High-frequency signal acquisition and compression processing method based on edge calculation

Abstract

The invention relates to the field and discloses a high-frequency signal acquisition and compression processing method based on edge calculation. Comprising the following steps: signal acquisition and complexity analysis: performing spectrum analysis on a high-frequency signal of target equipment at an edge node, and adaptively adjusting a sampling rate and a sampling window length according to signal complexity; feature extraction and cooperative compression: extracting a time-frequency feature vector of the signal, identifying a similar signal interval according to a coherence threshold between nodes, executing cross-node differential compression on the similar interval, and only transmitting a change signal segment; compressed code stream generation and feature restoration: adding spectrum feature information to the compressed code stream for local restoration of the sink node when the energy deviation exceeds a threshold value; model self-learning and dynamic evolution are carried out, and the cloud updates compression model parameters according to error information reported by the nodes and distributes the compression model parameters according to scenes; data are transmitted credibly, and cloud edge bidirectional verification is realized by using Hash fingerprints and digital signatures.

Description

Technical Field

[0001] This invention relates to the field of signal acquisition and data compression processing, specifically to a high-frequency signal acquisition and compression processing method based on edge computing. Background Technology

[0002] Industrial production sites commonly deploy numerous high-frequency sensors for vibration, current, and acoustics to monitor equipment operating status. These signals are sampled at high frequencies, generate large amounts of data, and update rapidly. Traditional edge-based simple acquisition and cloud-based centralized analysis methods are prone to causing link congestion, increased storage pressure, and analysis delays in scenarios with multiple concurrent devices, making it difficult to support timely fault warnings and status assessments.

[0003] In existing technologies, most solutions use a fixed sampling rate to collect high-frequency signals, then compress them locally at a single edge node or gateway before uploading them to the cloud. This approach has the following problems: First, the sampling parameters are independent of signal complexity, resulting in a large amount of redundant data during stable signal phases and insufficient resolution during abrupt changes. Second, there is a lack of coordination between edge nodes; compression is performed independently at each node, failing to utilize the correlation between signals, leading to the repeated uploading of similar data and consuming bandwidth and computing resources. Third, the compression and transmission links are mostly unidirectional processes; if errors occur during encoding or transmission, there is a lack of effective reconstruction verification and repair mechanisms, making it difficult to guarantee the reliability of the reconstructed signal.

[0004] As the scale of edge nodes expands and the frequency of cloud-edge interactions increases, the integrity and reliability of acquired data during transmission become increasingly prominent issues. Some systems still rely on simple verification or centralized recording methods, making it difficult to detect data tampering or loss in intermediate stages in a timely manner. Overall, existing technologies still have room for improvement in areas such as adaptive acquisition of high-frequency signals, multi-node collaborative compression, reconstruction quality assurance, and reliable transmission. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a high-frequency signal acquisition and compression processing method based on edge computing, thereby solving the technical problems existing in the prior art.

[0006] The above-mentioned technical objective of the present invention is achieved through the following technical solution: A method for high-frequency signal acquisition and compression processing based on edge computing includes the following steps: Signal acquisition and complexity analysis: High-frequency signals from the target device are acquired at edge nodes, and the sampled signals are subjected to spectral analysis to calculate the instantaneous spectral entropy to obtain a complexity index. The sampling rate and sampling window length are adaptively adjusted according to the changes in the complexity index so that high-complexity signals are acquired at a higher sampling rate and low-complexity signals are acquired at a lower sampling rate. Feature extraction and collaborative compression: The sampled signal is input into the feature extraction module to extract the time-frequency feature vector; multiple edge nodes calculate the inter-node signal correlation coefficient matrix based on the time synchronization signal, and identify similar signal intervals according to the coherence threshold; cross-node differential compression is performed on the similar signal intervals to transmit only the changing signal segments; Compressed bitstream generation and reversible repair: spectral feature information for reconstruction error detection is embedded in the compressed bitstream. When the energy retention rate of the reconstructed signal is lower than the threshold, the edge nodes perform local repair based on the feature information to compensate for signal distortion. The model self-learning and evolution process uploads the error distribution generated during compression and reconstruction to the cloud model repository and updates the parameters of the edge compression model based on the error gradient. The cloud updates the model according to different signal feature scenarios, realizing the self-learning and dynamic evolution of the edge compression model. Trusted data upload involves generating hash fingerprints and verifying the integrity of compressed data blocks to ensure the security and trustworthiness of compressed data during transmission from the edge to the cloud.

[0007] Preferably, spectral analysis is performed on the sampled signal at the edge node to obtain energy distribution information for each time slice; The complexity index of the signal is calculated based on the degree of change in energy distribution, and the complexity index is compared with a preset complexity threshold. When the complexity index is higher than the threshold, the control sampling module increases the sampling rate and shortens the sampling window length. When the complexity index is lower than the threshold, the control sampling module reduces the sampling rate and extends the sampling window length; By adaptively adjusting the sampling rate and sampling window, the sampling signal output by the edge node is kept in balance between time resolution and frequency resolution.

