A method for edge-side industrial data lightweight preprocessing caching

Abstract

The application provides a kind of edge side industrial data lightweight pretreatment cache method, belong to industrial data technical field, the application is through Gray code anti-interference coding to improve data acquisition quality, adaptive sampling rate adjustment and weak abnormal feature weight vector realizes dynamic sampling, wavelet decomposition and high frequency fault feature sparse matrix accurately locates fault position, semantic perception compression encoder cooperates with dispersion evaluation matrix and carries out differential coding, and network transmission graph model optimizes transmission path, this overall systematic technical scheme from signal acquisition, adaptive processing, intelligent compression, dynamic cache management to network optimization each link, solves the technical problems that the cache occupancy rate is too high and the data transmission delay is large caused by the difficulty of efficient pretreatment of multi-source signals of industrial equipment at edge side.

Description

Technical Field

[0001] This invention belongs to the field of industrial data technology, and more specifically, relates to a method for lightweight preprocessing and caching of industrial data at the edge. Background Technology

[0002] In industrial IoT and edge computing applications, industrial equipment needs to collect multi-source signals such as vibration, temperature, pressure, and current through multiple sensors. These signals generate massive amounts of data with high update frequencies. Traditional signal processing methods typically employ a fixed sampling rate to uniformly sample all signals, followed by frequency domain analysis using Fourier transform or wavelet transform, and finally data compression using standard lossy or lossless coding algorithms. However, these traditional methods have serious limitations in the actual deployment of edge computing nodes. Fixed sampling strategies prevent the amount of data collected from dynamically adjusting to signal characteristics. For low-frequency signals in stable operating states, a fixed sampling rate generates a large amount of redundant data, unnecessarily consuming the limited cache resources of edge nodes. For fault signals containing abrupt changes, a fixed sampling rate may fail to capture high-frequency details, affecting the accuracy of fault diagnosis. Furthermore, traditional general compression coding methods use the same coding strategy for different types of signal features, failing to differentiate processing based on fault characteristics. This results in either excessive loss of fault feature information or an unsatisfactory overall data compression ratio, ultimately leading to high cache occupancy and increased network transmission latency. In other words, existing technologies have technical problems such as difficulty in efficiently implementing adaptive sampling and intelligent compression coding of industrial signals at the edge, resulting in high buffer pressure and high transmission latency. Summary of the Invention

[0003] In view of this, the present invention provides a method for lightweight preprocessing caching of industrial data at the edge, which can solve the technical problems of excessively high cache occupancy and large data transmission delay in the real-time preprocessing of multi-source heterogeneous signals of industrial equipment at the edge in the prior art.

[0004] This invention is implemented as follows: A method for lightweight preprocessing and caching of industrial data at the edge includes: connecting industrial sensors to an edge computing node to collect signals and inputting Gray code analog signals into a digital anti-interference encoder to output the original digital signal sequence; inputting an adaptive sampling rate adjustment module to adjust the sampling rate according to a weak anomaly feature weight vector to output an adaptive sampling signal sequence; inputting a wavelet decomposition parallel computing unit to perform a four-layer wavelet transform and marking fault feature coefficients and normal signal coefficients according to a high-frequency fault feature sparse matrix; inputting a semantically aware compression encoder to query a ladder-style fault mode library vector to select the fault level and using lossless Huffman coding, arithmetic coding, and quantization compression coding; monitoring the cache occupancy rate parameter to dynamically adjust the number of parallel processing threads and cache space; performing a fast Fourier transform on the encoded data to extract the frequency band energy ratio value and performing filtering to obtain filtered data; writing the data into an edge-side circular buffer and constructing a network transmission graph model; and using Dijkstra's shortest path algorithm to transmit the data to complete the caching.

[0005] Among them, the Gray code analog-to-digital anti-interference encoder refers to an encoder that uses Gray code encoding to convert analog sensor signals into digital signals. The analog signal is converted into natural binary code by an analog-to-digital converter, and then converted into Gray code by a Gray code conversion circuit.

[0006] Among them, the weak anomaly feature weight vector refers to the numerical sequence used to quantify the importance of small anomaly changes in the signal. The construction steps are to collect historical signal data under normal operating conditions to obtain the mean sequence and standard deviation sequence, calculate the deviation value of the current sampled signal and divide it by the standard deviation value to obtain the standardized deviation sequence, and perform sliding window processing to calculate the sum of the absolute values ​​of the standardized deviations within the window.

[0007] The adaptive sampling rate adjustment module calculates the local variance and gradient rate of change, and judges the signal state based on the weight vector of weak anomaly features. When both the local variance and gradient rate of change are in the low value range (local variance is [0, 0.03] and gradient rate of change is [0, 0.05]), the sampling rate is reduced; when they are in the medium value range (local variance is (0.03, 0.12] and gradient rate of change is (0.05, 0.18]), the sampling rate is maintained; and when they are in the high value range (local variance is (0.12, 1] and gradient rate of change is (0.18, 1]), the sampling rate is increased.

[0008] Among them, 0.03 and 0.12 are derived from the statistical analysis of 100 sets of operating data under normal working conditions. 0.03 corresponds to the 25th percentile and 0.12 corresponds to the 75th percentile. 0.05 and 0.18 are derived from the gradient analysis of historical fault data. 0.05 corresponds to the upper limit of the 95% confidence interval of normal signal gradient changes and 0.18 corresponds to the typical threshold of slight abnormal signal gradient changes.

[0009] Among them, the high-frequency fault feature sparse matrix refers to a two-dimensional sparse matrix used to identify the distribution of fault features in the high-frequency coefficients after wavelet decomposition. The construction steps are to establish an industrial equipment fault mode database, perform wavelet decomposition on the signal samples to count the positions of high-frequency detail coefficients, establish a matrix framework with the time sampling point number as the row index, assign a value of 1 to the fault feature position and assign a value of 0 to the other positions.

[0010] Among them, the wavelet decomposition parallel computing unit performs four-level wavelet transform on the adaptive sampled signal sequence, outputs high-frequency detail coefficients of each level and low-frequency approximation coefficients of the fourth level, identifies the position of coefficients containing fault features based on the high-frequency fault feature sparse matrix, and marks the fault feature coefficients and normal signal coefficients.

[0011] Among them, the tiered fault mode library vector refers to the set of fault feature template vectors organized according to the severity of the fault. The construction steps are to divide the fault type into minor fault level, moderate fault level and severe fault level, and to perform feature extraction and cluster analysis on the fault samples to calculate the center vector as the typical fault mode vector.

[0012] Among them, the semantic-aware compression encoder calculates the similarity value between the fault feature coefficients and each mode vector in the ladder fault mode library vector and selects the fault level corresponding to the mode vector with the largest similarity value. Lossless Huffman coding is used for the fault feature coefficients, arithmetic coding is used for the coefficients in the fourth layer low-frequency approximation coefficients with gradient absolute values ​​greater than the threshold, and quantization compression coding is used for the normal signal coefficients.

[0013] The discreteness evaluation matrix refers to a two-dimensional matrix used to evaluate the discreteness of signal coefficients and guide the adjustment of quantization step size. The construction steps are as follows: divide the coefficient sequence of each layer into data blocks according to the time window, calculate the discreteness index for each data block, normalize it, and then sum it by weight to obtain the comprehensive discreteness value and map it to the quantization step size adjustment factor.

[0014] In this process, the quantization step size of the quantization compression encoding is dynamically adjusted based on the adjustment factor at the corresponding position in the dispersion evaluation matrix. The mapping rule for the adjustment factor is that the overall dispersion value is in the low numerical range. When the signal coefficients are concentrated and have low dispersion, the adjustment factor is set to its first value (an empirical value of 1.8, using a larger quantization step size to increase the compression ratio), which falls within the medium value range. (Indicating moderate signal coefficient dispersion) is the second value (empirical value is 1.0, maintaining standard quantization step size), which is in the high value range. (This indicates that the signal coefficients are distributed widely and have a high degree of dispersion) is the third value (the empirical value is 0.6, which uses a smaller quantization step size to retain more details).

