Dense monitoring Internet of Things remote estimation method based on multi-cell semantic enhancement

Abstract

The invention discloses a dense monitoring Internet of Things remote estimation method based on multi-cell semantic enhancement, and belongs to the field of wireless communication. The method comprises the following steps: firstly, an edge device observes a reasoning target, encodes an observation value according to a semantic codebook, and transmits the observation value to an edge server; secondly, the edge server carries out vector quantization on the received signal and sends a quantized code word index to a cloud server through a forward link; and finally, the cloud server performs dequantization and joint decoding on the received quantization index to obtain an estimated value of the reasoning target in the multiple cells, and a remote estimation task is completed. In addition, a loss function is constructed based on an information bottleneck theory and a straight-through estimation gradient approximation method, and the system is trained to be optimal by adopting a two-stage training strategy. According to the method, semantic information extraction of the remote estimation task in a multi-cell scene is realized, the performance of the remote estimation task can be improved while code word redundancy is inhibited, and the utilization efficiency of spectrum resources is remarkably improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of wireless communication, and particularly relates to a dense monitoring Internet of Things remote estimation method based on multi-cell semantic enhancement. BACKGROUND

[0002] In recent years, with the rapid development of the Internet of Things, remote estimation has been widely applied in the fields of smart agriculture, hydrological and meteorological monitoring, industrial automation and intelligent transportation, and has become an important basis for realizing real-time perception and intelligent decision-making. In actual dense monitoring Internet of Things remote estimation scenarios, a large number of sensing devices are often distributed in multiple cells for monitoring dynamic changes of inference targets in a wide area and continuously generating and transmitting massive data. In order to reduce delay and improve data processing efficiency, the fog network access (Fog Radio Access Network, F-RAN) architecture emerges as the times require, which provides a decentralized network infrastructure by sinking computing and storage resources to the network edge.

[0003] In the multi-cell dense monitoring Internet of Things scenario, due to the large number of devices and high density, the spectrum resources in the F-RAN are extremely scarce, and the traditional bit-level-based communication method is difficult to meet the performance requirements of remote estimation. Therefore, the type-based multiple access (Type-Based Multiple Access, TBMA) technology is proposed [2][3]. In TBMA, all devices share a codebook, and each code word corresponds to a specific observation value interval. Unlike traditional time division multiple access, frequency division multiple access, code division multiple access and other multiple access protocols that recover messages device by device, TBMA classifies and transmits based on the type of observation value rather than the device itself, effectively reducing bandwidth occupation, and is considered as a semantic enhanced multiple access way.

[0004] However, in existing TBMA research, the codebook usually adopts a preset fixed codebook [4] or is generated by a neural network [5], which will cause redundant code words in the codebook, including semantically equivalent code words and task-independent code words. Such redundancy will introduce ambiguity in the semantic decoding process, increase the size and complexity of the codebook, and ultimately reduce the accuracy and robustness of remote estimation. Therefore, reducing code word redundancy is crucial to improving the semantic representation efficiency of the remote estimation system. On the other hand, under the multi-cell F-RAN architecture, the capacity of the front link connecting the edge server and the cloud server is usually limited, and the signal received by the edge server must be quantized and compressed. Existing works [5][6] mostly simplify the quantization error as additive noise, and do not propose an efficient quantization strategy for the edge server, which is difficult to effectively apply in actual systems. This not only affects the faithful transmission of semantic information, but also limits the further improvement of remote estimation accuracy.

[0005] Therefore, it is necessary to study a dense monitoring Internet of Things remote estimation method based on multi-cell semantic enhancement. By optimizing the design of semantic codebook, reducing the redundancy of code words, and combining effective quantization strategies, the spectral utilization efficiency and estimation accuracy of the system are improved, thereby meeting the remote estimation requirements of dense monitoring Internet of Things in multi-cell scenarios.

[0006] [1]Peng M, Zhang K. Recent advances in fog radio access networks: Performance analysis and radio resource allocation [J]. IEEE Access, 2016, 4: 5003-5009.

