Communication and calculation parallel method based on IQ domain processing, storage medium and product

Abstract

The invention discloses a communication and calculation parallel method based on IQ domain processing, a storage medium and a product, a sending end extracts multi-level features of input data through a neural network encoder and quantifies the multi-level features to generate a source feature bit stream; a first bit stream is generated through adding a protocol field, and a second bit stream is generated through channel coding and modulated into an IQ symbol to be sent. A receiving end supports dual-mode operation: in a calculation mode, a pre-stored prototype vector is used for direct classification in an IQ domain, and demodulating and decoding are bypassed; in a communication mode, protocol fields are demodulated, decoded and stripped in sequence, and original data are reconstructed through a neural network decoder. According to the method, the native parallelism of physical layer communication and calculation is realized, the end-to-end time delay is remarkably reduced, and the method is compatible with an existing wireless protocol and has high compatibility and high robustness in a dynamic environment; according to application scene requirements, a rapid calculation mode and a complete communication mode can be flexibly switched at a receiving end, or classification precision or reconstruction quality can be emphasized by adjusting a coding strategy.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication and intelligent computing technology, specifically relating to a parallel communication and computing method, storage medium and product based on IQ domain processing. Background Technology

[0002] The integration of wireless communication and intelligent computing is the core of future 6G and IoT applications. Existing systems generally adopt a serial paradigm of "communication first, then computation": the receiving end needs to receive and decode the data completely, recover the original information, and then input it into the computational model for inference. This paradigm has inherent contradictions: source compression to ensure communication quality may lose features that are crucial to inference; while transmitting complete original data will lead to high latency and low spectral efficiency.

[0003] Parallel frameworks, exemplified by "wireless computing," attempt to address this problem, but they typically rely on strict synchronization, ideal channel state information, or dedicated hardware, making them difficult to deploy in real-world dynamic, multipath environments. Therefore, a new method is urgently needed that can achieve efficient parallelism between communication and computation without compromising existing protocols or complex hardware modifications. Summary of the Invention

[0004] To address the aforementioned issues, the present invention aims to provide a parallel communication and computing method, storage medium, and product based on IQ domain processing. This method enables the receiving end to perform fast inference directly at the IQ symbol level during transmission at the same physical layer, while retaining complete data communication reconstruction capabilities. This significantly reduces end-to-end latency and improves overall system efficiency.

[0005] To achieve the above objectives, the technical solution adopted by the present invention includes:

[0006] A parallel communication and computation method based on IQ domain processing, including sender processing and receiver processing; Sending end processing includes: S1. Input the input data into the encoder, extract the feature representation of the input data, quantize the feature representation, and generate the source feature bitstream; S2. Add a protocol field to the source feature bitstream obtained in S1 to generate the first bitstream; S3. Perform channel coding on the first bitstream obtained in S2 to generate the second bitstream; S4. Modulate the second bitstream obtained in S3 into IQ symbols and send it; The receiving end processing includes: receiving the IQ symbols obtained from S4, and selecting the calculation mode or communication mode according to the requirements; The specific computation mode involves using pre-stored prototype vectors to directly perform classification reasoning on IQ symbols in the IQ domain and output the classification results. Communication modes include: S5. Demodulate the IQ symbols to obtain the received bit stream; S6. Perform channel decoding on the received bitstream obtained in S5 to recover the first bitstream estimate; S7. Remove the protocol field from the first bitstream estimate obtained in S6 to recover the source feature bitstream estimate; S8. Input the source feature bitstream estimate obtained in S8 into the decoder to reconstruct and output the estimated data of the input data.

[0007] Preferably, the encoder in S1 is a multi-layer convolutional neural network.

[0008] Preferably, the protocol fields in S2 include physical layer header information, MAC layer frame header information, and verification information; wherein, the physical layer header information includes at least a synchronization sequence or signaling field, and the MAC layer frame header information includes at least a frame control field or an address field.

[0009] Preferably, the channel coding in S3 uses LDPC code.