[0008] Preferably, the feature extraction and collaborative compression steps include: The adaptively sampled signal is input into the feature extraction module, which performs time-domain and frequency-domain analysis on the signal to obtain a time-frequency feature vector that reflects the signal amplitude changes and spectral distribution. Multiple edge nodes receive a unified time reference through a time synchronization module to align the timestamps of signals collected by each node; Based on the synchronized signal data, the correlation coefficient matrix between the signals of the nodes is calculated, and the similarity interval of the signals of each node is determined according to the preset coherence threshold. When a similar interval is determined, the collaborative compression module selects a node as a reference node, compresses and encodes the difference data between the signals of other nodes and the signal of the reference node, and transmits only the difference data fragment to the aggregation node. The output of the differential compression is used as input data for subsequent compressed bitstream generation and error correction steps.

[0009] Preferably, the compressed bitstream generation and feature repair steps include: The differentially compressed signal is encoded at the edge nodes to generate a compressed bitstream; Spectral feature information for reconstruction detection is appended to the compressed bitstream. The spectral feature information includes the dominant frequency position of the signal before compression and the corresponding energy percentage. At the aggregation node, the received compressed bitstream is decoded and reconstructed, and the energy difference between the reconstructed signal and the reference signal is calculated. When the energy difference exceeds the preset error threshold, the edge node is triggered to call the attached spectral feature information to perform local signal repair and compensate for the main frequency component of the reconstructed signal. The repaired signal data serves as input for the model's self-learning and dynamic evolution steps.

[0010] Preferably, the model self-learning and dynamic evolution steps include: Error information obtained from edge nodes during compression and reconstruction is uploaded to the cloud model repository; The cloud-based model repository calculates the parameter update amount based on the error information uploaded by each edge node, and adjusts the weight parameters of the compressed model according to the update amount.

[0011] Preferably, when the updated compression model meets the preset stability conditions, the model is distributed to each edge node for replacement; When the model is distributed in the cloud, edge nodes are classified into scenarios based on signal complexity distribution, spectral characteristics differences, or device operation type. Corresponding compression model parameters are assigned to nodes in different scenarios to maintain the convergence and adaptability of the compression model under different operating environments.

[0012] Preferably, the trusted data transmission step includes: After generating compressed data blocks at the edge nodes, a hash fingerprint value is calculated for each data block, and a digital signature is generated based on the node identity information. The hash fingerprint value and digital signature are recorded together in the ledger module of the edge node.

[0013] Preferably, when the compressed data block is uploaded to the cloud, the cloud verifies whether the hash fingerprint of the received data matches the fingerprint value in the ledger record module. If the verification is successful, the integrity of the transmitted data is confirmed, and the record is synchronously written to the cloud ledger. The ledger recording module is a lightweight distributed structure that supports bidirectional verification and record updates between edge nodes and the cloud.

[0014] In summary, the present invention has the following main beneficial effects: First, an adaptive sampling mechanism based on signal complexity is introduced at the edge nodes to automatically adjust the sampling rate and sampling window length according to the degree of signal variation under different operating conditions. High sampling accuracy is maintained for high-frequency signals with significant variations, while the sampling density is reduced for signals with relatively stable waveforms. This reduces the amount of data that needs to be processed and uploaded while ensuring that key waveforms and spectral details are not lost, alleviating the pressure on edge nodes and network links. This approach is suitable for online monitoring scenarios involving multiple nodes and long-term operation.

[0015] Secondly, cross-node collaborative differential compression is employed in the feature extraction and compression stages, and spectral feature information such as the main frequency and energy proportion are included in the compressed bitstream. The system utilizes the correlation between signals from different nodes, uploading only the data that differs from the reference node. At the aggregation end, necessary local repairs are performed on the reconstructed signal based on the spectral features. Compared to the method of independent compression and sequential uploading by each node, this scheme can reduce redundant data transmission and improve the quality of the compressed signal reconstruction under the same bandwidth conditions, making it more suitable for industrial field network environments with limited bandwidth and a large number of nodes.