[0007] [2]Liu K, Sayeed AM. Asymptotically optimal decentralized type-based detection in wireless sensor networks [C] / / 2004 IEEE International Conference on Acoustics, Speech, and Signal Processing. IEEE, 2004, 3: iii-873.

[0008] [3]Mergen G, Tong L. Type based estimation over multiaccess channels [J]. IEEE Transactions on Signal Processing, 2006, 54(2): 613-626.

[0009] [4]Anandkumar A, Tong L. Type-based random access for distributed detection over multiaccess fading channels [J]. IEEE Transactions on Signal Processing, 2007, 55(10): 5032-5043.

[0010] [5] Kassab R, Simeone O, Popovski P. Information-centric grant-free access for IoT fog networks: Edge vs. cloud detection and learning[J]. IEEE Transactions on Wireless Communications, 2020, 19(10): 6347-6361.

[0011] [6] Dommel J, Utkovski Z, Simeone O, et al. Joint source-channel coding for semantics-aware grant-free radio access in IoT fog networks[J]. IEEE Signal Processing Letters, 2021, 28: 728-732. SUMMARY

[0012] To solve the above problems, the application discloses a dense monitoring Internet of Things remote estimation method based on multi-cell semantic enhancement, which is called a vector quantization and information bottleneck-inspired type based multiple access method (VQ-IB-TBMA). The method constructs a remote estimation framework suitable for a multi-cell scene, and systematically considers the perception, transmission, quantization and inference process of semantic information. Specifically, the method designs a TBMA semantic codebook shared by multiple cells and multiple devices, and constructs a shared vector quantization (VQ) embedding space between the edge server and the cloud server. At the same time, based on the information bottleneck (IB) theory and the straight-through estimator (STE) gradient approximation method, a multi-cell collaborative inference loss function is established, and a two-stage training method is designed, so as to realize the joint optimization of the TBMA codebook and the VQ embedding space. Through the design, the application can significantly improve the accuracy of the remote estimation task in the multi-cell scene, and effectively reduce the communication spectrum resource consumption.

[0013] The specific steps are as follows:

[0014] Step one, the edge device observes the inference target, and based on the type of observation value, uses the preset semantic codebook to encode, maps the observation value to the corresponding code word;

[0015] The specific steps are as follows:

[0016] Step 101, multiple edge devices independently observe the inference target (Quantity of Interest, QoI) in the cell where they are located, and obtain the discretized observation value. In cell c∈{1,2,...,C}, multiple edge devices independently observe an inference target θ c ∈{θ (1) ,θ (2) ,...,θ (L)}, where L represents the number of types of QoI, and each device i∈{1,2,...,N c} obtains a measurement value The measurement value is drawn from the unknown conditional distribution p(x c ∣θ c ). It is assumed that each observation value is taken from a discrete table of size M {0,1,...,M-1}. In actual systems, if the observation value is a continuous variable, the observation value needs to be quantized to M order.

[0017] Step 102, input the observation value of the edge device into the semantic encoder, and convert it into a code word with specific semantic characteristics based on the shared semantic codebook, so as to be transmitted subsequently.

[0018] All edge devices share a semantic codebook which contains M code words. Each code word corresponds to an observation value m∈{0,1,...,M-1}. Each code word consists of K symbols and needs to satisfy the power constraint ||s m || 2 ≤E. These code words represent the semantic characteristics of the observation value. This process is implemented by a semantic encoder ε S (·), whose input is the observation value and the output is the corresponding code word Therefore, the signal transmitted by device i in cell c can be represented as:

[0019]

[0020] Step two, the encoded code word is transmitted to the edge server through the wireless channel, and the influence of large-scale fading, small-scale fading, additive noise, and interference between multiple cells is considered;

[0021] The specific steps are as follows:

[0022] Step 201, model the wireless channel between the edge devices and the edge server. Consider a wireless transmission model where each device's transmitted signal experiences small-scale fading and large-scale fading. Specifically, the wireless channel between device i in cell c1 and the edge server in cell c2 can be represented as:

[0023]

[0024] The small-scale fading component obeys a Rician distribution and can be modeled as:

[0025]

[0026] where μ h and denote the mean and variance of the intra-cell wireless channel, respectively, and μ g and denote the mean and variance of the inter-cell wireless channel, respectively.