[0015] Among them, the cache utilization rate parameter of the edge computing node is monitored. When the increase of the cache utilization rate parameter within a preset time window exceeds the first growth threshold, the number of parallel processing threads of the semantic-aware compression encoder is increased and additional cache space is pre-allocated. When the decrease exceeds the first decrease threshold, the number of parallel processing threads is reduced and excess cache space is reclaimed.

[0016] Specifically, the power spectral density distribution is obtained by performing a fast Fourier transform on the compressed encoded data, and the energy proportion values ​​of the first, second, and third frequency bands are extracted. When the energy proportion value of the first frequency band is the largest, the low-pass filter gain is increased and low-pass filtering is performed to obtain low-frequency filtered data.

[0017] Specifically, when the energy proportion of the second frequency band is the largest, the center frequency parameter of the bandpass filter is adjusted and bandpass filtering is performed to obtain intermediate frequency filtering data. When the energy proportion of the third frequency band is the largest, the gain of the high-pass filter is increased and high-pass filtering is performed to obtain high-frequency filtering data.

[0018] Among them, the network transmission graph model refers to a mathematical model that abstracts the network topology from edge nodes to cloud storage systems into a weighted directed graph. The construction steps are to take edge computing nodes, intermediate routing nodes and cloud storage nodes as graph vertices, network links as directed edges of the graph, and assign weight values ​​to each directed edge.

[0019] Specifically, the filtered data is written to the edge-side circular buffer, which contains a first buffer and a second buffer. When the data fill rate of the first buffer reaches full, the process switches to the second buffer to continue writing, while the data in the first buffer is marked as data to be transmitted.

[0020] In this process, a network transmission graph model is constructed, and the Dijkstra shortest path algorithm is used to calculate the shortest path sequence from the edge computing node to the cloud storage node. The shortest path sequence is defined as the path with the smallest sum of the weights of all directed edges on the path. The data to be transmitted is then transmitted to the cloud storage node to complete the data caching based on the shortest path sequence.

[0021] This invention addresses the technical challenges of high buffer occupancy and significant data transmission latency caused by the inefficient preprocessing of multi-source signals at the edge of industrial equipment. Specifically, it establishes an adaptive sampling rate adjustment module, dynamically optimizes the sampling strategy by introducing a weight vector of weak anomaly features, identifies fault feature locations using wavelet decomposition and a high-frequency fault feature sparse matrix, employs a semantically aware compression encoder for differentiated encoding based on fault level, and selects the optimal transmission path using a network transmission graph model. The adaptive sampling rate adjustment dynamically adjusts the sampling frequency based on the signal's local variance and gradient rate of change, reducing sampling to 40% in stable regions and increasing it to 180% in abnormal regions. This significantly reduces data redundancy while preserving key feature information, thereby reducing buffer pressure at the edge. The semantically aware compression encoder accurately locates fault features using a high-frequency fault feature sparse matrix, employs lossless Huffman coding to preserve complete diagnostic information for fault features, and uses high-compression-ratio quantization encoding for normal signals, achieving a balance between compression efficiency and fault feature fidelity. Furthermore, the adaptive quantization strategy based on the discreteness evaluation matrix and network transmission optimized by Dijkstra's shortest path algorithm further reduce buffer occupancy and transmission latency. In summary, the present invention solves the technical problems mentioned in the background art, such as the difficulty in efficiently preprocessing industrial signals at the edge, which leads to excessive buffer occupancy and large transmission delay. Attached Figure Description

[0022] Figure 1 This is a flowchart of the method of the present invention.

[0023] Figure 2 This is a comparison chart of the original signal and the adaptive sampling processing in the example.

[0024] Figure 3 This is a diagram illustrating the dynamic management process of cache utilization in this embodiment.

[0025] Figure 4 This is a graph showing the changes in data volume at each stage of compression in the example. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0027] like Figure 1 The diagram shows a flowchart of a method for lightweight preprocessing and caching of industrial data at the edge, provided by the present invention. This method includes the following steps:

[0028] S01. Connect multiple industrial sensors to the edge computing node to collect vibration signals, temperature signals, pressure signals and current signals of industrial equipment. Input the vibration signals, temperature signals, pressure signals and current signals into the Gray code analog-to-digital anti-interference encoder. The Gray code analog-to-digital anti-interference encoder converts the analog signals into Gray code digital signals and outputs the original digital signal sequence.

[0029] S02. The original digital signal sequence is input into the adaptive sampling rate adjustment module. The adaptive sampling rate adjustment module calculates the local variance and gradient rate of change of the original digital signal sequence. The signal state is determined according to the weight components at the corresponding time points in the weak anomaly feature weight vector. When the local variance ∈ [0, 0.03] and the gradient rate of change ∈ [0, 0.05], the sampling rate is reduced to 40% of the original sampling rate and a low-frequency sampling signal sequence is output. When the local variance ∈ (0.03, 0.12] or the gradient rate of change ∈ (0.05, 0.18], the current sampling rate is maintained and a standard sampling signal sequence is output. When the local variance ∈ (0.12, 1] or the gradient rate of change ∈ (0.18, 1], the sampling rate is increased to 180% of the original sampling rate and a high-frequency sampling signal sequence is output.

[0030] S03. The low-frequency sampling signal sequence, the standard sampling signal sequence, and the high-frequency sampling signal sequence are merged into an adaptive sampling signal sequence. The adaptive sampling signal sequence is input into a wavelet decomposition parallel computing unit. The wavelet decomposition parallel computing unit performs a four-level wavelet transform on the adaptive sampling signal sequence and outputs the first-level high-frequency detail coefficients, the second-level high-frequency detail coefficients, the third-level high-frequency detail coefficients, the fourth-level high-frequency detail coefficients, and the fourth-level low-frequency approximation coefficients. Based on the high-frequency fault feature sparse matrix, the coefficients containing fault features are identified. The coefficients located at the fault feature positions among the first-level high-frequency detail coefficients, the second-level high-frequency detail coefficients, the third-level high-frequency detail coefficients, and the fourth-level high-frequency detail coefficients are marked as fault feature coefficients, and the remaining coefficients are marked as normal signal coefficients.

[0031] S04. Input the fault feature coefficients and the normal signal coefficients into the semantic-aware compression encoder. The semantic-aware compression encoder queries the ladder-style fault mode library vector, calculates the similarity value between the fault feature coefficients and each mode vector in the ladder-style fault mode library vector, selects the fault level corresponding to the mode vector with the largest similarity value as the current fault level, uses lossless Huffman coding on the fault feature coefficients to obtain fault feature coding data, uses arithmetic coding on the coefficients in the fourth layer low-frequency approximation coefficients with gradient absolute values ​​greater than the threshold 0.15 to obtain mutation point coding data, and uses quantization compression coding on the normal signal coefficients to obtain stationary region coding data. The quantization step size of the quantization compression coding is dynamically adjusted according to the adjustment factor at the corresponding position in the discreteness evaluation matrix.

[0032] S05. Monitor the cache utilization rate and processor utilization rate of the edge computing nodes. When the cache utilization rate increases by more than 35% within 5 minutes, increase the number of parallel processing threads of the semantic-aware compression encoder to 150% of the original number of threads and pre-allocate additional cache space to 130% of the original cache space. When the cache utilization rate decreases by more than 30% within 5 minutes, reduce the number of parallel processing threads of the semantic-aware compression encoder to 70% of the original number of threads and reclaim excess cache space to 75% of the original cache space.