[0016] Furthermore, this invention reports compression and reconstruction errors through edge nodes, establishes a model repository in the cloud for continuous updates to the compression model, and distributes adapted models based on signal characteristics and device types. Simultaneously, during data upload, hash fingerprints, digital signatures, and lightweight ledger records are used to verify the integrity and origin of the compressed data. Therefore, on the one hand, the compression model maintains good adaptability and compression accuracy as operating conditions change; on the other hand, it improves the security and traceability of data in the transmission link, providing a reliable data foundation for subsequent fault diagnosis and maintenance decisions. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1 refer to Figure 1 A high-frequency signal acquisition and compression processing method based on edge computing includes the following steps: Signal acquisition and complexity analysis: High-frequency signals from the target device are acquired at edge nodes, and the sampled signals are subjected to spectral analysis to calculate the instantaneous spectral entropy to obtain a complexity index. The sampling rate and sampling window length are adaptively adjusted according to the changes in the complexity index so that high-complexity signals are acquired at a higher sampling rate and low-complexity signals are acquired at a lower sampling rate. Feature extraction and collaborative compression: The sampled signal is input into the feature extraction module to extract the time-frequency feature vector; multiple edge nodes calculate the inter-node signal correlation coefficient matrix based on the time synchronization signal, and identify similar signal intervals according to the coherence threshold; cross-node differential compression is performed on the similar signal intervals to transmit only the changing signal segments; Compressed bitstream generation and reversible repair: spectral feature information for reconstruction error detection is embedded in the compressed bitstream. When the energy retention rate of the reconstructed signal is lower than the threshold, the edge nodes perform local repair based on the feature information to compensate for signal distortion. The model self-learning and evolution process uploads the error distribution generated during compression and reconstruction to the cloud model repository and updates the parameters of the edge compression model based on the error gradient. The cloud updates the model according to different signal feature scenarios, realizing the self-learning and dynamic evolution of the edge compression model. Trusted data upload involves generating hash fingerprints and verifying the integrity of compressed data blocks to ensure the security and trustworthiness of compressed data during transmission from the edge to the cloud.

[0020] Signal Acquisition and Complexity Analysis: This module includes the following functional units: The signal acquisition unit is used to acquire high-frequency analog signals from the target device and perform analog-to-digital conversion. The spectrum analysis unit is used to perform short-time spectrum transformation on the sampled signal to obtain energy distribution information for each time slice. The complexity calculation unit is used to calculate the signal complexity index based on the degree of change in energy distribution; The sampling control unit is used to adaptively adjust the sampling rate and sampling window length based on the relationship between the complexity index and a preset threshold.

[0021] The signal acquisition unit at the edge node uses an analog-to-digital converter to digitally sample the high-frequency analog signals generated by the target device. The sampled data is stored in a local buffer at fixed time intervals to ensure data continuity and no loss.

[0022] The spectrum analysis unit uses Short-Time Fourier Transform (STFT) or similar time-frequency analysis algorithms to calculate the spectral intensity of the buffered signal frame by frame, obtaining the energy distribution function of each time slice.

[0023] in, This represents the time slice index, used to identify the time position corresponding to the current calculation window; This represents the frequency index, corresponding to the position of the transformed discrete frequency component. For time slices In frequency The energy value at a given point is calculated as the square of the spectral amplitude at that point. These are the complex spectral coefficients obtained through short-time Fourier transform; To reflect the relative distribution characteristics of energy in the frequency domain, the energy spectrum is normalized to obtain the normalized form of the energy distribution: ; in, This is a normalized energy distribution; the numerator in this formula is the energy at a certain frequency, and the denominator is the sum of the energies of all frequencies within that time slice. The value range is 0 to 1, and it satisfies the normalization condition. =1.

[0024] Normalized energy distribution This describes the proportion of energy of each frequency component of the signal within a given time slice. It is a fundamental parameter for subsequent calculations of spectral entropy and complexity exponent. If the signal energy in a certain time slice is concentrated in a few frequencies, it indicates that the signal structure is simple. The signal energy is relatively large at some frequencies and approaches zero at others; if the signal energy is uniformly distributed across multiple frequencies, then... If the values ​​are relatively close, it indicates a high level of signal complexity. Complexity calculation unit, through It can quantitatively reflect the spectral energy dispersion of a signal, providing a basis for calculating the instantaneous spectral entropy in subsequent steps.

[0025] in, for The spectral entropy is obtained over time; and the entropy value is normalized to obtain a complexity index ranging from [0,1]. To suppress the influence of instantaneous fluctuations, exponential smoothing can be applied to the entropy sequence to obtain the final complexity index. The smoothing coefficient ranges from 0.6 to 0.95 and is used to control the response speed.

[0026] To ensure the complexity index The judgment results are accurate; a preset threshold set needs to be established before the system is deployed. It is used to distinguish between high and low ranges of signal complexity.

[0027] The threshold can be determined using statistical analysis: Multiple sets of high-frequency signal samples are collected under typical operating conditions, and the distribution of the complexity index is calculated for each set of samples; after merging the complexity results of all samples, the 25th percentile is taken as the low complexity threshold. Take the 75th percentile as the high complexity threshold. .

[0028] If the average value of the complexity index drifts significantly during long-term operation, the threshold can be fine-tuned based on the average value over a recent period to adapt to changes in signal characteristics. In practical applications, complexity index The value range is 0 to 1, and the threshold set is generally set as follows: =0.3~0.4; =0.6~0.7.

[0029] By setting thresholds, the system can accurately distinguish between high-complexity and low-complexity signals, and achieve adaptive adjustment of the sampling rate and sampling window, thereby taking into account both the time resolution and frequency resolution of the signal.