[0027] The large-scale fading component takes into account the distance-dependent path loss and log-normal shadowing, and has the form:

[0028]

[0029] where d (unit: km) denotes the distance between device i in cell c1 and the edge server in cell c2. The path loss model is The shadowing term obeys a log-normal distribution with mean 0 and standard deviation 8 dB.

[0030] Step 202, considering channel fading, additive white Gaussian noise, and inter-cell interference, the received signal at the edge server can be represented as:

[0031]

[0032] where is the additive white Gaussian noise, which is assumed to be independent and identically distributed across cells with noise power

[0033] Step 3, considering the fronthaul link capacity constraint, the edge server performs vector quantization on the received signal and maps it to a discrete index in the embedding space;

[0034] The cloud server and the edge server share the embedding space, which can be represented as where q is the number of quantization bits, and D = 2 q is the size of the discrete latent space, is the embedding vector in the embedding space, and y is the corresponding codebook indexc ∈{1,2,...,D}. The vector quantization process can be represented as:

[0035]

[0036] wherein, denotes a vector quantizer.

[0037] Step four, the cloud server receives the vector index from the edge server, dequantizes it to recover the quantized vector, and jointly estimates the QoI in all cells through the semantic decoder;

[0038] The specific steps are as follows:

[0039] Step 401, the cloud server dequantizes the received quantization index according to the shared embedding space, and converts it into the corresponding quantized vector. The process can be represented as:

[0040]

[0041] wherein, denotes a dequantizer.

[0042] Step 402, the semantic decoder receives the quantized vector of each cell edge server, and performs remote estimation to predict the inference target of each cell. A semantic decoder based on artificial neural network with parameter π is used to jointly estimate the QoI in all cells:

[0043]

[0044] Step five, a joint loss function based on information bottleneck theory and vector quantization is constructed. By combining the variational information bottleneck method and the direct estimation approximate gradient propagation, the balance between estimation accuracy and code redundancy is achieved;

[0045] The specific steps are as follows:

[0046] Step 501, to balance the estimation accuracy and code redundancy, and solve the problem that mutual information is difficult to calculate directly, the information bottleneck principle is introduced, and the variational information bottleneck method is used to construct the variational information bottleneck loss function

[0047]

[0048] wherein, the first term is the distortion term, defined as:

[0049]

[0050] This term is used to measure the cross-entropy loss between the predicted value and the true value of the inference target, so as to improve the inference accuracy. The second term is the code rate term, defined as:​

[0051]

[0052] This term is used to promote more compact and structured semantic representation, thereby reducing the redundancy between codewords. The parameter β ≥ 0 is used to adjust the balance between inference accuracy and information compression.

[0053] Step 502, for the non-derivative problem introduced by the quantization process, the straight-through estimation method is used to approximate the gradient propagation, and the vector quantization loss function is constructed

[0054]

[0055] Wherein, sg[·] represents the gradient stop operator; the first term is used to align the embedding vector with the received signal, mainly training the embedding space Q; the second term is used to promote the stable update of the embedding space Q, thereby ensuring the reliability of the training process.