[0033] S06. Merge the fault feature coding data, the abrupt change point coding data, and the stable region coding data into compressed coding data. Perform a fast Fourier transform on the compressed coding data to obtain the power spectral density distribution. Extract the energy proportion values ​​of the 20Hz to 80Hz frequency band, the 100Hz to 300Hz frequency band, and the 350Hz to 600Hz frequency band from the power spectral density distribution. When the energy proportion value of the 20Hz to 80Hz frequency band is greater than the energy proportion value of the 100Hz to 300Hz frequency band and also greater than the energy proportion value of the 350Hz to 600Hz frequency band, increase the low-pass filter gain to 1.4 times the original gain and apply the compressed coding data... Low-pass filtering is performed to obtain low-frequency filtered data. When the energy proportion of the 100Hz to 300Hz frequency band is greater than the energy proportion of the 20Hz to 80Hz frequency band and greater than the energy proportion of the 350Hz to 600Hz frequency band, the center frequency parameter of the bandpass filter is adjusted to 200Hz and bandpass filtering is performed on the compressed encoded data to obtain intermediate-frequency filtered data. When the energy proportion of the 350Hz to 600Hz frequency band is greater than the energy proportion of the 20Hz to 80Hz frequency band and greater than the energy proportion of the 100Hz to 300Hz frequency band, the gain of the high-pass filter is increased to 1.6 times the original gain and high-pass filtering is performed on the compressed encoded data to obtain high-frequency filtered data.

[0034] S07. The low-frequency filtered data, the mid-frequency filtered data, and the high-frequency filtered data are merged into filtered processing data. The filtered processing data is written into an edge-side circular buffer, which includes a first buffer and a second buffer. When the data fill rate of the first buffer reaches 100%, the process switches to the second buffer to continue writing. At the same time, the data in the first buffer is marked as data to be transmitted. A network transmission graph model is constructed, in which edge computing nodes, intermediate routing nodes, and cloud storage nodes are used as graph vertices, and network links between nodes are used as graph edges. The graph edge weights are calculated based on the transmission delay measurement, bandwidth utilization measurement, and packet loss rate measurement of each link. The shortest path sequence from the edge computing node to the cloud storage node is calculated using Dijkstra's shortest path algorithm. The data to be transmitted is then transmitted to the cloud storage node via the network according to the shortest path sequence to complete data caching.

[0035] It should be explained that the three interval boundary values ​​of 0.03 and 0.12 for the local variance value in step S02 are derived from the statistical analysis of vibration, temperature, pressure, and current signals under normal operating conditions. The quantiles of the local variance value are calculated by collecting 100 sets of normal operating data. 0.03 corresponds to the 25th quantile, representing the upper limit of the signal being in a highly stable state, and 0.12 corresponds to the 75th quantile, representing the lower limit of the signal starting to show significant fluctuations. The three interval boundary values ​​of 0.05 and 0.18 for the gradient rate of change value in step S02 are derived from the gradient analysis of historical fault data. 0.05 corresponds to the upper limit of the 95% confidence interval for normal signal gradient changes, and 0.18 corresponds to the typical threshold for slight abnormal signal gradient changes. The ratio of reducing the sampling rate to 40% and increasing the sampling rate to 180% in step S02 comes from the sampling theorem and the balance of computational resources. A sampling rate of 40% can meet the safety margin of twice the sampling requirement of the highest frequency component of a stable signal, while a sampling rate of 180% can capture the high-frequency transient characteristics of abrupt signals and maintain a 30% anti-aliasing margin.

[0036] It should be explained that the gradient absolute value threshold of 0.15 in step S04 is derived from the statistical analysis of the fourth-layer low-frequency approximation coefficients. It is calculated by adding three times the standard deviation to the mean of the gradient absolute values ​​of the fourth-layer low-frequency approximation coefficients of a normal signal. This threshold of 0.15 effectively distinguishes between stable changes and abrupt changes. The cache utilization growth threshold of 35% and the decrease threshold of 30% in step S05 are derived from empirical values ​​for edge computing node cache management. A 35% increase indicates a significant increase in cache pressure requiring resource expansion, while a 30% decrease indicates significant redundancy in cache resources requiring reclamation. The 5-minute time window in step S05 is derived from the periodicity analysis of industrial data streams; 5 minutes can cover a complete working cycle for most devices. The percentages for adjusting the number of parallel processing threads to 150% and 70% and the cache space to 130% and 75% in step S05 are derived from load balancing tests. The increased percentages of 150% and 130% can cope with sudden data traffic while avoiding excessive resource consumption, while the decreased percentages of 70% and 75% can release redundant resources while maintaining basic processing capacity.

[0037] It should be explained that the division of the three frequency bands in step S06—20Hz to 80Hz, 100Hz to 300Hz, and 350Hz to 600Hz—is based on the frequency distribution of typical fault characteristics in industrial equipment. 20Hz to 80Hz corresponds to imbalance and loosening faults in low-speed rotating parts; 100Hz to 300Hz corresponds to gear meshing and bearing faults in medium-speed transmission systems; and 350Hz to 600Hz corresponds to electromagnetic faults and high-frequency vibrations in high-speed motors. The 1.4x gain of the low-pass filter, the 200Hz center frequency of the band-pass filter, and the 1.6x gain of the high-pass filter in step S06 are derived from frequency domain compensation requirements. The 1.4x gain compensates for the attenuation of low-frequency signals during compression encoding, 200Hz is the geometric center frequency of the 100Hz to 300Hz band, and the 1.6x gain compensates for energy loss of high-frequency signals during quantization.

[0038] A Gray code analog-to-digital anti-interference encoder refers to an encoder that uses Gray code encoding to convert analog sensor signals into digital signals. The working principle of this Gray code analog-to-digital anti-interference encoder is as follows: the analog signal is first converted into natural binary code by an analog-to-digital converter, and then the natural binary code is converted back into Gray code by a Gray code conversion circuit. The Gray code conversion rule is that the highest bit of the Gray code is the same as the highest bit of the natural binary code, and the remaining bits of the Gray code are equal to the result of the XOR operation between the corresponding bit in the natural binary code and the bit higher. The core characteristic of Gray code is that adjacent values ​​differ by only one binary bit. When the sensor output analog signal fluctuates near the quantization boundary, the transition amplitude of the Gray code is much smaller than that of the natural binary code. This Gray code analog-to-digital anti-interference encoder can reduce the encoding fluctuation caused by quantization boundary jitter to a single-bit change, significantly reducing the number of transitions in the digital signal and improving the compression efficiency of subsequent compression algorithms. Gray code has a natural single-bit error detection capability. When a single-bit error occurs during transmission or storage, it only causes a deviation of one quantization level rather than multiple transitions in the natural binary code, enhancing the anti-interference capability of edge-side data acquisition.

[0039] A weak anomaly feature weight vector is a numerical sequence used to quantify the importance of minute anomaly changes in a signal. The construction steps of the weak anomaly feature weight vector are as follows: First, historical signal data under normal operating conditions is collected as a benchmark dataset. Statistical analysis is performed on the benchmark dataset to obtain the mean and standard deviation sequences for each time point. Then, the deviation value between the current sampled signal and the mean sequence is calculated. The deviation value is divided by the corresponding standard deviation value to obtain the standardized deviation sequence. Next, a sliding window processing is applied to the standardized deviation sequence, with a window length of 50 sampling points. The sum of the absolute values ​​of the standardized deviations within each window is calculated, and this sum of absolute values ​​is used as the weak anomaly feature weight at the center point of the window. Finally, the weak anomaly feature weights of all the window center points are combined to form the weak anomaly feature weight vector. Each element in the weak anomaly feature weight vector represents the degree to which the signal deviates from the normal state at the corresponding time point. A larger element value indicates a higher probability of anomaly features appearing at the corresponding time point, requiring a higher sampling rate to capture anomaly details. A smaller element value indicates a stable signal, allowing for a lower sampling rate to reduce the amount of data. The weak anomaly feature weight vector provides a quantitative basis for adaptive sampling rate adjustment, and the optimal allocation of sampling resources is achieved by dynamically monitoring the numerical changes of the weak anomaly feature weight vector.