[0030] The sampling control unit will increase the complexity index With preset threshold set This is used to determine the current complexity of the signal. when > When the signal complexity is high, the control sampling module increases the sampling rate and shortens the sampling window. when < When the signal complexity is low, the control sampling module reduces the sampling rate and extends the sampling window. when When the sample is within the threshold range, the current sampling parameters remain unchanged. During the sampling rate reduction process, the system synchronously adjusts the cutoff frequency of the anti-aliasing filter to avoid spectral folding.

[0031] When the sampling rate is increased, the window overlap rate is kept constant to maintain signal continuity and consistency of time-frequency resolution.

[0032] The sampled signal output by the edge nodes has high temporal resolution in high-complexity stages and high frequency resolution in low-complexity stages. This signal data is directly input into the feature extraction and collaborative compression module, achieving seamless integration between the modules.

[0033] Feature extraction and collaborative compression module: Feature extraction and collaborative compression are implemented on edge computing nodes to identify signal similarity across multiple edge nodes. Cross-node differential compression is performed based on a coherence threshold to reduce data redundancy and transmission bandwidth consumption. This module is integrated with the signal acquisition and complexity analysis module, and the input is a high-frequency signal sequence after adaptive sampling. The feature extraction and collaborative compression module includes the following functional units: The feature extraction unit is used to perform time-domain and frequency-domain analysis on the sampled signal and extract time-frequency feature vectors that reflect the changes in signal amplitude and spectral distribution. The time synchronization unit is used to receive a unified time reference signal and align the timestamps of the signals collected by each edge node. The coherence identification unit is used to calculate the correlation coefficient matrix between nodes based on the time-synchronized signals and to identify similar signal intervals according to a preset coherence threshold. The cooperative differential compression unit is used to select a reference node within a similar signal range, perform differential compression on the signals of the remaining nodes, and transmit only the changed segments.

[0034] The adaptively sampled signal is input into the feature extraction unit, let the first... Each edge node in the time slice The sampling sequence within is ;in , The time slice length; Perform a short-time Fourier transform on each node:

[0035] in, For window functions; Number of transformation points; Frequency index; Indicate the time slice index; obtain the energy spectrum. And normalize to obtain the energy distribution:

[0036] in, Indicates time slice Inner Frequency point energy percentage; The following features are extracted in the time and frequency domains: Root mean square value:

[0037] Spectral centroid:

[0038] Spectral entropy:

[0039] Forming nodes Time-frequency eigenvectors:

[0040] in, For the first Each frequency point, Sampling rate; Indicates signal strength; Reflects the position of the main frequency; This vector represents the spectral dispersion; it reflects the amplitude and spectral changes of the signal at different time slices and is used to determine the correlation between subsequent nodes.

[0041] Each edge node receives a unified time reference signal (NTP, GPS, or IEEE1588) through a time synchronization unit to align the timestamps of the sampled signals. If slight drift exists, linear interpolation is used to correct it, ensuring that all time slices are within the same time frame. They correspond to the same physical time.

[0042] Calculate the correlation coefficient matrix for the synchronized signal:

[0043] in, The mean; This indicates the correlation between node i and node j in time slice t; when ≥ And continue for no less than Within a given time slice, the signal intervals are identified as similar signal intervals; in, The coherence threshold is selected as 0.75–0.85; The coherent duration window size is selected as 3 to 5; to avoid frequent switching, the system employs a hysteresis-based decision mechanism, with the exit condition being... < -0.05.

[0044] Within the similar signal interval, from the current node set Select a reference node The selection rule is based on the principle of minimizing residual energy.

[0045] in, This represents the mean square error of the signals at the two nodes; For each non-reference node Calculate the residual signal: And perform discrete cosine transform: Only retain the first few coefficients that account for 90% to 95% of the energy, or keep them fixed. M The largest coefficient is used for quantization and entropy encoding: Only if the residual energy satisfies: Only then will the system transmit the corresponding compressed segment, in which The residual energy threshold is set (preferably 2% to 5% of the reference node signal energy). If the residual energy is below the threshold, no transmission is performed to avoid reporting invalid data; the output compressed segment... Reference node identifier They are then fed into the compressed bitstream generation and error correction module.

[0046] Compressed bitstream generation and feature restoration: Compressed bitstream generation and feature restoration are achieved within an edge-convergence architecture. The input consists of the reference node identifier and residual compressed fragment, while the output is the restored time-domain signal data, which is then used for subsequent model self-learning and dynamic evolution. Reference node identifier and its original time slice signal ; Each non-reference node residual compressed fragment Corresponding time slice index ; The spectral feature information packet is added to each time slice along with the bitstream, and includes the main frequency position and energy percentage; Repaired reconstructed signal As input for the model's self-learning and dynamic evolution steps; It consists of 1) a code stream generation and packaging unit; 2) a feature information embedding unit; 3) a decoding and reconstruction unit; 4) an error detection unit; and 5) a local spectrum repair unit.

