[0145] The global pooling method is used to pool the semantic features m GThis module downsamples the spatial dimensions, effectively reducing data dimensionality while retaining key semantic information. This enhances the global representation of features, making them more suitable for subsequent attention mechanism processing. This allows the model to dynamically adjust feature attention weights, highlighting the features most relevant to the current task. By strengthening the focus on key features, the model's expressiveness and generalization performance are significantly improved, thereby enhancing the overall stability and adaptability of the semantic communication system. This module is closely integrated with the signal-to-noise ratio (SNR), enabling the processed features to adapt to different channel environments, forming adaptive channel environment features.

[0146] First, the SNR adaptation module is used to adjust the feature m G Perform global average pooling; feature m G for The details are as follows:

[0147] The global average pooling method M(·) is used to combine the input semantic features m G Processing is performed to extract the global feature F; for the input feature m G The element m in ij , the calculation formula is as follows:

[0148]

[0149] Among them, h×w represents the size of the feature, m ij is feature m G The elements in , F represents the global information feature;

[0150] The global information feature F is concatenated with the signal-to-noise ratio (SNR) information γ to form a comprehensive feature representation F′. This operation aims to combine global feature information and channel state information to enhance the environmental adaptability of the feature. The concatenated intermediate features are as follows:

[0151]

[0152] Next, F′ is input into a nonlinear deep neural network to generate a scaling factor α, which is calculated as follows:

[0153]

[0154] in, Represent the PReLU layer and the Sigmoid layer respectively, w1, w2, b1, and b2 are the parameters of the linear layer; this process dynamically adjusts the scaling factor α through nonlinear transformation to adapt to different channel conditions, thereby optimizing the expressive power of the features.

[0155] Perform the scaling factor α and the input semantic feature m of the SNR adaptation module GElement-by-element multiplication of to realize the calculation of channel environment adaptive feature m′, based on input feature m G Re-adapt feature m′:

[0156] m′=α⊙m G ,

[0157] Finally, the comprehensive feature F′ is output after being processed by MLP and multiplied element by element with the adaptive feature m′ to obtain the final output feature m A ,The process can be expressed as:

[0158] m A =MLP(F′)⊙m′.

[0159] These processed features are encoded by the SNR adaptive deep compressed sensing encoder and expressed as:

[0160] y k =f θ (x k ,γ),

[0161] Among them, x k is the relevant semantic information output by the industrial semantic recognition network, γ is the channel SNR, and f θ (·) represents the mapping function from original data to compressed data, and θ is the model training parameter of the encoder.

[0162] The encoded industrial semantic features y k Transmitted through noisy physical channel industry semantics to the SNR adaptive deep compressed sensing decoder module; including:

[0163] The extracted industrial semantic features y k The signal is transmitted to the SNR adaptive deep compressed sensing decoder module through the additive white Gaussian noise (AWGN) channel, i.e., the noisy physical channel, as shown below:

[0164]

[0165] Among them, h∈C is the channel coefficient, n is the channel noise, and it obeys the complex Gaussian distribution n~CN(0,σ 2 ), is the noisy semantic feature received after passing through the noisy physical channel;

[0166] To measure the quality of the physical channel, SNR is defined as follows:

[0167]

[0168] Among them, P mrepresents the average power of semantic features, P n Represents the noise power.

[0169] Decode and reconstruct the received industrial semantic features to recover the fine semantic information in the original data; including:

[0170] The SNR adaptive deep compressed sensing decoder module includes a residual transposed convolution block, an SNR adaptive module, and a deep reconstruction block. Through the synergistic effect of these components, the SNR adaptive deep compressed sensing decoder module can accurately restore and present the important semantic information in the original data, thereby achieving effective decoding and reconstruction of the data. After decoding by the residual transposed convolution block and the SNR adaptive module, the deep decompression module performs deep reconstruction to obtain the final reconstructed data. The specific process is:

[0171] First, the received After the residual transposed convolution block, it is used to achieve feature upsampling and recovery, and the resolution of the feature map is restored to a higher level through the transposed convolution operation, thereby achieving feature refinement and reconstruction.