[0040] A high-frequency fault feature sparse matrix is ​​a two-dimensional sparse matrix used to identify the distribution of fault features in high-frequency coefficients after wavelet decomposition. The construction steps of the high-frequency fault feature sparse matrix are as follows: First, an industrial equipment fault mode database is established, containing the characteristic frequency ranges and amplitude thresholds corresponding to various typical faults. Then, wavelet decomposition is performed on normal signal samples and fault signal samples respectively, and the positions of coefficients exceeding the amplitude threshold in each layer of high-frequency detail coefficients are counted. These coefficient positions are plotted on the time and frequency axes to form a two-dimensional coordinate system. Next, a matrix framework is established using the time sampling point index as the row index and the combination of the wavelet decomposition layer number and the frequency sub-band number as the column index. For the positions where characteristic coefficients appear in the fault signal samples, the corresponding element in the matrix is ​​assigned a value of 1, and the remaining positions are assigned a value of 0. Finally, a compressed sparse row storage format is used to record the non-zero elements in the matrix and their position indices to form the high-frequency fault feature sparse matrix. In the high-frequency fault feature sparse matrix, the wavelet coefficients corresponding to the positions with a value of 1 contain fault feature information and need to be preserved using lossless or low-distortion encoding. The positions with a value of 0 correspond to normal signal coefficients and are allowed to use lossy encoding with a high compression ratio. The sparsity of the high-frequency fault feature sparse matrix is ​​reflected in the local concentration of fault feature signals in the time-frequency domain. Most elements in the matrix are 0, and non-zero elements only appear at the time of fault occurrence and in the frequency band of fault occurrence. The sparsity characteristic enables the compression encoder to adopt differentiated compression strategies for different regions, which can significantly reduce the amount of data while ensuring that the key features required for fault diagnosis are not destroyed.

[0041] A tiered fault mode library vector refers to a set of fault feature template vectors organized according to the severity of the fault. The construction steps of the tiered fault mode library vector are as follows: First, based on historical fault records of industrial equipment and expert knowledge, fault types are divided into three levels: minor fault, moderate fault, and severe fault. Wavelet decomposition and feature extraction are performed on fault samples within each level. The extracted features include the energy proportion of each frequency band, peak frequency position, time-domain envelope shape parameters, and statistical moment parameters. Then, cluster analysis is performed on the feature vectors of all fault samples within the same level, and the center vector of each cluster is calculated as the typical fault mode vector for the corresponding level. Next, the typical fault mode vectors of each level are arranged sequentially according to the severity of the fault from low to high, forming a tiered vector sequence. Finally, a fault type label, characteristic frequency band range, and recommended compression quality factor parameter are added to each vector to form the complete tiered fault mode library vector. Each vector in the stepped fault mode library vector set represents a standard feature of a fault mode with a certain severity. The feature component values ​​in the vector reflect the performance of the corresponding fault in different frequency bands and time domain statistics. The stepped arrangement enables the semantic-aware compression encoder to quickly determine the severity of the fault based on the similarity between the current signal features and the vectors at each level, and then match the corresponding compression coding strategy. For the signal segment corresponding to the severe fault level, a higher coding quality is used to retain complete diagnostic information, while for the signal segment corresponding to the minor fault level, a higher compression ratio is used to reduce the amount of data, thus achieving a hierarchical balance between compression efficiency and diagnostic information fidelity.

[0042] The discreteness evaluation matrix is ​​a two-dimensional matrix used to evaluate the discreteness of signal coefficients and guide the adjustment of quantization step size. The steps for constructing the dispersion evaluation matrix are as follows: First, the coefficient sequences of each layer after wavelet decomposition are divided into multiple data blocks according to the time window. Each data block contains 256 continuous coefficients. Then, a dispersion index is calculated for each data block. The dispersion index includes the range, interquartile range, and coefficient of variation. The range, interquartile range, and coefficient of variation are normalized and then weighted and summed to obtain the comprehensive dispersion value. The normalization method is to divide each index value by the maximum value of the corresponding index in all data blocks, with weights set to 0.4, 0.3, and 0.3, respectively. Next, a matrix is ​​established with the data block number as the row index and the wavelet decomposition layer number as the column index. The comprehensive dispersion value of each data block in the corresponding layer is filled into the corresponding position in the matrix. Finally, the matrix elements are mapped to quantization step size adjustment factors according to the distribution of the comprehensive dispersion values. The mapping rule is that the adjustment factor is 1.8 when the comprehensive dispersion value is less than 0.2, 1.0 when the comprehensive dispersion value is ∈ [0.2, 0.5], and 0.6 when the comprehensive dispersion value is greater than 0.5. Each element in the discreteness evaluation matrix corresponds to the signal coefficient discreteness of a time window and frequency layer. Larger element values ​​indicate a more dispersed coefficient distribution in the corresponding region, requiring a smaller quantization step size to retain more detail. Smaller element values ​​indicate a more concentrated coefficient distribution, allowing for a larger quantization step size to improve the compression ratio. This discreteness evaluation matrix provides an adaptive parameter adjustment basis for quantization compression coding. By dynamically optimizing the quantization strategy through real-time updates of matrix element values, the compression algorithm can flexibly adjust the compression intensity based on the local statistical characteristics of the signal. This achieves a higher compression ratio in stable, low-dispersion regions and retains sufficient accuracy in abruptly high-dispersion regions, thus achieving a dynamic balance between data reduction and information fidelity.

[0043] The network transmission graph model is a mathematical model that abstracts the network topology from edge nodes to cloud storage systems into a weighted directed graph. The construction steps of the network transmission graph model are as follows: First, edge computing nodes, intermediate routing nodes, and cloud storage nodes are respectively used as vertices of the graph, and the network links between nodes are used as directed edges of the graph. Then, a weight value is assigned to each directed edge. The weight value comprehensively considers three factors: the transmission delay measurement value, the bandwidth utilization measurement value, and the packet loss rate measurement value. The weight value is calculated by multiplying the normalized values ​​of the transmission delay measurement value, the bandwidth utilization measurement value, and the packet loss rate measurement value by weight coefficients of 0.5, 0.3, and 0.2, respectively, and then summing them. The normalization method is to divide the measured value of each parameter by the historical maximum value of the corresponding parameter. Next, Dijkstra's shortest path algorithm is used to calculate the shortest path sequence from the edge computing node to the cloud storage node. The shortest path sequence is defined as the path with the smallest sum of the weight values ​​of all directed edges on the path. During the algorithm iteration process, the shortest distance estimate and predecessor node record of each vertex are maintained. Finally, the forwarding route sequence of data packets is determined based on the calculated shortest path sequence. The network transmission graph model is applied to optimize data transmission paths. By periodically updating the weight values ​​of directed edges, it adapts to the dynamic changes in network status, avoids sending data to congested or high-latency links, and ensures that compressed data can be transmitted to cloud storage nodes with the lowest latency and highest reliability, thereby improving the data interaction efficiency between the edge and the cloud.

[0044] The technical solution of this invention has the following advantages: it solves the problem of insufficient real-time processing capability of high-frequency sampled data by using an adaptive sampling rate adjustment module and a wavelet decomposition parallel computing unit; it achieves a balance between compression efficiency and fault feature preservation by using a semantically aware compression encoder and a high-frequency fault feature sparse matrix; it improves the anti-interference capability of data acquisition by using a Gray code analog-to-digital anti-interference encoder; it optimizes and reduces data transmission latency by combining a network transmission graph model and Dijkstra's shortest path algorithm; and the overall solution achieves lightweight preprocessing and efficient caching of industrial data at the edge.

[0045] As an optional technical solution, the present invention also provides a method for implementing a lightweight preprocessing cache system for edge-side industrial data using a computer. The computer is equipped with a readable storage medium, which stores program instructions. When the program instructions are run on the computer, they execute the aforementioned method for lightweight preprocessing cache of edge-side industrial data.

[0046] The specific implementation methods of the above steps are described in detail below.

[0047] The specific implementation of step S01 is as follows: Multiple industrial sensors are deployed on the edge computing node to collect vibration, temperature, pressure, and current signals from industrial equipment in real time. The collected analog signals are converted into natural binary code by an analog-to-digital converter, and then converted into Gray code digital signals by a Gray code conversion circuit. Gray code encoding utilizes the characteristic that adjacent values ​​differ by only one binary bit, which can significantly reduce the encoding jump amplitude when the sensor output analog signal fluctuates near the quantization boundary. It limits the encoding fluctuation caused by quantization boundary jitter to a single-bit change range, improving the compression efficiency of subsequent compression algorithms. Gray code also has a natural single-bit error detection capability. When a single-bit error occurs during data transmission or storage, it only causes a deviation of one quantization level. Compared with the multi-level jumps of natural binary code, it can effectively enhance the anti-interference capability of data acquisition. The digital signal after Gray code conversion forms the original digital signal sequence.