[0056] Step 503, the above variational information bottleneck loss function and vector quantization loss function are integrated to obtain the final joint loss function

[0057]

[0058] Step six, the above-mentioned modules of the application are trained to obtain the optimal parameters. In order to avoid the adverse effects of straight-through estimation in the gradient approximation propagation process, a two-stage training strategy is designed, including the pre-training stage of the codebook and the decoder, and the joint optimization stage of the system;

[0059] The specific steps are as follows:

[0060] Step 601, pre-training the codebook and decoder without considering the influence of quantization. Assuming that the signal received by the edge server can be transmitted to the cloud server without distortion, temporarily ignoring the quantization process. Under this assumption, the variational information bottleneck loss function is used to optimize the codebook S and the decoder D π (·). Then, the codewords in the codebook are clustered to reduce the number of codewords and redundancy. The above optimization process is repeated for the compressed codebook and the decoder, and finally the pre-trained codebook S and the decoder D π (·) are obtained, laying a foundation for the subsequent stage.

[0061] Step 602, using the pre-trained codebook and decoder, introducing the quantization constraint, and optimizing the whole system through the joint loss function. The pre-trained codebook S and the decoder D π(·) as the initial parameters, and introduce the embedding space Q. The edge server quantizes the received signal according to the embedding space Q to obtain the quantization index y c and transmits to the cloud server; the cloud server dequantizes the index and decodes through the decoder to realize remote estimation of QoI. In the training process, the joint loss function is used to jointly optimize the codebook S, the embedding space Q and the decoder D π , so as to realize the optimization of the overall performance of the system.

[0062] The advantages of the present application are:

[0063] 1. A dense monitoring Internet of Things remote estimation method based on multi-cell semantic enhancement, which realizes semantic information extraction of remote estimation tasks in a multi-cell scenario.

[0064] 2. A dense monitoring Internet of Things remote estimation method based on multi-cell semantic enhancement, which effectively improves the performance of remote estimation tasks while considering code word redundancy suppression.

[0065] 3. A dense monitoring Internet of Things remote estimation method based on multi-cell semantic enhancement, which improves the utilization efficiency of spectrum resources. BRIEF DESCRIPTION OF DRAWINGS

[0066] Figure 1 is a typical multi-cell dense monitoring Internet of Things remote estimation scenario diagram of the present application;

[0067] Figure 2 is a dense monitoring Internet of Things remote estimation method system implementation block diagram in the corresponding multi-cell scenario of the present application;

[0068] Figure 3 is a dense monitoring Internet of Things remote estimation method flowchart in the corresponding multi-cell scenario of the present application;

[0069] Figure 4 is the error rate comparison of the present application and other remote estimation methods under different signal-to-noise ratios;

[0070] Figure 5 is the error rate comparison of the present application and other remote estimation methods under different interference intensities;

[0071] Figure 6 is the spectrum utilization rate comparison of the present application and other remote estimation methods under different device quantities; DETAILED DESCRIPTION

[0072] In order for those skilled in the art to more clearly understand and implement the present application, the technical solutions of the embodiments of the present application will be described in detail below with specific embodiments and with reference to the accompanying drawings.

[0073] The present application proposes a multi-cell semantic enhancement-based dense monitoring Internet of Things remote estimation method, called Vector Quantization and Information Bottleneck-Inspired Type Based Multiple Access (VQ-IB-TBMA). The present application can be applied to the multi-cell joint remote estimation scene in the dense monitoring Internet of Things system as shown in the figure. Figure 1 In this scene, the edge devices in multiple cells observe the inference target, encode it and transmit it to the edge server in the cell, the edge server transmits the received signal to the cloud server through the front transmission link, and the cloud server performs remote estimation on the inference target in multiple cells according to the received signal, thereby completing the remote estimation task.

[0074] The system implementation block diagram of the present application is shown in the figure. Figure 2 First, the edge device observes the inference target θ c , and encodes the observation value according to the semantic codebook S to obtain the transmission code word and transmits it to the edge server, which is affected by channel fading, noise interference and multi-cell interference; second, the edge server vector quantizes the received signal y c , and sends the quantized code word index y c to the cloud server through the front transmission link; finally, the cloud server dequantizes and jointly decodes the received quantized index to obtain the estimation value of the inference target in multiple cells , and completes the remote estimation task. In addition, the present application constructs a loss function based on the information bottleneck theory and the direct estimation gradient approximation method, and adopts a two-stage training strategy to train the above system to the optimal, and fixes the system parameters and applies them to the actual remote estimation task.