[0010] Preferably, the modulation in S4 is QAM modulation or PSK modulation that maps the second bitstream to complex symbols.

[0011] Preferably, the demodulation in S5 is either hard-decision demodulation or soft-decision demodulation, which corresponds to the modulation method used in S4.

[0012] Preferably, the channel decoding in S6 is a decoding algorithm corresponding to the channel coding method used in S3.

[0013] Preferably, the prototype vectors used in the calculation mode are obtained in advance and stored at the receiving end in the following way: During the training phase, the sample data of each predetermined category are processed by the sending end S1~S4 in sequence to obtain the corresponding reference IQ symbols, and prototype vectors representing each category are calculated based on the reference IQ symbols.

[0014] A computer-readable storage medium stores a computer program that, when executed by a processor, implements the IQ domain-based parallel communication and computation method disclosed in this application.

[0015] A computer program product includes a computer program / instructions that, when executed by a processor, implement the IQ domain-based parallel communication and computation method disclosed in this application.

[0016] Compared with the prior art, the advantages of the present invention are: (1) The communication and computing parallel method, storage medium and product based on IQ domain processing of the present invention achieves native parallel communication and computing throughout the process. By directly reasoning in the IQ domain, the physical layer achieves "integrated computing", which greatly reduces end-to-end processing latency.

[0017] (2) The communication and computing parallel method, storage medium and product based on IQ domain processing of the present invention can directly embed the generated bit stream into the existing wireless communication protocol stack (such as Wi-Fi) through the protocol alignment mechanism, without the need to modify the hardware or strictly synchronize, adapt to dynamic channels and mobile environments, and have high compatibility and high robustness.

[0018] (3) The communication and computing parallel method, storage medium and product based on IQ domain processing of the present invention adopts a multi-level feature fusion source coding strategy, which simultaneously retains discriminative information for high-precision reasoning and structural information for high-fidelity reconstruction in the generated bit stream.

[0019] (4) The communication and computing parallel method, storage medium and product based on IQ domain processing of the present invention can flexibly switch between "fast computing" and "complete communication" modes at the receiving end according to the application scenario requirements, or adjust the coding strategy to focus on classification accuracy or reconstruction quality. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the following detailed description to explain the invention, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the basic structure of the convolutional neural network used in this invention.

[0021] Figure 2(a) shows two images transmitted by the transmitter, taken from the EuroSAT and CIFAR-10 datasets respectively. Figure 2(b) shows the reconstruction effect at the receiver when only the fifth layer features of the encoder are transmitted. Figure 2(c) shows the reconstruction effect at the receiver when both the third and fifth layer features of the encoder are transmitted simultaneously.

[0022] Figure 3 This diagram illustrates the comparison of inference time between the proposed solution and the traditional discrete architecture on five datasets, showing the inference time performance of both under GPU and CPU computing conditions.

[0023] Figure 4 This paper presents the performance of the present invention in reconstructing and classifying five datasets under three QAM modulation schemes.

[0024] Figure 5 This diagram illustrates the performance comparison between the present invention and JPEG at different compression ratios, showing the comparison results for compression ratios ranging from 1.5 to 2.5.

[0025] Figure 6 This is a schematic diagram of the performance of the present invention under different channel coding configurations. The six sets of parity check matrix configurations cover different combinations of code length and code rate. The configurations A to F correspond to (324, 648), (486, 648), (648, 1296), (972, 1296), (972, 1944) and (1458, 1944), respectively.

[0026] Figure 7 This diagram illustrates the performance of the present invention under different communication protocols, where protocols A to E correspond to 802.11n (20MHz), 802.11n (40MHz), 802.11ac (40MHz), 802.11ac (80MHz), and 802.11ax, respectively.

[0027] Figure 8 This is a schematic diagram illustrating the performance of the present invention under different communication distance conditions, wherein the distance between the transmitter and the receiver ranges from 10m to 100m.

[0028] Figure 9 This is a schematic diagram illustrating the performance of the present invention in a multi-user scenario, where the number of users is 5.