[0047] For residual signals The compressed segment is obtained through transformation, coefficient selection, quantization, and entropy coding:

[0048] in, This is the discrete cosine transform. Select the retention factor based on energy percentage or number; Scalar quantization; Entropy coding; For compressing bitstreams; To perform reconstruction detection and necessary local repairs at the aggregation end, spectral feature information packets are embedded along with each time-slice bitstream at the edge side. :

[0049] in, The number of clock frequencies recorded; For the first The position of the dominant frequency is determined by the reference node or the peak value of the power spectrum of the node; Take 0-1, the first The percentage of each main frequency energy is defined as:

[0050] Pack Attached as a lightweight structure to the bitstream header, containing a time slice index. Reference node identifier Sampling rate, window length; The main frequency and energy ratio can be estimated by the reference node and then distributed as a unified value, or they can be estimated locally by each node. In this embodiment, it is preferred that the reference node uniformly estimates them in the coherent interval to ensure consistency.

[0051] The aggregation node receives the original fragment from the reference node. Residual fragments at each node Decode the residual fragments, and then reconstruct the nodes according to the difference relationship. Time-domain signal:

[0052] To determine whether local repair needs to be triggered, the energy difference or energy retention rate between the reconstructed signal and the reference signal is calculated: Energy difference:

[0053] Energy retention rate:

[0054] in, , At that time, local repair was initiated. This is the error threshold; The energy retention rate threshold is used; if the reconstructed energy deviates significantly from the reference level, it indicates that there is information loss or quantization deviation in the compression, transmission, and decoding links during this time slice, and targeted compensation is needed using the embedded master frequency information.

[0055] After the aggregation node receives the compressed bitstream uploaded by the edge nodes, the system first decodes and reconstructs the data. Specifically, the aggregation node decodes and restores the compressed residual fragments of each slave node into the corresponding time-domain signal based on the data from the reference node. In this way, the system can recover the original signal profile of each node.

[0056] Subsequently, the aggregation node performs energy detection on the decoded signal. The system calculates the energy difference between the reconstructed signal and the reference signal. If the difference exceeds a set threshold, it indicates that some valid information may have been lost during compression or transmission. In this case, the system performs local repair based on the spectral feature information packets transmitted along with the bitstream.

[0057] The spectral characteristic information packet here is actually a lightweight data segment appended by the edge nodes during compression to describe the main frequency components of the original signal. The packet contains the center frequency of each dominant frequency and the corresponding energy percentage. For example, if the dominant frequency of the signal is around 12 kHz, accounting for 30% of the total energy, the system will append this information as a reference value to the compressed data header.

[0058] When the sink node detects a large energy difference, it reads the dominant frequency parameter from the data packet, finds the corresponding frequency band in the spectrum of the reconstructed signal, and appropriately amplifies the signal within this band to restore its energy to the target level indicated in the data packet. This method is equivalent to compensating for the lost dominant frequency component within a local range without altering other unrelated frequency bands, thus repairing the signal without introducing new noise or distortion.

[0059] To prevent over-repair, the system sets an upper limit on the gain, for example, a maximum amplification of no more than 6 dB. When multiple main frequencies are close together, the system automatically merges the repair ranges to prevent spectral abrupt changes caused by overlapping amplification. After repair, the system then restores these signal segments into a continuous time-domain signal through superposition and smooth splicing.

[0060] Throughout the process, the energy threshold and main frequency compensation parameters can be adaptively adjusted. For example, the error threshold is typically set between 5% and 15%, and the energy retention rate threshold can be set between 85% and 95%. This ensures that the system maintains good signal restoration capabilities even at high compression rates, while also avoiding the computational burden caused by frequent triggering of repairs.

[0061] Through the above steps, the convergence node finally obtains the repaired high-fidelity signal, which is then passed to the model self-learning and dynamic evolution for subsequent error statistics and model parameter optimization.

[0062] Model self-learning and dynamic evolution: Building upon compressed bitstream generation and feature repair, a model self-learning and dynamic evolution mechanism is further proposed to achieve continuous optimization and scene-adaptive updates of the edge-end compressed model. By constructing a parameter feedback path between the cloud model repository and edge nodes, distributed learning and collaborative evolution of model parameters are achieved, thereby maintaining the convergence and generalization ability of the compressed model under different device operating environments.

[0063] This step mainly consists of the following units: Error acquisition and reporting unit: Deployed at each edge node, used to record error information in real time during the compression and reconstruction process; Cloud-based model repository unit: used to receive error data uploaded from multiple nodes, calculate parameter update amounts, and perform centralized optimization; Model distribution and replacement unit: used to distribute updated versions to each node after the model has stabilized; Scene classification and adaptive configuration unit: used to assign adaptive models to nodes based on different signal characteristics and device types.