[0048] The specific implementation of step S02 is as follows: The adaptive sampling rate adjustment module processes the original digital signal sequence and uses two quantitative indicators, local variance calculation and gradient rate of change analysis, to evaluate the signal state. First, the current signal's operating state is determined based on the weight components at corresponding time points in the weak anomaly feature weight vector. The weak anomaly feature weight vector is constructed by statistically analyzing historical signal data under normal operating conditions and using a sliding window method (window length of 50 sampling points), where each element represents the degree to which the signal deviates from the normal state at the corresponding time point. When the signal's local variance value... [0, 0.03] and gradient rate of change value When the value is [0, 0.05], the signal is determined to be in a highly stable state. The system reduces the sampling rate to 40% of the original sampling rate and outputs a low-frequency sampling signal sequence; when the local variance value is... (0.03, 0.12] or gradient rate of change value When the value is (0.05, 0.18), the signal is determined to be in a normal changing state, the current sampling rate is maintained, and a standard sampled signal sequence is output; when the local variance value is... (0.12, 1] or gradient rate of change value When the value is (0.18, 1), an abnormal fluctuation in the signal is determined, the sampling rate is increased to 180% of the original sampling rate, and a high-frequency sampled signal sequence is output. This method is designed based on the sampling theorem. A sampling rate of 40% can meet the safety margin requirement of twice the sampling of the highest frequency component of a stable signal, while a sampling rate of 180% can capture the high-frequency transient characteristics of abrupt signals and maintain a 30% anti-aliasing margin, thus achieving an optimized balance between sampling resources and acquisition quality.

[0049] The specific implementation of step S03 is as follows: The low-frequency sampled signal sequence, the standard sampled signal sequence, and the high-frequency sampled signal sequence are merged into an adaptive sampled signal sequence, which is then input into a wavelet decomposition parallel computing unit for multi-level frequency domain analysis. The wavelet decomposition parallel computing unit performs four-level wavelet transform on the adaptive sampled signal sequence, obtaining high-frequency detail coefficients and low-frequency approximation coefficients at each level. The four-level decomposition outputs high-frequency detail coefficients at the first, second, third, and fourth levels, as well as low-frequency approximation coefficients at the fourth level. The system then uses a high-frequency fault feature sparse matrix to identify the positions of coefficients containing fault features. This sparse matrix is ​​constructed based on an industrial equipment fault mode database. A two-dimensional coordinate system is established by combining the time sampling point index, the wavelet decomposition level, and the frequency sub-band. The positions where feature coefficients appear in the fault signal samples are marked as 1, and the remaining positions are marked as 0. Based on the identification results of the sparse matrix, coefficients located at fault feature positions are marked as fault feature coefficients, and the remaining coefficients are marked as normal signal coefficients, achieving hierarchical isolation between fault-related information and stable background information.

[0050] The specific implementation of step S04 is as follows: Fault feature coefficients and normal signal coefficients are differentially encoded using a semantically aware compression encoder. The encoder first queries a tiered fault pattern library vector, which is divided into three levels according to fault severity: minor fault, moderate fault, and severe fault. Each level obtains a corresponding typical fault pattern vector through cluster analysis. The encoder calculates the similarity value between the current fault feature coefficient and each pattern vector, selects the fault level corresponding to the pattern vector with the highest similarity value as the current fault level, and then matches the corresponding compression strategy. Lossless Huffman coding is used for fault characteristic coefficients to ensure the complete preservation of key diagnostic information. For coefficients in the fourth-level low-frequency approximation coefficients with gradient absolute values ​​greater than a threshold of 0.15, arithmetic coding is used to obtain abrupt change point encoding data. This threshold is obtained by calculating the mean of the gradient absolute values ​​of the fourth-level low-frequency approximation coefficients of the normal signal plus three times the standard deviation, effectively distinguishing between stationary changes and abrupt changes. For normal signal coefficients, quantization compression coding is used to obtain stationary region encoding data. The quantization step size is dynamically adjusted according to the adjustment factor at the corresponding position in the dispersion evaluation matrix. The dispersion evaluation matrix is ​​based on the dispersion of the signal across 256 sampling point data blocks. When the comprehensive dispersion value is less than 0.2, the adjustment factor is 1.8; when the value is less than 0.2, the adjustment factor is 1.8. The adjustment factor is 1.0 when the value is [0.2, 0.5], and 0.6 when the value is greater than 0.5.

[0051] The specific implementation of step S05 is as follows: The system monitors the cache utilization rate and processor utilization rate parameters of the edge computing nodes to dynamically adapt to changes in data traffic. A 5-minute time window is used for evaluation, which covers a complete working cycle of most industrial equipment. When the cache utilization rate parameter increases by more than 35% within 5 minutes, it indicates a significant increase in cache pressure. The system responsively increases the number of parallel processing threads of the semantic-aware compression encoder to 150% of the original number of threads, while pre-allocating additional cache space to 130% of the original cache space. The 150% and 130% increase ratios can cope with sudden data traffic while avoiding excessive resource consumption. When the cache utilization rate parameter decreases by more than 30% within 5 minutes, it indicates significant redundancy in cache resources. The system correspondingly reduces the number of parallel processing threads to 70% of the original number of threads and reclaims excess cache space to 75% of the original cache space. The 70% and 75% reduction ratios can effectively release redundant resources while maintaining basic processing capabilities.

[0052] The specific implementation of step S06 is as follows: The system merges fault feature coding data, abrupt change point coding data, and stable region coding data into compressed coding data, performs a fast Fourier transform on it to obtain frequency domain characteristics, and calculates the power spectral density distribution. Frequency band division is designed based on the frequency distribution of typical fault characteristics of industrial equipment, and the system extracts 20–80 MHz from the power spectral density distribution. Energy percentage of frequency band, 100-300 The energy proportion of the frequency band and 350-600 The energy proportions of each frequency band correspond to imbalance and loosening faults in low-speed rotating components, gear meshing and bearing faults in medium-speed transmission systems, and electromagnetic faults and high-frequency vibrations in high-speed motors. When 20–80... When the frequency band energy proportion is at its maximum, increase the low-pass filter gain to 1.4 times the original gain and perform low-pass filtering to obtain low-frequency filtered data; when 100~300 When the frequency band energy proportion is at its maximum, adjust the center frequency parameter of the bandpass filter to 200. (i.e., 100-300) The geometric center frequency of the frequency band is determined and bandpass filtering is performed to obtain intermediate frequency filtered data; when 350~600 When the frequency band energy ratio is at its maximum, increase the high-pass filter gain to 1.6 times the original gain and perform high-pass filtering to obtain high-frequency filtered data.

[0053] The specific implementation of step S07 is as follows: The system merges low-frequency filtered data, mid-frequency filtered data, and high-frequency filtered data into filtered processing data, and writes it into the edge-side circular buffer to achieve efficient data caching. The circular buffer adopts a dual-buffering mechanism design, including a first buffer and a second buffer. When the data fill rate of the first buffer reaches 100%, it automatically switches to the second buffer to continue writing, while marking the data in the first buffer as data to be transmitted. The system constructs a network transmission graph model based on the network topology, using edge computing nodes, intermediate routing nodes, and cloud storage nodes as graph vertices, and network links between nodes as graph edges. The graph edge weight values ​​are calculated based on the transmission delay measurement, bandwidth utilization measurement, and packet loss rate measurement of each link. The weight value calculation adopts a normalized weighted summation method, assigning weight coefficients of 0.5, 0.3, and 0.2 to the three parameters respectively. The normalization process divides the measured value of each parameter by the historical maximum value of the corresponding parameter. The system uses Dijkstra's shortest path algorithm to calculate the shortest path sequence from the edge computing node to the cloud storage node. The algorithm iteratively maintains the shortest distance estimate and predecessor node record for each vertex, and finally determines the transmission route with the minimum sum of the weights of all directed edges on the path. The data to be transmitted is then transmitted to the cloud storage node through the network according to the shortest path sequence to complete the data caching.