[0172] Then, the SNR adaptive module is introduced. The SNR adaptive module introduces an attention mechanism so that the model can focus on specific features or regions, thereby improving the importance of features and the performance of the model. The attention feature block helps the model better understand and learn the key information in the data. The feature m obtained by the SNR adaptive module is used as the A is sent on a physical channel;

[0173] Finally, deep reconstruction is performed in the deep decompression module; this module is typically used to reconstruct and restore the original data in the decoded feature space. Through deep reconstruction, the model learns from the compressed feature representation and reconstructs a result that is closer to the original data, thus achieving data restoration and recovery, and obtaining the final reconstructed data.

[0174] Through this series of steps, the model is able to obtain the final reconstructed data. This data undergoes a series of processing and reconstruction steps to preserve important information and restore the original data as accurately as possible, thus achieving an effective data compression and reconstruction process. Finally, the decoder receives the recovered data, and the relationship between the original relevant semantic information and the reconstructed semantic information can be transformed into an optimization problem.

[0175] The final reconstructed data is expressed by the following formula:

[0176]

[0177] Where γ is the channel SNR, Implement the mapping of the received compressed data to the restored original data for the decoding function, are the training parameters of the SNR adaptive deep compressed sensing decoder module;

[0178] The relationship between the original relevant semantic information and the reconstructed semantic information is transformed into an optimization problem, which is expressed as:

[0179]

[0180] s.ty k =f θ (x k ,γ)

[0181] Among them, x k is important semantic information, γ is the channel SNR, f θ (·) is the encoding function that realizes the mapping from original data to compressed data, θ is the training parameter of the SNR adaptive deep compressed sensing encoder module, Implement the mapping of the received compressed data to the restored original data for the decoding function, are the training parameters of the SNR adaptive deep compressed sensing decoder module.

[0182] Training the industrial semantic communication model to obtain a trained industrial semantic communication model; including:

[0183] In the industrial semantic communication model, the optimization goal is to minimize the loss function. For each training sample i, the input of the industrial semantic communication model is z i , the loss function L is expressed as:

[0184]

[0185] Where N is the number of samples in the training set, z i is the input data, x i is the processed output;

[0186] Train an optimal strategy and use the gradient descent algorithm to update the system parameters θ, Until the loss function converges, it can be expressed as:

[0187]

[0188] s.ty k =f θ (x k ,γ)

[0189] Input the test set to verify the industrial semantic communication model; the trained industrial semantic communication model is obtained. The details are as follows:

[0190] During the verification process, the performance of the model is evaluated by calculating the accuracy on the test set. Below, accuracy is used to evaluate the performance, and the formula is:

[0191]

[0192] where 1(·) is the exponential function. When the predicted value is equal to the true input x k it returns 1, otherwise it returns 0.

[0193] Combined with Figure 5 , under the AWGN channel, the performance comparison with the ADJSCC and DeepJSCC-V semantic communication systems. The simulation results show that under consistent experimental conditions, the present invention has achieved improvements under the vast majority of signal-to-noise ratio conditions. In particular, consider the performance tests under different SNRs in the AWGN channel condition. Among them, R = CR = k / n, where k < n represents the data dimension at the sending end, and k represents the input channel data dimension. Other models are also restricted under the same CR, and consider the performance tests under different SNRs in the AWGN channel condition.

[0194] Input the industrial semantic information and semantic markers restored at the receiving end into the industrial AI model for analysis and processing. This system can be used for industrial product quality monitoring and can display the monitoring results using an interactive interface.

[0195] Example 3

[0196] A computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the industrial semantic communication method based on adaptive deep compressive sensing described in Example 1 or 2.

[0197] Example 4

[0198] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the steps of the industrial semantic communication method based on adaptive deep compressive sensing described in Example 1 or 2.