[0064] During differential compression and feature restoration at edge nodes, the system calculates the error metric between the compressed and reconstructed signals, primarily including mean square error (MSE) and energy retention rate (EPR). The error signal is defined as follows:

[0065] in, For nodes The original signal; The corresponding reconstruction signal; Sample-level error signals reflect reconstruction deviations; Calculate the mean square error based on the error signal: and energy retention rate

[0066] The system uses time slices as units to... , Indicators such as compression ratio and clock frequency deviation are encapsulated into an error statistics vector: ,in This refers to the compression ratio; To reconstruct the difference between the main frequency of the reconstructed signal and the original signal, this error statistical vector is periodically uploaded to the cloud model repository via a lightweight message channel. The upload frequency can be dynamically adjusted according to bandwidth and model training rhythm, typically set to synchronize once every 30 to 60 minutes.

[0067] The cloud-based model repository aggregates error data uploaded from various edge nodes, performing parameter updates and model evolution. A centralized parameter vector is maintained in the cloud. ,in This is the set of weights for the current compression model. This represents the number of model parameters; after receiving error data from multiple nodes, the cloud calculates the global gradient update amount.

[0068] in, The number of nodes participating in the upload; For nodes Weighting coefficients; For nodes The local loss function is generally defined as:

[0069] in, , These are weighting coefficients used to balance distortion and energy preservation; Cloud-based press-to-update model parameters:

[0070] in, Using the learning rate, after the update, the system calculates the average error convergence trend of the model across multiple signal types using the validation set. Once the model meets the stability conditions, the cloud model repository encapsulates the updated weight package into a new version of the model and distributes it to each edge node.

[0071] The distribution is carried out in batches to prevent bandwidth surges caused by network-wide synchronization. Each node executes the following process upon receiving the new model: Perform local consistency checks to ensure version numbers match; Load the new model in the background and perform transition tests (i.e., run the old and new models simultaneously for a period of time and compare the output differences). When the new model outperforms the old model and the system does not detect any anomalies, replace the old model and update the local cache.

[0072] Trusted data upload: This ensures that compressed data generated by edge nodes is not tampered with or omitted during the upload process to the cloud, and enables integrity verification and traceability recording in the cloud. By setting up hash fingerprints, digital signatures, and distributed ledger modules, bidirectional trusted verification of the data transmission process is achieved.

[0073] After the edge nodes complete the generation of the compressed bitstream, the system divides the compression result into multiple independent data blocks in chronological order. After each data block is generated, the edge node first performs a hash operation on its content to obtain a unique fingerprint value. This fingerprint is equivalent to the identity label of the data block, which can determine whether the data has been modified in subsequent verification.

[0074] Subsequently, the system digitally signs the generated hash fingerprint based on the node's own identity information. The digital signature is encrypted using the node's private key, and only a cloud system with the corresponding public key can verify the validity of the signature, thereby ensuring the authenticity and trustworthiness of the data source and preventing forgery or impersonation during uploading.

[0075] After the hash fingerprint and digital signature are generated, the edge node writes both pieces of information into its local ledger module. The ledger module records the fingerprint, signature, node number, and generation time of each data entry in chronological order. Records are linked in a chain, meaning each subsequent record contains the hash digest of the preceding record, forming an immutable chain. This structure ensures the integrity of the ledger records; even if a single data block is deleted or modified, the chain verification relationship will be broken and detected by the system.

[0076] When an edge node uploads a compressed data block to the cloud, the cloud recalculates the hash fingerprint of the received data and compares it with the fingerprint stored in the node's ledger. If they match, it means the data has not been modified during transmission. Subsequently, the cloud verifies the digital signature using the node's public key to confirm that the uploaded data indeed originated from the corresponding node. Only when both the hash fingerprint and signature verification pass, does the system mark the data block as valid and synchronously write the record to the cloud ledger.

[0077] The cloud ledger and edge ledger use a two-way synchronization method. Every set time interval (e.g., 5 to 10 minutes), the system automatically performs two-way verification. The cloud randomly selects a portion of records and sends a verification request to the edge nodes. The nodes return the corresponding hash digests for comparison. If the comparison results match, it indicates that the ledger status is synchronized and the data is valid. If they do not match, a retransmission or anomaly alarm mechanism is triggered.

[0078] To improve operational efficiency, the ledger module in this embodiment adopts a lightweight design, storing only necessary information such as data summaries and signatures, without saving all compressed data content, thereby reducing storage pressure and improving synchronization speed.

[0079] Through this structure, the system can form a reliable data transmission loop between the edge and the cloud. Even if there is network latency or node downtime in the transmission link, post-event data verification and traceability can be achieved through ledger records.

[0080] During the signal acquisition and complexity analysis phase, edge nodes acquire high-frequency signals from the target device in real time. The nodes perform spectral analysis on the sampled signals, calculating the energy distribution of the signal across different time slices, and obtaining a complexity index reflecting the degree of signal variation. The system automatically adjusts the sampling rate and sampling window length based on changes in the complexity index, allowing high-complexity signals to be acquired at a higher sampling rate and stable signals at a lower sampling rate. This significantly reduces redundant sampling while ensuring that critical details are not lost.

[0081] The beneficial effects are that this adaptive sampling mechanism achieves a dynamic balance between time resolution and frequency resolution, enabling the acquisition system to take into account both signal integrity and data economy, and avoids the bandwidth waste and key information distortion problems that exist in traditional fixed sampling methods.