[0029] Figure 10 This is a schematic diagram illustrating the performance of the present invention in a mobile scenario, where the receiving end moves at speeds of 5, 10, 15, 20, and 25 km / h.

[0030] Figure 11 This diagram illustrates the performance of the invention under different channel environments, where the signal-to-noise ratio (SNR) ranges from -20dB to 10dB, and shows the test results on five datasets.

[0031] Figure 12 This diagram illustrates the test results of the present invention under the performance limits of classification and reconstruction, showing its performance on five datasets.

[0032] Figure 13 is a schematic diagram of the performance of the present invention in complex scenarios, wherein: Figure 13(a) shows the location of 10 test points in an outdoor road environment and the performance of each point; Figure 13(b) shows the location of 10 test points in three-dimensional space and the performance of each point; Figure 13(c) shows the location of 10 test points in an indoor room environment and the performance of each point. Detailed Implementation

[0033] The invention is not limited to the specific embodiments described below. All equivalent modifications made based on the technical solutions of this application fall within the protection scope of this invention. Unless otherwise specified, all components and devices in this invention utilize components and devices known in the prior art.

[0034] Example This embodiment discloses a parallel communication and computation method based on IQ domain processing. For example... Figure 1 As shown, this embodiment uses a five-layer convolutional neural network as the basic structure of the encoder, but it should be understood that the core idea of ​​the present invention is not limited to this specific network structure.

[0035] The specific data flow and processing steps of the method in this embodiment are as follows: (1) Processing flow at the sending end: Step 1: Obtain input data X (e.g., a 32×32 pixel RGB image). Input X into a pre-trained five-layer convolutional neural network encoder. This encoder extracts multi-level feature representations of X. Specifically, it extracts structural features fs (256×2×2, flattened to a 1024-dimensional vector) from the third convolutional layer and discriminative features fd (512×1×1, flattened to a 512-dimensional vector) from the fifth convolutional layer. Concatenate fs and fd to obtain a joint feature representation f (1536-dimensional). Then quantize the joint feature representation f. Linearly normalize each element value of f to the interval [0, 255] (i.e., corresponding to quantization bits q=8), and round to the nearest integer to obtain an integer feature vector f_int. Convert each integer value in f_int into an 8-bit binary sequence, arrange them in order, and generate the source feature bitstream B1 (1536×8 bits in length).

[0036] Specifically, the basic structure of a convolutional neural network is as follows: Figure 1As shown, the input is a 32×32 three-channel image. First, it passes through the first convolutional module Conv1, which combines batch normalization (BN) and ReLU activation to convert the number of feature channels to 64, resulting in a 16×16 feature map. Then, it passes through the second convolutional module Conv2+BN+ReLU and the third convolutional module Conv3+BN+ReLU, increasing the number of feature channels to 128 and 256 respectively, reducing the feature map size to 8×8 and 4×4 accordingly. Next, it passes through the fourth convolutional module Conv4+BN+ReLU and the fifth convolutional module Conv5+BN+ReLU, finally increasing the number of feature channels to 512, resulting in a 2×2 feature representation. Simultaneously, average pooling is performed on the 256-channel, 4×4 feature map output from the third layer, downsampling it to a 256-channel, 2×2 feature representation. Therefore, this invention obtains a global discriminative feature of size 512×1×1 and a structural feature of size 256×2×2 during the feature extraction stage. The 512×1×1 feature is flattened into a 512-dimensional vector, and the 256×2×2 feature is flattened into a 1024-dimensional vector. These two features are then concatenated to obtain a joint feature representation of length 1536. The 512-dimensional feature is primarily used for classification reasoning, while the 1024-dimensional feature is primarily used for data reconstruction. Together, they constitute a source feature representation that can simultaneously support communication and computation. At this point, the encoder has completed feature compression of the input data.

[0037] To enable transmission in wireless communication systems, this invention performs bit-level quantization on the 1536-dimensional feature by uniformly quantizing it according to q bits per dimension, resulting in a source feature bit stream B1 with a length of 1536×8.