[0054] It should be noted that the key technical ideas of this invention include: First, the collaborative implementation of an adaptive sampling rate adjustment module and wavelet decomposition achieves hierarchical processing in the frequency domain. This module dynamically adjusts the sampling rate based on the signal's local variance and gradient rate of change, employing differentiated sampling strategies for different signal characteristics. This provides multi-level precision input signals for subsequent wavelet transforms, significantly reducing data redundancy in low-stationarity regions compared to fixed sampling rate schemes, while preserving transient features in high-anomaly regions. Second, the combination of semantically aware compression coding and a sparse matrix of high-frequency fault features achieves a balance between compression efficiency and fault fidelity. Based on the hierarchical structure of a fault mode library, this scheme can dynamically adjust the coding strategy according to the severity of the fault. Lossless coding is used for key fault information, while high compression ratio coding is used for stable backgrounds. Compared to a uniform compression scheme, this significantly reduces the amount of data while ensuring the integrity of key diagnostic information. Third, the combination of a network transmission graph model and Dijkstra's algorithm achieves dynamic routing optimization. This solution comprehensively considers multiple factors such as link latency, bandwidth, and packet loss rate. By periodically updating weight values, it adapts to dynamic changes in network status. Compared with static routing solutions, it can effectively prevent data from flowing to congested links and ensure that data is transmitted to the cloud with the lowest latency and the highest reliability.

[0055] It should be noted that this invention also solves the following technical problem: In traditional analog-to-digital conversion, excessive encoding jumps caused by quantization boundary jitter lead to insufficient anti-interference capability in data acquisition. This invention employs a Gray code analog-to-digital anti-interference encoder. Utilizing the characteristic that adjacent numerical codes in Gray code differ by only one binary bit, it limits encoding fluctuations caused by quantization boundary jitter to a single-bit variation range, significantly reducing the number of jumps in the digital signal and improving the effectiveness of subsequent compression algorithms. Simultaneously, Gray code has a natural single-bit error detection capability; when a single-bit error occurs during transmission or storage, it only causes a deviation of one quantization level, unlike natural binary codes which produce multi-level jumps, thereby enhancing the anti-interference capability of edge-side data acquisition.

[0056] Furthermore, this invention addresses the technical problem of low efficiency in dynamic cache management of edge computing nodes. Traditional solutions typically employ static cache configurations, which cannot cope with sudden changes in industrial data traffic. This invention establishes a dynamic cache expansion and reclamation mechanism based on a 5-minute time window by real-time monitoring of cache occupancy and processor occupancy parameters. When cache pressure increases, the number of parallel processing threads is increased to 150%, and the pre-allocated cache is increased to 130% to handle sudden traffic surges. When pressure decreases, the number of threads is reduced to 70%, and the cache is reclamated to 75% to release redundant resources. This dynamic resource management strategy avoids data loss due to cache overflow and also prevents long-term excessive resource consumption, thus improving the overall processing efficiency of edge nodes.

[0057] Specifically, the principle of this invention is as follows: The solution is based on the principle of local adaptability in signal processing and the objective optimization strategy of information theory. Multi-source signals from industrial equipment exhibit time-varying characteristics during actual operation: during normal and stable operation, signal energy is concentrated in the low-frequency band and changes slowly; during abnormal and fault phases, signals change rapidly and generate rich high-frequency components. Traditional fixed sampling cannot adapt to this dynamic characteristic. The adaptive sampling rate adjustment module, by calculating local variance and gradient change rate in real time, dynamically judges the signal state based on the weight vector of weak anomaly features, ensuring that the sampling rate always tracks changes in signal complexity, thereby minimizing redundant data while retaining necessary information. The multi-resolution characteristics of wavelet decomposition allow signal components of different frequencies to be separated into different decomposition layers. The sparse matrix of high-frequency fault features fully utilizes the time-frequency local concentration of fault information to accurately identify the coefficient positions requiring high-fidelity processing. The semantic-aware compression encoder judges the fault level based on a ladder-style fault mode library vector, achieving differentiated encoding based on the semantic features of the signal rather than simply the amplitude. This maintains high accuracy for key diagnostic features and uses a high compression ratio for stable signals, resulting in a significant reduction in buffer occupancy and a significant reduction in data transmission latency.

[0058] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0059] The specific implementation of step S01 is as follows: Multiple industrial sensors are deployed on the edge computing node to collect vibration, temperature, pressure, and current signals from industrial equipment in real time. The collected analog signals are converted into natural binary code via an analog-to-digital converter. A Gray code conversion circuit converts the natural binary code into Gray code digital signals. The mathematical expression of the conversion rule is as follows:

[0060] ;

[0061] ;

[0062] In the formula, For Gray code Bit, The first of the natural binary code Bit, The first of the natural binary code Bit, It is the XOR operator. , The encoding bit length is typically 12 or 16. The digital signal converted by Gray code forms the original digital signal sequence.

[0063] The specific implementation of step S02 is as follows: The adaptive sampling rate adjustment module processes the original digital signal sequence, first calculating the local variance value, the calculation formula of which is as follows:

[0064] ;

[0065] In the formula, For the first Local variance values ​​of each sampling point For the first The signal value at each sampling point For the first The mean of the signal within a local window centered on each sampling point is calculated using the following formula: , This is the local window length, with an empirical value of 100 sampling points.

[0066] Next, the gradient rate of change is calculated using the following formula:

[0067] ;

[0068] In the formula, For the first The gradient rate of change value at each sampling point For the first The signal value at each sampling point For the first The signal value at each sampling point The maximum value in the signal sequence. It is the minimum value in the signal sequence.

[0069] In the process of constructing the weight vector of weak anomaly features, the formula for calculating the standardized deviation sequence is as follows:

[0070] ;

[0071] In the formula, For the first Standardized deviation of each sampling point The current sampled signal is the first Each sample point value, For the benchmark dataset in the The mean value of each sampling point was calculated from historical signal data under normal operating conditions. For the benchmark dataset in the 1st The standard deviation of each sampling point.

[0072] The formula for calculating the weights of weak anomaly features is as follows:

[0073] ;

[0074] In the formula, For the first The weights of weak anomaly features at each sampling point are calculated, with a sliding window length of 50 sampling points and the window center point being the [number]th sampling point. One sampling point, For the first in the window The standardized deviation value of each sampling point.

[0075] according to and The numerical judgment sampling rate adjustment strategy, when and At that time, the adjusted sampling rate was:

[0076] ;

[0077] In the formula, To adjust the post-sampling rate, Original sampling rate, in units of .

[0078] when or At that time, the adjusted sampling rate was:

[0079] .

[0080] The specific implementation of step S03 is as follows: the low-frequency sampling signal sequence, the standard sampling signal sequence, and the high-frequency sampling signal sequence are merged into an adaptive sampling signal sequence according to time order, and then input into the wavelet decomposition parallel computing unit to perform a four-level wavelet transform. The high-frequency detail coefficients and low-frequency approximation coefficients of layer wavelet decomposition are expressed as follows: and ,in The elements of the high-frequency fault feature sparse matrix are defined as follows:

[0081] ;

[0082] In the formula, For the high-frequency fault feature sparse matrix in the first place Line number Column elements, The time sampling point number, This is a combined index of wavelet decomposition layer number and frequency subband number.

[0083] The specific implementation of step S04 is as follows: The semantic-aware compression encoder calculates the similarity value between the fault feature coefficients and each mode vector in the ladder fault mode library vector. The similarity calculation formula is as follows:

[0084] ;

[0085] In the formula, The fault characteristic coefficient vector and the first The similarity value of each pattern vector. The first fault characteristic coefficient vector One portion, For the first The first pattern vector One portion, This is the vector dimension, typically 64.

[0086] The formula for calculating the absolute value of the gradient of the fourth-layer low-frequency approximation coefficients is as follows:

[0087] ;

[0088] In the formula, For the first The absolute value of the gradient of the fourth-layer low-frequency approximation coefficients. The fourth layer low-frequency approximation coefficient sequence One coefficient, For the first Each coefficient. When At that time, the coefficient is arithmetically encoded to obtain the mutation point encoded data.