[0082] Next, the system proceeds to the feature extraction and collaborative compression stage. The adaptively sampled signal is input into the feature extraction module, where the system extracts feature parameters reflecting the signal's characteristics from both the time and frequency domains, such as the root mean square value, dominant frequency position, and spectral entropy. These features not only describe the signal's energy structure but also reflect the degree of similarity between signals at different nodes.

[0083] Multiple edge nodes receive a unified time reference through a time synchronization module, align the acquired signals in time, and then calculate the correlation coefficient matrix between nodes. A coherence threshold is used as the criterion to identify similar signal intervals. Within these similar intervals, the system automatically selects the node with the most representative signal change as the reference node, performs differential compression on the signals from other nodes, and transmits only the portions that differ from the reference node.

[0084] The beneficial effects are as follows: this collaborative compression method can eliminate duplicate data among multiple nodes to the greatest extent, significantly reducing the overall transmission volume while ensuring reconstruction accuracy. Compared with traditional independent compression methods, this invention achieves collaborative compression and network load reduction in distributed edge systems through a cross-node differential mechanism.

[0085] After compression, the system enters the compressed bitstream generation and feature restoration stage. Edge nodes encode the compression result into a bitstream and append characteristic information such as the signal's main frequency and energy percentage to the bitstream header. Upon receiving the compressed bitstream, the aggregation node decodes the data and restores the signal. When the system detects that the energy retention rate of the reconstructed signal is lower than a preset threshold, it automatically invokes the appended spectral feature information to perform local restoration, compensating for lost energy components in the main frequency neighborhood.

[0086] The beneficial effects are as follows: through this self-detection and local repair mechanism, the signal restoration quality can be significantly improved without increasing the transmission burden, and the compressed signal can still maintain high fidelity under high compression ratio. It is particularly suitable for application scenarios with high requirements for spectrum integrity, such as industrial vibration and acoustic monitoring.

[0087] The system then enters the model self-learning and dynamic evolution phase. Each edge node records its error information during compression and reconstruction, such as mean squared error, energy retention rate, and compression ratio, and periodically uploads it to the cloud model repository. The cloud aggregates error data from multiple nodes, calculates the update amount of model parameters based on this data, and performs self-learning adjustments to the weights of the compressed model. Once the model converges and reaches a stable condition, the cloud distributes the optimized model to each edge node. During distribution, the cloud also classifies scenarios based on signal characteristics and device type, assigning different parameters to different application scenarios to ensure that each node maintains a high level of model matching during actual operation.

[0088] The beneficial effects are as follows: This self-learning mechanism enables continuous model evolution, allowing the system to automatically adapt to changes in signal characteristics during long-term operation, reducing the need for manual maintenance. Through a collaborative architecture of edge feedback and centralized cloud optimization, both compression accuracy and operational stability are significantly improved.

[0089] Finally, the system executes the trusted data transmission phase. After generating compressed data blocks at the edge nodes, the system calculates a hash fingerprint for each data block and generates a digital signature, then writes this information into the node's local ledger recording module. When data is uploaded to the cloud, the cloud recalculates the hash fingerprint of the received data and compares it with the ledger record; if they match, it confirms that the data has not been tampered with during transmission, and the record is synchronously written to the cloud ledger. This ledger is a lightweight distributed structure that supports bidirectional verification and updates between the edge and the cloud.

[0090] The beneficial effects are as follows: Through the dual mechanism of hashing and signing, the entire process from data generation to upload is verifiable, ensuring the security and authenticity of compressed data. The lightweight ledger design reduces the computational burden on edge devices and maintains recording continuity even when there is a disconnection or delay at the cloud edge, effectively improving the reliability and traceability of the system.

[0091] In summary, the overall working principle of this invention is as follows: signal complexity drives sampling optimization, coherence analysis enables cross-node collaborative compression, spectral features assist in repair to improve restoration quality, error learning drives model evolution, and a trusted ledger mechanism ensures data security. The various modules form a closed-loop architecture encompassing intelligent acquisition, collaborative compression, self-learning optimization, and secure transmission, ensuring both the system's real-time performance and accuracy while significantly reducing network and storage resource consumption.