[0038] Step 2: Following the frame format of the target wireless communication protocol (such as IEEE 802.11n), add protocol fields for frame synchronization, length indication, and error checking before and after B1 to generate the first bit stream B2.

[0039] The 1536×8 bitstream obtained in step 1 is encapsulated according to the frame format of existing wireless communication protocols. Specifically, this bitstream is used as the payload, and protocol fields, including but not limited to frame header, frame trailer, synchronization sequence, length field, and check field, are filled before and after it to meet the requirements of standard data frame structure, so that the bitstream generated by this invention can be directly recognized and processed by the existing wireless communication physical layer and link layer.

[0040] The frame header is used by the receiver to synchronize and parse the frame structure, while the frame trailer or checksum field is used for frame integrity detection. Since this invention does not change the protocol field layout or its encoding method, it is seamlessly compatible with existing wireless standards (such as 802.11) without requiring modifications to the wireless hardware and protocol stack.

[0041] Step 3: Using the selected channel coding scheme (such as LDPC code with a code rate of 1 / 2), the first bit stream B2 is encoded, and redundancy check information is added to generate a second bit stream B3 with stronger anti-interference capability.

[0042] Specifically, the information is first divided into several segments according to a preset bit length. Then, each segment is channel-coded using a low-density parity-check code (LDPC code) to add redundant bits to the original information bits for error detection and correction at the receiver. The LDPC coding is performed according to a predetermined code rate and parity check matrix rules. Each encoded codeword is then combined sequentially to form the final transmitted bit stream, thereby enhancing the reliability of the bit stream in noisy channels and ensuring that the receiver can recover the original information through the corresponding decoding algorithm.

[0043] Step 4: Using a digital modulation method (such as 64QAM), the second bit stream B3 is mapped to a complex IQ symbol sequence S_tx and transmitted via the radio frequency front end.

[0044] Specifically, this experiment uses 64-order quadrature amplitude modulation (64QAM) for modulation, with every 6 bits mapped to a complex symbol, where the real and imaginary parts correspond to discrete values ​​of amplitude and phase, respectively. Through this modulation process, the discrete bit stream is converted into a continuous sequence of complex symbols to adapt to the transmission characteristics of the wireless channel and to provide input for demodulation and channel decoding at the receiver.

[0045] (2) Receiver processing flow: The receiving end receives the IQ symbol sequence S_rx after it has been transmitted through the channel and performs mode selection: depending on the application requirements (such as needing extremely low latency classification results or needing a complete image), it selects to enter either computation mode or communication mode.

[0046] Path 1: Calculation Mode Retrieves the prototype vector library corresponding to the transmitter encoder from local storage. Each prototype vector in this library corresponds to a data category, which is essentially a mathematical representation (e.g., class center) of the ideal IQ symbol sequence obtained after the complete transmitter processing (S101-S105) of each category's samples. Directly calculates the distance (e.g., Euclidean distance) between S_rx and each prototype vector in the IQ domain, and outputs the category with the smallest distance as the classification result. This path does not require executing subsequent communication mode steps.

[0047] Path 2: Communication Mode Step 5: Demodulate the IQ symbol sequence S_rx (e.g., 64QAM hard decision demodulation), determine the corresponding bit based on the position of the received symbol in the constellation diagram, and generate the received bit stream B3' (corresponding to B3 at the transmitting end).

[0048] Step 6: Perform channel decoding (such as LDPC decoding) on ​​B3', use the parity bits to correct transmission errors, remove redundant parity bits, and recover the first bit stream estimate B2' (corresponding to B2 at the sending end).

[0049] Specifically, the receiving end uses the same channel coding rules as the sending end to perform channel decoding processing on the demodulated bit stream, remove redundant parity bits and correct bit errors generated during transmission, thereby obtaining the recovered valid information bit stream, providing input for subsequent reconstruction or classification reasoning.

[0050] Step 7: Based on the known protocol format, identify and remove all protocol fields added in step S103 from B2' to recover the core source feature bitstream estimation B1' (corresponding to B1 of the sending end).