[0075] As shown in the figure, the specific steps of the present application are as follows: Figure 3

[0076] Step one, the edge device observes the inference target, and based on the type of the observation value, encodes it using the preset semantic codebook to map the observation value to the corresponding code word;

[0077] The specific steps are as follows:

[0078] Step 101, multiple edge devices independently observe the inference target in the cell where they are located, and obtain the discretized observation value. In the cell c∈{1,2,…,C}, multiple edge devices independently observe an inference target θ c ∈{θ (1) ,θ (2) ​..., θ (L) , where L denotes the number of types of QoI, each device i ∈ {1, 2,..., N c , obtains one measurement The measurement is drawn from an unknown conditional distribution p(x c | θ c ). Assume each observation is taken from a discrete table of size M {0, 1,..., M-1}. In a real system, if the observation is a continuous variable, the observation needs to be quantized to M levels.

[0079] Step 102, input the observation of the edge device into the semantic encoder, and convert it into a code word with specific semantic features based on the shared semantic codebook, so as to be transmitted subsequently.

[0080] All edge devices share a semantic codebook which contains M code words. Each code word corresponds to an observation m ∈ {0, 1,..., M-1}. Each code word is composed of K symbols and needs to satisfy the power constraint ||s m || 2 ≤ E. These code words represent the semantic features of the observation. This process is implemented by a semantic encoder ε S (·), whose input is the observation and the output is the corresponding code word Therefore, the signal transmitted by device i in cell c can be represented as:

[0081]

[0082] Step two, transmit the encoded code word to the edge server through the wireless channel, and comprehensively consider the influence of large-scale fading, small-scale fading, additive noise, and interference between multiple cells;

[0083] The specific steps are as follows:

[0084] Step 201, model the wireless channel between the edge device and the edge server. Consider a wireless transmission model in which the signal transmitted by each device will experience small-scale fading and large-scale fading. Specifically, the wireless channel between device i in cell c1 and the edge server in cell c2 can be represented as:

[0085]

[0086] The small-scale fading component obeys the Rice distribution and can be modeled as:

[0087]

[0088] where μh and denote the mean and variance of the intra-cell wireless channel, respectively, μ g and denote the mean and variance of the inter-cell wireless channel, respectively.

[0089] The large-scale fading component considers the distance-dependent path loss and log-normal shadowing, which is given by

[0090]

[0091] where denotes the distance between device i in cell c1 and the edge server in cell c2 (unit: km). The path loss model is The shadowing term follows a log-normal distribution with mean 0 and standard deviation 8 dB.

[0092] Step 202, considering the channel fading, additive white Gaussian noise, and inter-cell interference, the signal received by the edge server can be represented as

[0093]

[0094] where, is the additive white Gaussian noise, which is assumed to be independent and identically distributed across cells with noise power

[0095] Step 3, considering the capacity limitation of the fronthaul link, the edge server performs vector quantization on the received signal and maps it to a discrete index in the embedding space;

[0096] The cloud server and the edge server share the embedding space, which can be represented as where q is the number of quantization bits, and D = 2 q is the size of the discrete latent space, is the embedding vector in the embedding space, and its corresponding code word index is y c ∈ {1, 2,..., D}. The vector quantization process can be represented as:

[0097]

[0098] where, denotes the vector quantizer.

[0099] Step 4, the cloud server receives the vector index from the edge server, performs dequantization to recover the quantized vector, and jointly estimates the QoI in all cells through a semantic decoder;

[0100] The specific steps are as follows:

[0101] Step 401, the cloud server according to the shared embedding space, the received quantization index is dequantized, and the corresponding quantization vector is converted. The process can be represented as:

[0102]

[0103] wherein, represents the dequantizer.