[0089] In constructing the dispersion evaluation matrix, the range is calculated for each data block. Its formula is:

[0090] ;

[0091] In the formula, For the first The range of each data block For the first data block One coefficient, For the first A set of coefficient indices contained in each data block.

[0092] Interquartile range The calculation formula is:

[0093] ;

[0094] In the formula, For the first Interquartile range of data blocks For the first The 75th percentile of the coefficient sequence of each data block It is the 25th percentile. For the first A sequence of coefficients for each data block.

[0095] coefficient of variation The calculation formula is:

[0096] ;

[0097] In the formula, For the first The coefficient of variation of each data block For the first The standard deviation of the coefficient sequence of each data block is calculated using the following formula: , For the first The mean of the coefficient sequence of each data block is calculated using the following formula: 256 is the number of consecutive coefficients contained in each data block.

[0098] The formula for calculating the overall dispersion value is as follows:

[0099] ;

[0100] In the formula, For the first The overall dispersion value of each data block. The maximum range among all data blocks. This represents the maximum interquartile range across all data blocks. This represents the maximum value of the coefficient of variation across all data blocks.

[0101] Quantization step size adjustment factor Based on the comprehensive dispersion value Sure:

[0102] .

[0103] The specific implementation of step S05 is as follows: The system monitors the cache utilization rate and processor utilization rate parameters of the edge computing nodes in real time. The calculation formula for the increase in cache utilization rate is as follows:

[0104] ;

[0105] In the formula, This represents the increase in cache usage over a 5-minute period. The current cache size, in units of , This represents the cache size 5 minutes ago, in units of... , Total cache capacity, in units of .

[0106] when At that time, the number of parallel processing threads was adjusted to:

[0107] ;

[0108] In the formula, This is the adjusted number of parallel processing threads. This represents the original number of threads.

[0109] The pre-allocated cache space has been adjusted to:

[0110] ;

[0111] In the formula, The adjusted cache space, in units of , Original cache space, unit: .

[0112] The formula for calculating the decrease in cache utilization is:

[0113] ;

[0114] In the formula, This represents the decrease in cache usage over a 5-minute period. At that time, the number of parallel processing threads was adjusted to Reclaim cache space and adjust to .

[0115] The specific implementation of step S06 is as follows: Perform a fast Fourier transform on the compressed encoded data to obtain the power spectral density distribution. ,in Frequency, unit: The formulas for calculating the energy proportion of each frequency band are as follows:

[0116] ;

[0117] ;

[0118] ;

[0119] In the formula, 20-80 The energy percentage of the frequency band, 100-300 The energy percentage of the frequency band, 350-600 The energy percentage of the frequency band, The highest frequency of the signal, in units of .

[0120] when and At that time, the low-pass filter gain is adjusted as follows:

[0121] ;

[0122] In the formula, To adjust the low-pass filter gain, This is the gain of the original low-pass filter.

[0123] when and At that time, the center frequency of the bandpass filter is adjusted to:

[0124] ;

[0125] In the formula, The center frequency of the bandpass filter, in units of .

[0126] when and At that time, the high-pass filter gain is adjusted as follows:

[0127] ;

[0128] In the formula, To adjust the high-pass filter gain, This is the gain of the original high-pass filter.

[0129] The specific implementation of step S07 is as follows: The system constructs a network transmission graph model, and the calculation formula for the graph edge weights is as follows:

[0130] ;

[0131] In the formula, For nodes To the node The directed edge weights, This is a measurement of the transmission delay of this link, in units of... , The historical maximum transmission delay across all links, in units of , This is a measurement of the bandwidth utilization of this link. This represents the historical maximum bandwidth utilization across all links. This is the measured packet loss rate of the link. This represents the historical maximum packet loss rate across all links.

[0132] The mathematical representation of Dijkstra's shortest path algorithm is: for any vertex in the graph Maintain its shortest distance estimate Initial distance of source node ,in The source node is the distance of other nodes. The iterative update process is as follows:

[0133] ;

[0134] In the formula, For nodes The current shortest distance estimate, For nodes To the node The edge weights are determined, and the algorithm iterates until the shortest distance of all nodes is determined, ultimately obtaining the shortest path sequence from the edge computing node to the cloud storage node.

[0135] It should be explained that this invention achieves accurate judgment of signal state through the dynamic calculation of local variance and gradient rate of change values ​​in the adaptive sampling rate adjustment module. Based on the weight vector of weak anomaly features, it guides the adaptive adjustment of the sampling rate within the range of 40% to 180%, ensuring both low-volume acquisition of stable signals and high-fidelity capture of abnormal signals. The semantic-aware compression encoder achieves intelligent identification of fault levels by matching a ladder-style fault mode library vector through similarity calculation. Differentiated encoding strategies are adopted for different levels of fault features, and the quantization step size is dynamically adjusted using the comprehensive dispersion value in the dispersion evaluation matrix to achieve an optimal balance between compression efficiency and information fidelity. The cache management mechanism dynamically adjusts the number of parallel processing threads and cache space allocation by real-time monitoring of cache occupancy growth and decline, avoiding resource waste and handling sudden data traffic. The frequency domain adaptive filtering selects the corresponding filter and adjusts the gain parameters by comparing the energy proportion of different frequency bands in the power spectral density distribution, specifically compensating for the energy attenuation of signals in each frequency band during the compression encoding process. The network transmission graph model calculates edge weights by comprehensively considering transmission latency, bandwidth utilization, and packet loss rate. It optimizes data transmission paths using the Dijkstra shortest path algorithm, minimizes end-to-end transmission latency, and improves transmission reliability, enabling lightweight preprocessing and efficient caching of industrial data at the edge.

[0136] It should be noted that the variables involved in this invention are explained in detail in Tables 1 and 2.

[0137] Table 1. Variable Explanation Table (Part 1)

[0138]

[0139] Table 2. Variable Explanation Table (Part Two)

[0140]

[0141] To better understand and implement this invention, the following is a specific application scenario of the invention, Example 2: The technical team applied edge-side industrial data lightweight preprocessing and caching technology in a rotating machinery equipment monitoring scenario. This equipment is a high-speed centrifugal pump with a rated operating speed of 3600 r / min, and is equipped with multiple sensors for real-time status monitoring. The technical solution is deployed on an edge computing node, responsible for processing four data streams from an accelerometer, a temperature sensor, a pressure sensor, and a current sensor.

[0142] In the initial stage, a Gray code analog-to-digital anti-interference encoder was used to convert the analog signal to eliminate encoding jitter near the quantization boundary. The original sampling frequency was set to 10 kHz, the data acquisition time was 5 minutes, and the generated raw data volume was approximately 300 MB. Gray code encoding reduced the number of transitions caused by single-bit errors by 87%, improving the processing efficiency of subsequent compression algorithms.

[0143] The adaptive sampling rate adjustment module processes the acquired data stream. Based on statistical analysis of historical normal operation data, a weight vector of weak anomaly features is established. During the sampling process, the local variance and gradient rate of change values ​​monitored by the module are divided into three stages: in the low-frequency stable stage (local variance = 0.018, gradient rate of change = 0.032), the sampling rate is reduced to 4000 Hz; in the standard stage (local variance = 0.067, gradient rate of change = 0.098), the original sampling rate of 10 kHz is maintained; in the high-frequency anomaly stage (local variance = 0.156, gradient rate of change = 0.225), the sampling rate is increased to 18 kHz. Figure 2 As shown, after adaptive sampling, the total data volume was reduced from 300 MB to 204 MB, saving 32% of the original data storage space.

[0144] As shown in Table 3, the adaptive sampling signal is input into the wavelet decomposition parallel computing unit for four-level discrete wavelet transform:

[0145] Table 3. Characteristics of Coefficients at Each Level of Wavelet Decomposition

[0146]

[0147] The high-frequency fault feature sparse matrix identified fault features concentrated in the 100 Hz to 300 Hz frequency band, corresponding to 2847 coefficient positions in the high-frequency detail coefficients of the second and third layers. Among them, the coefficients marked as fault features accounted for 8.6% of all coefficients, while the fault information concentration reached 92%.