[0051] Specifically, the receiving end locates and deletes the frame header, trailer, and other protocol control fields according to the preset frame structure, thereby obtaining a bit stream containing only valid data information, which provides input for subsequent reconstruction or classification reasoning modules.

[0052] Step 8: Input B1' into the decoder that is paired with the encoder trained at the transmitting end. The decoder first parses B1' into 8-bit groups, reconstructs it into integer feature vectors, then maps it back to continuous feature space through inverse normalization, and finally reconstructs the estimated data X' through operations such as deconvolution, thus completing the communication.

[0053] Specifically, the receiving end uses the decoder trained in step 1, corresponding to the encoder, to map the net payload bitstream back to the original data space, completing the source decoding process and thus realizing end-to-end data communication reconstruction.

[0054] Comparison of experimental results: The experimental setup for this embodiment is as follows: Hardware and Datasets. All models in this invention were trained on a workstation equipped with an AMD 3950X CPU and an NVIDIA RTX A4000 GPU. The training setup used a 1×10... 3 The learning rate, batch size of 256, and epochs were used. We evaluated the performance of this invention on five datasets, including handwritten digits (MNIST), fashion objects (FashionMNIST), natural images (CIFAR-10), remote sensing images (EuroSAT), and medical images (BloodMNIST).

[0055] Experimental Setup. In a controlled experimental environment, we implemented the transmitter (Tx) and receiver (Rx) on the USRPX410 platform. The system operated at a frequency of 2.4 GHz with a bandwidth of 20 MHz, configured according to the 802.11n standard, and used 64-QAM modulation to generate a standard Wi-Fi signal. Both Tx and Rx were connected to omnidirectional antennas with a gain of approximately 15 dBi, and the default transmission distance was set to 5 m. We used EuroSAT as the default dataset with LDPC coding parameters of (324,648). To evaluate the effectiveness of this invention, we used classification accuracy, peak signal-to-noise ratio (PSNR), and bits per pixel (BPP) as metrics. PSNR is widely used in image and video compression, where values ​​above 25 dB generally indicate acceptable reconstruction quality, and values ​​above 30 dB signify high-quality reconstruction. BPP quantifies the average number of bits required per pixel and serves as a standard measure of compression efficiency, where lower values ​​correspond to higher compression ratios.

[0056] The communication and computation parallelism method, storage medium, and product based on IQ domain processing presented in this embodiment will be evaluated from the following n aspects: (1) Effectiveness evaluation of joint feature fusion between intermediate and terminal layers: To verify the effectiveness of the multi-level feature fusion strategy in this invention, we compared and evaluated the scheme of "constructing the bitstream using only high-level end features" with the scheme of "simultaneously introducing intermediate structural features and high-level end features for joint encoding". Figure 2(a) shows two images transmitted by the transmitter, taken from the EuroSAT and CIFAR-10 datasets respectively. Figure 2(b) shows the reconstruction effect at the receiver when only the fifth layer features of the encoder are transmitted. Figure 2(c) shows the reconstruction effect at the receiver when the third and fifth layer features of the encoder are transmitted simultaneously.

[0057] It can be observed that the reconstructed image in Figure 2(b) relying solely on high-level features at the end suffers from significant deficiencies in texture details, edge structure, and local shape preservation, exhibiting overly smoothed or structurally blurred appearance. In contrast, the reconstruction result in Figure 2(c) employing a multi-level feature fusion strategy shows significant improvements in contour, texture, and color consistency, more closely resembling the original input image. This phenomenon indicates that the local structural information contained in the intermediate layer features plays a crucial role in data reconstruction, while the high-level features at the end primarily contribute to global semantic expression. Combining both helps improve reconstruction quality while maintaining inference performance.

[0058] Furthermore, unlike traditional feature fusion networks that employ a full-layer jump-connection structure, this invention selects only mid-level features and high-level terminal features from the encoding network for fusion, avoiding the introduction of excessive redundant information and additional computational burden during transmission and reconstruction. On the one hand, this strategy preserves discriminative semantic features that contribute to classification; on the other hand, it preserves local structural information used for reconstruction, thereby reducing computational overhead and improving source encoding efficiency while ensuring reconstruction quality and inference performance.