[0104] Step 402, the semantic decoder receives the quantization vector of each cell edge server and performs remote estimation to predict the inference target of each cell. The semantic decoder based on artificial neural network with parameter π jointly estimates the QoI in all cells:

[0105]

[0106] Step five, construct a joint loss function based on information bottleneck theory and vector quantization. By combining the variational information bottleneck method and the straight-through estimation approximate gradient propagation, the balance between estimation accuracy and code redundancy is achieved.

[0107] The specific steps are as follows:

[0108] Step 501, to balance the estimation accuracy and code redundancy, and solve the problem that mutual information is difficult to calculate directly, the information bottleneck principle is introduced, and the variational information bottleneck method is used to construct the variational information bottleneck loss function

[0109]

[0110] wherein, the first term is the distortion term, defined as:

[0111]

[0112] This term is used to measure the cross-entropy loss between the predicted value of QoI and the true value, so as to improve the inference accuracy. The second term is the code rate term, defined as:

[0113]

[0114] This term is used to promote more compact and structured semantic representation, so as to reduce the redundancy between code words. Parameter β≥0 is used to adjust the balance between inference accuracy and information compression.

[0115] Step 502, to solve the non-differentiable problem introduced by the quantization process, the straight-through estimation method is used to approximate the gradient propagation, and the vector quantization loss function is constructed

[0116]

[0117] Wherein, sg[·] represents the gradient stop operator; the first term is used to align the embedding vector with the received signal, mainly training the embedding space Q; the second term is used to promote the stable update of the embedding space Q, so as to ensure the reliability of the training process.

[0118] Step 503, the above-mentioned variational information bottleneck loss function and vector quantization loss function are integrated to obtain a final joint loss function

[0119]

[0120] Step six, the above-mentioned modules are trained to obtain optimal parameters. In order to avoid the adverse effects of straight-through estimation on the gradient approximation propagation process, a two-stage training strategy is designed, including a pre-training stage of the codebook and the decoder, and a joint optimization stage of the system.

[0121] The specific steps are as follows:

[0122] Step 601, pre-training the codebook and the decoder without considering the influence of quantization. Assuming that the signal received by the edge server can be transmitted to the cloud server without distortion, temporarily ignoring the quantization process. Under this assumption, the variational information bottleneck loss function is used to optimize the codebook S and the decoder D π (·). Subsequently, the code words in the codebook are clustered to reduce the number of code words and redundancy. The pre-trained codebook S and the decoder D π (·) are obtained by repeating the above optimization process on the compressed codebook and the decoder, laying a foundation for the subsequent stage.

[0123] Step 602, using the pre-trained codebook and decoder, introducing the quantization constraint, and optimizing the entire system through the joint loss function. The pre-trained codebook S and the decoder D π (·) are used as initial parameters, and the embedding space Q is introduced. The edge server quantizes the received signal according to the embedding space Q to obtain the quantization index y c and transmits it to the cloud server; the cloud server dequantizes the index and decodes it through the decoder to realize remote estimation of the QoI. In the training process, the joint loss function is used to jointly optimize the codebook S, the embedding space Q and the decoder D π (·), so as to realize the optimization of the overall performance of the system.

[0124] The application compares the performance of the proposed method and other remote estimation algorithms under the same parameter conditions based on real air quality monitoring data sets. The data set used is derived from daily air quality monitoring records in 360 major cities in China from 2014 to 2020. The experiment selects 8 air quality monitoring stations in the urban area of Beijing as observation equipment, and the PM2.5 concentration collected by the station in real time as the observation value. According to the spatial distribution of the station, it is divided into 2 small areas, and the air quality level in the small area is taken as the reasoning target.

[0125] Figure 4 The remote estimation error rates of the application and the comparison algorithm under different signal-to-noise ratios are shown. The experimental results show that as the signal-to-noise ratio increases, the remote estimation error rates of all methods decrease, but the proposed method is significantly better than the traditional edge remote estimation algorithm under each signal-to-noise ratio, which shows that the application has better robustness to noise.