[0148] During the semantically aware compression coding stage, the technical team queried the stepped fault mode library vectors and found that the similarity between the current signal features and the minor fault level was 0.78. Lossless Huffman coding was used for the fault feature coefficients; arithmetic coding was used for abrupt changes in the fourth-layer low-frequency approximation coefficients where the absolute gradient value exceeded 0.15 (a total of 156); and adaptive quantization compression coding was used for the normal signal coefficients, with the quantization step size adjustment factor dynamically set to 1.0 based on the discreteness evaluation matrix. After completing the compression coding, the data size was further reduced to 48 MB.

[0149] During the cache management phase, the memory occupancy of the edge computing nodes increased by 52% within the first 5 minutes, exceeding the 35% growth threshold. Therefore, the system increased the number of parallel processing threads of the semantic-aware compression encoder to 150% (from 4 threads to 6 threads) and pre-allocated additional cache space to 130% (from 512 MB to 665 MB). Figure 3 As shown, this measure effectively controlled the growth of cache utilization, reducing the peak from 68% to 61%.

[0150] The filtering module performed power spectral density analysis on the compressed encoded data. Energy distribution statistics showed that the energy share was 16% in the 20 Hz to 80 Hz band, 54% in the 100 Hz to 300 Hz band, and 30% in the 350 Hz to 600 Hz band. Since the mid-frequency band had the largest energy share, the system adjusted the center frequency of the bandpass filter to 200 Hz and performed bandpass filtering on the compressed encoded data. Figure 4 As shown, the filtered mid-frequency data is combined with the low-frequency and high-frequency filtered data to form the final filtered data, which totals 52 MB.

[0151] A double-buffering mechanism is used when writing data to the edge-side circular buffer: once the first buffer is 100% full, the process switches to the second buffer to continue writing, with the first buffer marked as data to be transmitted. The network transmission graph model parameterizes the network structure consisting of five nodes: the edge computing node, two intermediate routing nodes, and the cloud storage node. Based on the measured transmission latency, bandwidth utilization, and packet loss rate of each link, the system calculates the weight of each link and uses Dijkstra's shortest path algorithm to calculate the optimal routing sequence. Finally, the data to be transmitted is uploaded to the cloud storage node with minimal transmission latency. The entire processing cycle, from data acquisition to successful caching, takes 352 seconds.

[0152] Compared to traditional fixed sampling rates and uniform compression strategies, this technical solution introduces an adaptive mechanism during the sampling stage to dynamically adjust the sampling frequency based on signal characteristics. This avoids oversampling of stable low-frequency signals and undersampling of abnormal high-frequency signals. Through the architecture design of wavelet decomposition parallel computing units, multi-level frequency domain decomposition is achieved, allowing fault features to be clearly revealed in high-frequency detail coefficients, significantly improving the separation between fault features and normal signals. By finely dividing the signal region using a sparse matrix of high-frequency fault features, the semantically aware compression encoder can employ differentiated encoding strategies for different regions. It maintains high fidelity for coefficients containing key fault diagnosis information and uses a high compression ratio for stable and normal regions, achieving an effective balance between compression efficiency and information preservation. Through a dynamic cache management mechanism, the system can automatically expand or reclaim computing resources based on real-time load, avoiding resource waste or overflow problems that may occur with traditional fixed cache configurations. While ensuring the integrity of fault diagnosis information, the entire technical solution reduces data storage by 84% compared to the traditional method of directly storing raw data, significantly reducing storage pressure and network transmission burden on the edge side.

[0153] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for lightweight preprocessing and caching of edge-side industrial data, characterized in that, include: Industrial sensors are connected to edge computing nodes to collect signals, and the collected signals are input into a Gray code analog-to-digital anti-interference encoder to output a raw digital signal sequence. The raw digital signal sequence is then input into an adaptive sampling rate adjustment module, which adjusts the sampling rate according to a weak anomaly feature weight vector to output an adaptive sampling signal sequence. The adaptive sampling signal sequence is then input into a wavelet decomposition parallel computing unit, which performs a four-layer wavelet transform to mark fault feature coefficients and normal signal coefficients based on a high-frequency fault feature sparse matrix. The fault feature coefficients and normal signal coefficients are then input into a semantic-aware compression encoder. The fault level is selected by querying a ladder-style fault mode library vector, and lossless Huffman coding, arithmetic coding, and quantization compression coding are used. The cache occupancy rate parameter is monitored to dynamically adjust the number of parallel processing threads and cache space. The encoded data is subjected to a fast Fourier transform to extract the frequency band energy ratio value and is filtered to obtain filtered data. This filtered data is written into an edge-side circular buffer, and a network transmission graph model is constructed. The Dijkstra shortest path algorithm is used to transmit the data to a cloud storage node for caching. The high-frequency fault feature sparse matrix refers to a two-dimensional sparse matrix used to identify the distribution of fault features in the high-frequency coefficients after wavelet decomposition. The construction steps are as follows: establish an industrial equipment fault mode database, perform wavelet decomposition on signal samples to count the positions of high-frequency detail coefficients, establish a matrix framework with the time sampling point number as the row index, assign a value of 1 to the fault feature position and assign a value of 0 to the other positions; the wavelet decomposition parallel computing unit performs a four-level wavelet transform on the adaptive sampled signal sequence, outputs the high-frequency detail coefficients of each level and the low-frequency approximation coefficients of the fourth level, and identifies the positions of coefficients containing fault features based on the high-frequency fault feature sparse matrix, and marks the fault feature coefficients and normal signal coefficients. The tiered fault mode library vector refers to a set of fault feature template vectors organized according to the severity of the fault. The construction steps are as follows: fault types are divided into minor fault level, moderate fault level, and severe fault level; feature extraction and cluster analysis are performed on fault samples to calculate the center vector of each cluster as a typical fault mode vector; a semantic-aware compression encoder calculates the similarity value between the fault feature coefficients and each mode vector in the tiered fault mode library vector and selects the fault level corresponding to the mode vector with the largest similarity value; lossless Huffman coding is used for the fault feature coefficients; arithmetic coding is used for the coefficients in the fourth-level low-frequency approximation coefficients where the absolute value of the gradient is greater than the threshold; and quantization compression coding is used for the normal signal coefficients. The weak anomaly feature weight vector refers to a numerical sequence used to quantify the importance of small anomaly changes in a signal. The construction steps are as follows: collect historical signal data under normal operating conditions to obtain the mean sequence and standard deviation sequence; calculate the deviation value of the current sampled signal and divide it by the standard deviation value to obtain the standardized deviation sequence; perform sliding window processing to calculate the sum of the absolute values ​​of the standardized deviations within the window; use the sum of the absolute values ​​as the weak anomaly feature weights of the window center points; and form the weak anomaly feature weight vector by combining the weak anomaly feature weights of all window center points.

2. The method according to claim 1, characterized in that, Gray code analog-to-digital anti-interference encoders refer to encoders that use Gray code encoding to convert analog sensor signals into digital signals. The analog signal is converted into natural binary code by an analog-to-digital converter, and then converted into Gray code by a Gray code conversion circuit.

3. The method according to claim 2, characterized in that, The quantization step size of the quantization compression coding is dynamically adjusted according to the adjustment factor at the corresponding position in the dispersion evaluation matrix. The mapping rule of the adjustment factor is that when the comprehensive dispersion value is in the low value range, the adjustment factor is the first value; when it is in the medium value range, it is the second value; and when it is in the high value range, it is the third value.

4. The method according to claim 3, characterized in that, The discreteness evaluation matrix is ​​a two-dimensional matrix used to evaluate the discreteness of signal coefficients and guide the adjustment of quantization step size. The construction steps are as follows: divide the coefficient sequence of each layer into data blocks according to the time window, calculate the discreteness index for each data block, normalize it, and then sum it with weights to obtain the comprehensive discreteness value and map it to the quantization step size adjustment factor.

Images