[0059] (2) Reasoning time assessment: To evaluate the low-latency advantage of this invention, we simulated and compared the computation time of this invention with that of a traditional discrete architecture on five datasets using the same hardware parameters and trained models. In the discrete architecture, end-to-end latency includes IQ reception, demodulation, channel and source decoding, and CNN-based image classification. In contrast, this invention uses Euclidean distance to perform lightweight prototype classification directly in the IQ domain. Figure 1 As shown, the encoder generates 1536 features from the final convolutional layer. These features are quantized to 8 bits, encapsulated with a protocol field, and further protected by LDPC encoding with a configuration of (324,648), resulting in approximately 4,000 IQ data points per sample. Experiments were conducted on CPU and GPU platforms. Figure 3 The diagram shows a comparison of inference time between the proposed solution and the traditional discrete architecture on five datasets. The results show that the inference time of the proposed solution is only 0.18ms per sample on the CPU and only 8μs per sample on the GPU, which is two orders of magnitude faster than the discrete architecture, thus demonstrating its significant latency advantage.

[0060] (3) Comprehensive performance evaluation: We conducted comprehensive performance tests on five datasets (MNIST, FashionMNIST, CIFAR-10, EuroSAT, and BloodMNIST) and three QAM modulation schemes. Figure 4 The reconstruction quality and classification accuracy performance of this invention under different modulation orders are presented. Experimental results show that, under all datasets and modulation schemes, this invention can consistently maintain high inference accuracy and reconstruction quality: the classification accuracy is consistently no less than 90.5%, reaching a maximum of 98.5%; the PSNR of image reconstruction remains above 26.4 dB, reaching a maximum of 32.5 dB. These results demonstrate that the source coding mechanism based on multi-level feature fusion and IQ domain inference proposed in this invention can achieve robust end-to-end communication and computation integration performance under different data distributions and channel configurations.

[0061] (4) Compression ratio assessment: To evaluate the impact of the number of features (i.e., compression ratio) on reconstruction and classification performance, we tested the invention at seven BPP settings ranging from 1.5 to 4.5 with a step size of 0.5, and compared it with JPEG as a baseline method. Figure 5 The performance of this invention and JPEG under different compression ratios is demonstrated. As shown in the figures, although the PSNR of this invention is slightly lower than that of JPEG, it remains consistently above 30dB, indicating reliable reconstruction quality. Furthermore, the reconstruction quality improves with increasing BPP in both methods.

[0062] It is worth noting that, under all compression ratio settings, the classification accuracy of this invention is consistently significantly higher than that of JPEG, with the highest difference exceeding 62.4%. This is because the transformation and quantization process of JPEG weakens inter-class separability and intra-class consistency, thereby affecting downstream inference performance; while the multi-level feature fusion and bit-level quantization strategy of this invention can retain task-related discriminative information, thus maintaining a high classification accuracy at a lower bit cost.

[0063] In summary, this invention achieves a better overall trade-off between image fidelity and task-related performance.

[0064] (5) System adaptability evaluation under different LDPC encoding configurations: To verify the universality and robustness of this invention under different channel coding configurations, we selected six representative sets of LDPC code parameters (k,n) for performance evaluation. For example... Figure 6 As shown, the six configurations (configurations A to F) cover different combinations of code length and code rate, specifically including: configuration A (324, 648), configuration B (486, 648), configuration C (648, 1296), configuration D (972, 1296), configuration E (972, 1944), and configuration F (1458, 1944). Experimental results show that under all the above configurations, the classification accuracy of the system is consistently maintained at around 96%. This result indicates that the alignment method proposed in this application has good channel independence, meaning that the system's classification performance is insensitive to changes in the specific parameters of LDPC encoding and can adapt to various application scenarios with different code lengths and code rates.