[0199] Example 5

[0200] An industrial semantic communication system based on adaptive deep compressive sensing includes:[[ID=3D]]

[0201] A data acquisition and preprocessing module, configured to: preprocess the data set and divide the obtained preprocessed data set into a training set and a test set;

[0202] A model construction and training module, configured to: construct an industrial semantic communication model; and train the industrial semantic communication model to obtain a trained industrial semantic communication model;

[0203] The industrial semantic communication module is configured to perform industrial semantic communication using a trained industrial semantic communication model. This includes accurately extracting industrial semantic information from industrial tool wear data. Industrial semantic information is divided into two categories: important and unimportant. Important semantic information includes three key stages: initial wear, stable wear, and accelerated wear, which are crucial for tool wear monitoring and prevention. Unimportant semantic information includes idling, which is not of substantial value for tool wear monitoring. Key semantic information is deeply compressed and industrial semantic features are extracted. The received industrial semantic features are then efficiently decoded and reconstructed to recover the fine-grained semantic information in the original data.

[0204] A method for establishing an industrial semantic communication system based on adaptive deep compressed sensing includes the following steps:

[0205] S1: Analyze and process the data set, and divide the preprocessed data set into a training set and a test set;

[0206] S2: Establish an industrial semantic communication system based on adaptive deep compressed sensing;

[0207] S3: Initialize the industrial semantic communication system based on adaptive deep compressed sensing and train each part of the system using the processed training set. To train the industrial semantic recognition network, the cross entropy function is used as the objective function and optimized using the ADAM optimizer.

[0208] S4: During the training process, the gradient descent algorithm is used to update the system parameters θ, Until the loss function converges. When the loss function converges, the training will stop and the corresponding parameter set will be saved for subsequent use;

[0209] S5: Use the test set to verify the trained industrial semantic communication system and determine the optimal parameter set by analyzing the test results.

Claims

1. An industrial semantic communication method based on adaptive deep compressed sensing, characterized in that: include: The data set is preprocessed and divided into a training set and a test set; the data set includes spindle vibration data of machine tool tool wear tests with different semantics; Build an industrial semantic communication model; Training the industrial semantic communication model to obtain a trained industrial semantic communication model; Conduct industrial semantic communication through trained industrial semantic communication models; The industrial semantic communication model includes: industrial semantic recognition module, SNR adaptive deep compressed sensing encoder module, transmission module, SNR adaptive deep compressed sensing decoder module; The industrial semantic recognition module extracts industrial semantic information from industrial tool wear data. The industrial semantic information is divided into two categories: important semantic information and non-important semantic information. The important semantic information includes the three key stages of initial wear, stable wear, and accelerated wear; the non-important semantic information includes the idling state. The SNR-adaptive deep compressed sensing encoder module compresses key semantic information and extracts industrial semantic features; In the transmission module, the extracted industrial semantic features are transmitted to the SNR adaptive deep compressed sensing decoder module through the noisy physical channel industrial semantics; The SNR-adaptive deep compressed sensing decoder module decodes and reconstructs the received industrial semantic features to recover the fine semantic information in the original data.

2. The industrial semantic communication method based on adaptive deep compressed sensing according to claim 1, characterized in that: Extracting industrial semantic information from industrial tool wear data; including: The industrial semantic recognition module includes a residual module, a multi-head self-attention module, a fully connected layer, and an output layer; The input data of the industrial semantic recognition module is the preprocessed vibration signal collected by the three-axis accelerometer, including vibration signals in the X, Y, and Z directions. The residual module includes an initial convolution layer, a multi-level convolution layer, and a residual block. The initial convolution layer and the multi-level convolution layer both include a convolution layer, a batch normalization layer, and a ReLU activation function. In the residual module, two convolution operations are performed through the initial convolution layer and the multi-level convolution layer, and the ReLU activation function is applied after each convolution operation; After the initial convolutional layer processing, low-level features are extracted, and nonlinearity is introduced through the ReLU activation function; then, high-level features are extracted layer by layer through multi-level convolutional layers, and the feature extraction capability is enhanced through the residual block; The features extracted by the industrial semantic recognition module are mapped to the multi-head self-attention module. Key, Query, and Value features are generated through convolution. The dot product of Query and Key is used to calculate the similarity score between features and normalized to generate attention weights. The attention weights are used to perform weighted summation on the Value features to generate an output feature map containing global context information. Multiple self-attention modules work in parallel, and the outputs of each attention head are finally fused into an overall feature map containing more comprehensive global semantic information, resulting in the multi-head self-attention output. The processed multi-head self-attention output, i.e., the high-dimensional feature vector, is subjected to a convolution operation to further extract semantic information related to the task. A learnable parameter Gamma is used for weighted fusion to form the final feature representation. The fused feature map is input into the fully connected layer for semantic recognition and classification, corresponding to different semantic states: non-important semantic information and important semantic information. Important semantic information is fed into the SNR-adaptive deep compressed sensing encoder module.