[0126] Figure 5 The remote estimation error rates of the application and the comparison algorithm under different inter-cell interference intensities The results show that due to the correlation of the inter-cell QoI, the inter-cell interference may be converted into performance gain in some cases. However, through the cloud joint remote estimation mechanism, the application can better learn and utilize this correlation, so it is significantly better than the traditional edge remote estimation algorithm under different interference conditions, showing better inter-cell interference robustness.

[0127] Figure 6 The spectral utilization rates of the application and the comparison algorithm under different numbers of devices in the small area are shown. The results show that the number of channels required by the traditional frequency division multiple access scheme increases linearly with the number of devices in the small area; and the spectral utilization rate of the traditional TBMA method is significantly lower than that of the proposed method due to the non-optimization of the code word redundancy. The results show that the application can effectively improve the utilization efficiency of spectral resources while ensuring the accuracy of remote estimation.

[0128] In summary, by implementing the application, a multi-cell semantic enhanced dense monitoring Internet of Things remote estimation method is realized, which realizes semantic information extraction of remote estimation tasks in a multi-cell scenario, effectively improves the performance of remote estimation tasks while considering code word redundancy suppression, and significantly improves the utilization efficiency of spectral resources.

[0129] The above is the preferred embodiment of the application. It should be noted that for those skilled in the art, without departing from the principles of the application, some improvements and refinements can be made, which are also considered within the protection scope of the application.