[0065] (6) Applicability assessment of different protocols: To evaluate the applicability of this invention under different wireless communication standards, we tested five protocol configurations (denoted as A to E) on the EuroSAT dataset, including: 802.11n (20MHz), 802.11n (40MHz), 802.11ac (40MHz), 802.11ac (80MHz), and 802.11ax. Figure 7This paper demonstrates the classification performance and reconstruction quality of the present invention under the aforementioned protocol conditions. Experimental results show that the present invention maintains stable performance under different protocol bandwidths and modulation configurations, with classification accuracy remaining between 96.3% and 97.2%, and reconstruction quality PSNR remaining stable at approximately 28.2 dB. These results indicate that the present invention can be adapted to multiple Wi-Fi standards without redesigning for specific protocol structures, demonstrating good protocol versatility and practical deployment potential.

[0066] (7) Communication distance assessment: To evaluate the impact of communication distance on the performance of this invention, we adjusted the distance between the transmitter and receiver from 10m to 100m and measured the classification accuracy, bit error rate (BER) after LDPC decoding, and reconstruction quality (PSNR) respectively. Figure 8 The performance trend of this invention under different communication distances is demonstrated. Experimental results show that as the transmission distance increases, channel fading and noise enhancement lead to a gradual decline in performance: the classification accuracy decreases from 98.0% at 10m to 92.5% at 100m, and the BER after LDPC decoding increases accordingly; the PSNR of the reconstruction quality decreases from approximately 31dB to 22.1dB. Despite the performance degradation, this invention can still maintain a classification accuracy of over 90% and a reconstruction quality of over 22dB even at a transmission distance of 100m, indicating that this invention still has good robustness and usability over long communication distances.

[0067] (8) Multi-user performance consistency assessment: To verify the performance consistency of this invention in a multi-user environment, we selected five users with different IP addresses for testing. Figure 9 This paper demonstrates the classification accuracy and reconstruction quality performance of the present invention in a multi-user scenario. Experimental results show that the present invention maintains a stable classification accuracy (94.1% to 96.2%) and a reconstruction PSNR of approximately 28 dB across all users. These results indicate that the protocol alignment mechanism proposed in this invention can effectively adapt to multi-user concurrency conditions and maintain reliable and consistent system performance across different users.

[0068] (9) Robustness assessment in mobile scenarios: To verify the robustness of the invention in mobile scenarios, we tested five different speeds (5, 10, 15, 20 and 25 km / h) under the default experimental configuration. Figure 10 The performance of this invention in a mobile environment is demonstrated. Results show that increased speed has almost no impact on classification accuracy, and the PSNR remains stable across all speeds. This robustness is attributed to the standard communication mechanism employed in this invention, which compensates for channel changes caused by movement (such as synchronization, equalization, and error correction), thus maintaining reliable performance in mobile scenarios.

[0069] (10) Robustness assessment of the channel environment: To verify the robustness of the invention under different channel conditions, we tested the classification accuracy of five datasets within the SNR range of -20dB to 10dB. Figure 11 The performance of this invention under different channel environments is demonstrated. Results show that classification accuracy increases rapidly with increasing SNR; when SNR exceeds -4dB, the accuracy of all datasets exceeds 80%. This result demonstrates that the invention maintains good robustness even in the presence of channel noise.

[0070] (11) Limit performance evaluation: To demonstrate the flexibility of this invention under different application requirements, we evaluated its performance in a single-task scenario. By adjusting the source encoding module, this invention can be configured into two modes: one focusing on high-fidelity reconstruction and the other on high-precision classification. We tested the system on five standard datasets. Figure 12 The performance under different task objectives is demonstrated. Results show that in reconstruction-only mode, the PSNR exceeds 30 dB; and in classification-only mode, the classification accuracy exceeds 92%. These results demonstrate that the present invention can be flexibly adjusted according to task requirements to achieve high-fidelity reconstruction or high-precision classification, highlighting its versatility.