3. The industrial semantic communication method based on adaptive deep compressed sensing according to claim 1, characterized in that: The preprocessing process of the vibration signal collected by the triaxial accelerometer includes: filtering and denoising, trend removal, normalization, outlier removal, resampling, signal smoothing and data alignment.

4. The industrial semantic communication method based on adaptive deep compressed sensing according to claim 1, characterized in that: Compress important semantic information and extract industrial semantic features; including: The SNR adaptive deep compressed sensing encoder module includes a convolution module, a residual convolution block and an SNR adaptive module; the convolution module includes a one-dimensional convolution layer and a normalization layer; Important semantic features are input into the convolution module, and the deep compressed sensing module compresses the important semantic features. This includes: first, the important semantic features pass through a one-dimensional convolution layer and a normalization layer; then, the semantic features are further enhanced through a residual convolution block; In the SNR adaptive deep compressed sensing encoder, the important semantic information x k After mapping transformation, it is converted into industrial semantic features represented as y k , as shown below: y k =f θ (x k ,c), Among them, x k is the important semantic information extracted by the industrial semantic recognition module, γ is the channel SNR, and f θ (·) represents the mapping function from original data to compressed data, θ is the training parameter of the SNR adaptive deep compressed sensing encoder, and y k is the encoded industrial semantic feature; The encoded industrial semantic features are then processed through a residual convolution block to generate new semantic features, which are then passed to the SNR adaptive module for further processing. In the SNR adaptive module, global pooling is used to generate a global feature F, which is then combined with the current channel SNR = γ to form a new comprehensive feature F′. A nonlinear deep neural network is then used to generate a scaling factor, which is multiplied by the input feature to obtain a channel environment adaptive feature. This feature is then continuously learned and adjusted before being passed to the transmission module. Further preferably, it is assumed that the important semantic feature x output by the industrial semantic recognition module is expressed as: Where h×w represents the size of the feature, and c represents the number of channels. The important semantic feature x is compressed and extracted by the SNR-adaptive deep compressed sensing encoder module, and the compressed three-dimensional semantic feature data y is expressed as:

5. The industrial semantic communication method based on adaptive deep compressed sensing according to claim 1, characterized in that: In the SNR adaptive module, the semantic features output by the residual convolution block are expressed as As the input of the SNR adaptation module; The global pooling method is used to pool the semantic features m G Perform downsampling of spatial dimensions; First, the SNR adaptation module adjusts the semantic feature m G Perform global average pooling processing; the details are as follows: Use the global average pooling method M(·) to get the input semantic features m G Processing is performed to extract the global feature F; for the input feature m G The element m in ij , the calculation formula is as follows: Among them, h×w represents the size of the feature, m ij is feature m G The elements in , F represent the global information features; The global information feature F is concatenated with the signal-to-noise ratio information γ to form a comprehensive feature representation F′, as shown below: Next, F′ is input into a nonlinear deep neural network to generate a scaling factor α, which is calculated as follows: in, Represent the PReLU layer and Sigmoid layer respectively, w1, w2, b1, b2 are the parameters of the linear layer; Perform the scaling factor α and the input semantic feature m of the SNR adaptation module G Element-by-element multiplication of to realize the calculation of channel environment adaptive feature m′, based on input feature m G Re-adapt feature m′: m′=α⊙m G , Finally, the comprehensive feature F′ is output after being processed by MLP and multiplied element by element with the adaptive feature m′ to obtain the final output feature m A ,The process is expressed as: m A =MLP(F′)⊙m′。 6. The industrial semantic communication method based on adaptive deep compressed sensing according to claim 1, characterized in that: The encoded industrial semantic features y k Industrial semantics are transmitted to the SNR adaptive deep compressed sensing decoder module through a noisy physical channel; include: The extracted industrial semantic features y k The signal is transmitted to the SNR adaptive deep compressed sensing decoder module through the additive white Gaussian noise channel, i.e., the noisy physical channel, as shown below: Among them, h∈C is the channel coefficient, n is the channel noise, and it obeys the complex Gaussian distribution n~CN(0,σ 2 ), is the noisy semantic feature received after passing through the noisy physical channel; SNR is defined as follows: Among them, P m represents the average power of semantic features, P n represents the noise power; Furthermore, the received industrial semantic features are decoded and reconstructed to restore the fine semantic information in the original data; including: The SNR adaptive deep compressed sensing decoder module includes a residual transposed convolution block, an SNR adaptive module, and a depth reconstruction block; the residual transposed convolution block and the SNR adaptive module are decoded, and the depth decompression module is used to perform deep reconstruction to obtain the final reconstructed data. The specific process is: First, the received After the residual transposed convolution block, it is used to achieve feature upsampling and recovery, and the resolution of the feature map is restored to a higher level through the transposed convolution operation; Then, the SNR adaptive module is entered. The SNR adaptive module introduces the attention mechanism and the output feature m obtained by the SNR adaptive module is A is sent on a physical channel; Finally, deep reconstruction is performed through the deep decompression module to obtain the final reconstructed data; Further preferably, the final reconstructed data is expressed by the following formula: Where γ is the channel SNR, Implement the mapping of the received compressed data to the restored original data for the decoding function, are the training parameters of the SNR adaptive deep compressed sensing decoder module; The relationship between the original relevant semantic information and the reconstructed semantic information is transformed into an optimization problem, which is expressed as: s.ty k =f θ (x k ,γ) Among them, x k is important semantic information, γ is the channel SNR, f θ (·) is the encoding function that realizes the mapping from original data to compressed data, θ is the training parameter of the SNR adaptive deep compressed sensing encoder module, Implement the mapping of the received compressed data to the restored original data for the decoding function, are the training parameters of the SNR adaptive deep compressed sensing decoder module.

7. The industrial semantic communication method based on adaptive deep compressed sensing according to claim 1, characterized in that: Training the industrial semantic communication model to obtain a trained industrial semantic communication model; including: In the industrial semantic communication model, the optimization goal is to minimize the loss function. For each training sample i, the input of the industrial semantic communication model is z i , the loss function L is expressed as: Where N is the number of samples in the training set, z i is the input data, x i is the processed output; Train an optimal strategy and use the gradient descent algorithm to update the system parameters θ, Until the loss function converges, it can be expressed as: s.ty k =f θ (x k ,γ) Input the test set to verify the industrial semantic communication model; obtain the trained industrial semantic communication model.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the industrial semantic communication method based on adaptive deep compressed sensing described in any one of claims 1 to 7 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the industrial semantic communication method based on adaptive deep compressed sensing according to any one of claims 1 to 7 are implemented.

10. An industrial semantic communication system based on adaptive deep compressed sensing, characterized in that: include: The data acquisition and preprocessing module is configured to: preprocess the data set and divide the preprocessed data set into a training set and a test set; The model building and training module is configured to: build an industrial semantic communication model; and train the industrial semantic communication model to obtain a trained industrial semantic communication model; The industrial semantic communication module is configured to: perform industrial semantic communication using a trained industrial semantic communication model; extract industrial semantic information from industrial tool wear data; classify industrial semantic information into two categories: important semantic information and non-important semantic information; important semantic information includes three key stages: initial wear, stable wear, and accelerated wear; non-important semantic information includes idling status; compress key semantic information and extract industrial semantic features; The received industrial semantic features are decoded and reconstructed to recover the fine semantic information in the original data.

Images