Claims

1. A remote estimation method for dense monitoring IoT based on multi-cell semantic enhancement, the specific implementation steps of which are as follows: Step 1: The edge device observes the inference target and encodes it using a preset semantic codebook based on the type of the observation, mapping the observation to the corresponding codeword. The specific steps are as follows: Step 101: Multiple edge devices independently observe the inference target (Quantity of Interest, QoI) within their respective cells and obtain discretized observation values. In a cell c∈{1,2,…,C}, multiple edge devices independently observe one inference target θ. c ∈{θ (1) ,θ (2) ,…,θ (L) }, where L represents the number of types of QoI, and each device i∈{1,2,...,N} c Get a measurement value x i c, the measured value is from the unknown conditional distribution p(x) c |θ c The observations are extracted from a discrete table of size M, {0,1,...,M-1}. In a real system, if the observations are continuous variables, then the observations need to be quantized to order M. Step 102: Input the observations from the edge device into the semantic encoder and convert them into codewords with specific semantic features based on the shared semantic codebook for subsequent transmission. All edge devices share a semantic codebook. It contains M codewords. Each codeword... This corresponds to an observation m∈{0,1,...,M-1}. Each codeword consists of K symbols and must satisfy the power constraint ||s m || 2 ≤E. These codewords characterize the semantic features of the observations. This process is performed by the semantic encoder ε. S (·) is implemented, with the input being the observed values. The output is the corresponding codeword. Therefore, the signal transmitted by device i in cell c can be represented as: Step 2: Transmit the encoded codewords to the edge server via a wireless channel, taking into account the effects of large-scale fading, small-scale fading, additive noise, and interference between multiple cells. The specific steps are as follows: Step 201: Model the wireless channel between the edge device and the edge server. Consider a wireless transmission model where the signal transmitted by each device experiences both small-scale and large-scale fading. Specifically, the wireless channel between device i in cell c1 and the edge server in cell c2 can be represented as: Small-scale fading components Following a Rice distribution, it can be modeled as: in, μ h and Let μ represent the mean and variance of the wireless channels in the same cell, respectively. g and These represent the mean and variance of the wireless channels in different cells, respectively. The large-scale fading component takes into account distance-dependent path loss and log-normal shading, and its form is as follows: in (Unit: km) represents the distance between device i in cell c1 and the edge server in cell c2. The path loss model is as follows: Shadowed items It follows a log-normal distribution with a mean of 0 and a standard deviation of 8 dB. Step 202: Taking into account channel fading, additive white Gaussian noise, and multi-cell interference, the signal received by the edge server can be expressed as: in The noise is additive white Gaussian noise, assumed to be independently and identically distributed across all cells, with a noise power of... Step 3: Considering the capacity limitations of the fronthaul link, the edge server performs vector quantization on the received signal and maps it into a discrete index in the embedding space; The cloud server and the edge server share the embedded space, which can be represented as: Where the number of quantization bits is q, and D = 2 q It is the size of the discrete latent space. Let y be the embedding vector in the embedding space, and let y be the codeword index. c ∈{1,2,...,D}. The vector quantization process can be represented as: in, This represents a vector quantizer. Step 4: The cloud server receives the vector index from the edge server, dequantizes it to recover the quantized vector, and jointly estimates the QoI in all cells through the semantic decoder; The specific steps are as follows: Step 401: The cloud server dequantizes the received quantization index according to the shared embedding space, converting it into the corresponding quantization vector. This process can be represented as: in, This represents the dequantizer. Step 402: The semantic decoder receives the quantized vectors from the edge servers of each cell and performs remote estimation to predict the inference target of each cell. A semantic decoder based on an artificial neural network with parameter π is used. To jointly estimate the QoI across all cells: Step 5: Construct a joint loss function based on information bottleneck theory and vector quantization. By combining the variational information bottleneck method with the pass-through estimation approximate gradient propagation, a balance is achieved between estimation accuracy and codeword redundancy. The specific steps are as follows: Step 501: To balance estimation accuracy and codeword redundancy, and to address the difficulty of directly calculating mutual information, the information bottleneck principle is introduced, and a variational information bottleneck loss function is constructed using the variational information bottleneck method. The first term is the distortion term, defined as: This term measures the cross-entropy loss between the predicted and actual values ​​of the inference target, thereby improving inference accuracy. The second term is the code rate term, defined as: This term promotes more compact and structured semantic representations, thereby reducing redundancy between codewords. The parameter β≥0 is used to adjust the balance between inference accuracy and information compression. Step 502: To address the non-differentiability problem introduced by the quantization process, a pass-through estimation method is used to approximate gradient propagation, and a vector quantization loss function is constructed. Where sg[·] represents the gradient stopping operator; the first term is used to align the embedding vector with the received signal, mainly training the embedding space Q; the second term is used to promote the stable update of the embedding space Q, thereby ensuring the reliability of the training process. Step 503: Combining the variational information bottleneck loss function and the vector quantization loss function, the final joint loss function is obtained. Step Six: Train the above-mentioned module proposed in this invention to obtain optimal parameters. To avoid the adverse effects of pass-through estimation during gradient approximation propagation, a two-stage training strategy is designed, including a pre-training stage for the codebook and decoder, and a joint optimization stage for the system. The specific steps are as follows: Step 601: Pre-train the codebook and decoder without considering the impact of quantization. Assume the signal received by the edge server can be transmitted to the cloud server without distortion, and temporarily ignore the quantization process. Under this assumption, the variational information bottleneck loss function is adopted. For codebook S and decoder D π (·) Optimization is then performed. Subsequently, the codewords in the codebook are clustered to reduce the number of codewords and redundancy. The above optimization process is then repeated on the compressed codebook and decoder to finally obtain the pre-trained codebook S and decoder D. π (·), laying the foundation for subsequent stages. Step 602: Utilizing the pre-trained codebook and decoder, introduce quantization constraints and perform end-to-end optimization of the entire system through a joint loss function. The pre-trained codebook S and decoder D are used. π (·) is used as the initial parameter, and the embedding space Q is introduced. The edge server quantizes the received signal according to the embedding space Q to obtain the quantization index y. c The data is then transmitted to a cloud server; the cloud server dequantizes the index and decodes it using a decoder to achieve remote estimation of QoI. During training, a joint loss function is used. For the codebook S, the embedding space Q, and the decoder D π (·) Perform joint optimization to achieve the optimal overall system performance.

Images