[0071] (12) Applicability assessment for complex scenarios: To verify the applicability of this invention in various complex scenarios, we conducted experiments in three environments: outdoor roads, indoor rooms, and three-dimensional space, selecting 10 different test points in each environment. Figure 13 shows the performance of this invention in these complex scenarios. The results show that regardless of environmental changes, this invention maintains high classification accuracy (95% to 97%) and stable reconstruction quality (PSNR 27 to 28 dB). These results demonstrate the robustness of this invention under different propagation conditions, ensuring reliable classification and data reconstruction capabilities in practical deployments.

[0072] The preferred embodiments of this disclosure have been described in detail above with reference to the accompanying drawings. However, this disclosure is not limited to the specific details of the above embodiments. Within the scope of the technical concept of this disclosure, various simple modifications can be made to the technical solutions of this disclosure, and these simple modifications all fall within the protection scope of this disclosure.

[0073] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, this disclosure will not describe the various possible combinations separately.

[0074] Furthermore, the various implementation methods disclosed in this solution can be combined in any way, as long as they do not violate the spirit of this disclosure, they should also be regarded as the content invented by this disclosure.

Claims

1. A parallel communication and computation method based on IQ domain processing, characterized in that, This includes processing at the sending end and processing at the receiving end; The sending end processing includes: S1. Input the input data into the encoder, extract the feature representation of the input data, quantize the feature representation, and generate the source feature bitstream; S2. Add a protocol field to the source feature bitstream obtained in S1 to generate the first bitstream; S3. Perform channel coding on the first bitstream obtained in S2 to generate the second bitstream; S4. Modulate the second bitstream obtained in S3 into IQ symbols and send it; The receiving end processing includes: receiving the IQ symbol obtained by S4, and selecting a calculation mode or communication mode according to requirements; The specific computation mode involves using pre-stored prototype vectors to directly perform classification reasoning on the IQ symbols in the IQ domain and outputting the classification results. The communication modes include: S5. Demodulate the IQ symbols to obtain the received bit stream; S6. Perform channel decoding on the received bitstream obtained in S5 to recover the first bitstream estimate; S7. Remove the protocol field from the first bitstream estimate obtained in S6 to recover the source feature bitstream estimate; S8. Input the source feature bitstream estimate obtained in S8 into the decoder, reconstruct and output the estimated data of the input data.

2. The parallel communication and computation method based on IQ domain processing as described in claim 1, characterized in that, The encoder in S1 is a multi-layer convolutional neural network.

3. The parallel communication and computation method based on IQ domain processing as described in claim 1, characterized in that, The protocol fields in S2 include physical layer header information, MAC layer frame header information, and verification information; wherein, the physical layer header information includes at least a synchronization sequence or signaling field, and the MAC layer frame header information includes at least a frame control field or an address field.

4. The parallel communication and computation method based on IQ domain processing as described in claim 1, characterized in that, The channel coding in S3 uses LDPC codes.

5. The parallel communication and computation method based on IQ domain processing as described in claim 1, characterized in that, The modulation in S4 is QAM modulation or PSK modulation, which maps the second bit stream to complex symbols.

6. The parallel communication and computation method based on IQ domain processing as described in claim 5, characterized in that, The demodulation in S5 is either hard-decision demodulation or soft-decision demodulation, which corresponds to the modulation method used in S4.

7. The parallel communication and computation method based on IQ domain processing as described in claim 4, characterized in that, The channel decoding in S6 is a decoding algorithm corresponding to the channel coding method used in S3.

8. The parallel communication and computation method based on IQ domain processing as described in any one of claims 1-7, characterized in that, The prototype vectors used in the computation mode are specifically obtained in advance and stored at the receiving end in the following way: During the training phase, the sample data of each predetermined category are processed sequentially by the sending end S1~S4 to obtain the corresponding reference IQ symbols, and the prototype vectors representing each category are trained and calculated based on the reference IQ symbols.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the communication and computation parallel method based on IQ domain processing as described in any one of claims 1-8.

10. A computer program product, characterized in that, Includes a computer program / instruction, which, when executed by a processor, implements the communication and computation parallelization method based on IQ domain processing as described in any one of claims 1-8.

Images