[0092] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for high-frequency signal acquisition and compression processing based on edge computing, characterized in that, Includes the following steps: Signal acquisition and complexity analysis: High-frequency signals from the target device are acquired at edge nodes, and the sampled signals are subjected to spectral analysis to calculate the instantaneous spectral entropy to obtain a complexity index. The sampling rate and sampling window length are adaptively adjusted according to the changes in the complexity index so that high-complexity signals are acquired at a higher sampling rate and low-complexity signals are acquired at a lower sampling rate. Feature extraction and collaborative compression: The sampled signal is input into the feature extraction module to extract the time-frequency feature vector; multiple edge nodes calculate the signal correlation coefficient matrix between nodes based on the time synchronization signal, and identify similar signal intervals based on the coherence threshold; Perform cross-node differential compression on the similar signal intervals, transmitting only the changed signal segments; Compressed bitstream generation and reversible repair: spectral feature information for reconstruction error detection is embedded in the compressed bitstream. When the energy retention rate of the reconstructed signal is lower than the threshold, the edge nodes perform local repair based on the feature information to compensate for signal distortion. The model self-learning and evolution process uploads the error distribution generated during compression and reconstruction to the cloud model repository and updates the parameters of the edge compression model based on the error gradient. The cloud updates the model according to different signal feature scenarios, realizing the self-learning and dynamic evolution of the edge compression model. Trusted data upload involves generating hash fingerprints and verifying the integrity of compressed data blocks to ensure the security and trustworthiness of compressed data during transmission from the edge to the cloud.

2. The high-frequency signal acquisition and compression processing method based on edge computing according to claim 1, characterized in that, Perform spectral analysis on the sampled signal at the edge node to obtain energy distribution information for each time slice; The complexity index of the signal is calculated based on the degree of change in energy distribution, and the complexity index is compared with a preset complexity threshold. When the complexity index is higher than the threshold, the control sampling module increases the sampling rate and shortens the sampling window length. When the complexity index is lower than the threshold, the control sampling module reduces the sampling rate and extends the sampling window length; By adaptively adjusting the sampling rate and sampling window, the sampling signal output by the edge node is kept in balance between time resolution and frequency resolution.

3. The high-frequency signal acquisition and compression processing method based on edge computing according to claim 2, characterized in that, The feature extraction and collaborative compression steps include: The adaptively sampled signal is input into the feature extraction module, which performs time-domain and frequency-domain analysis on the signal to obtain a time-frequency feature vector that reflects the signal amplitude changes and spectral distribution. Multiple edge nodes receive a unified time reference through a time synchronization module to align the timestamps of signals collected by each node.

4. The high-frequency signal acquisition and compression processing method based on edge computing according to claim 6, characterized in that, Based on the synchronized signal data, the correlation coefficient matrix between the signals of the nodes is calculated, and the similarity interval of the signals of each node is determined according to the preset coherence threshold. When a similar interval is determined, the collaborative compression module selects a node as a reference node, compresses and encodes the difference data between the signals of other nodes and the signal of the reference node, and transmits only the difference data fragment to the aggregation node. The output of the differential compression is used as input data for subsequent compressed bitstream generation and error correction steps.

5. The high-frequency signal acquisition and compression processing method based on edge computing according to claim 4, characterized in that, The compressed bitstream generation and feature restoration steps include: The differentially compressed signal is encoded at the edge nodes to generate a compressed bitstream; Spectral feature information for reconstruction detection is appended to the compressed bitstream. The spectral feature information includes the dominant frequency position of the signal before compression and the corresponding energy percentage. At the aggregation node, the received compressed bitstream is decoded and reconstructed, and the energy difference between the reconstructed signal and the reference signal is calculated.

6. In the high-frequency signal acquisition and compression processing method based on edge computing according to claim 5, when the energy difference exceeds a preset error threshold, the edge node is triggered to call the attached spectral feature information to perform local signal repair and compensate for the main frequency component of the reconstructed signal; The repaired signal data serves as input for the model's self-learning and dynamic evolution steps.

7. The high-frequency signal acquisition and compression processing method based on edge computing according to claim 6, characterized in that, The model's self-learning and dynamic evolution steps include: Error information obtained from edge nodes during compression and reconstruction is uploaded to the cloud model repository; The cloud-based model repository calculates the parameter update amount based on the error information uploaded by each edge node, and adjusts the weight parameters of the compressed model according to the update amount.

8. The high-frequency signal acquisition and compression processing method based on edge computing according to claim 7, characterized in that, When the updated compressed model meets the preset stability conditions, the model is distributed to each edge node for replacement. When the model is distributed in the cloud, edge nodes are classified into scenarios based on signal complexity distribution, spectral characteristics differences, or device operation type. Corresponding compression model parameters are assigned to nodes in different scenarios to maintain the convergence and adaptability of the compression model under different operating environments.

9. The high-frequency signal acquisition and compression processing method based on edge computing according to claim 8, characterized in that, The trusted data transmission steps include: After generating compressed data blocks at the edge nodes, a hash fingerprint value is calculated for each data block, and a digital signature is generated based on the node identity information. The hash fingerprint value and digital signature are recorded together in the ledger module of the edge node.

10. The high-frequency signal acquisition and compression processing method based on edge computing according to claim 9, characterized in that, When compressed data blocks are uploaded to the cloud, the cloud verifies whether the hash fingerprint of the received data matches the fingerprint value in the ledger record module. If the verification is successful, the integrity of the transmitted data is confirmed, and the record is synchronously written to the cloud ledger. The ledger recording module is a lightweight distributed structure that supports bidirectional verification and record updates between edge nodes and the cloud.

Images