Big data analysis-oriented edge artificial intelligence calculation method

Abstract

The invention discloses a big data analysis-oriented edge artificial intelligence calculation method, which relates to the technical field of edge artificial intelligence, and comprises the following steps of: after a boundary aggregation condition is identified, calculating a data distribution compression ratio on the basis of a numerical mapping overlapping degree before and after truncation, a sample clustering overlapping rate of a boundary interval and an upper and lower adjacent section thereof, and a data distribution compression ratio; constructing a feature space overlapping ratio model for identifying whether a distribution folding trend exists or not; and on the basis of the boundary aggregation condition and the identification result of the distribution folding trend, constructing a credibility discrimination evaluation function, generating a credibility discrimination evaluation coefficient, and dividing the current input data into a high credibility area, a middle credibility area or a low credibility area. According to the method, the problems of data distribution folding and misjudgment caused by truncation processing in industrial edge calculation are solved, and dynamic evaluation of input credibility and adaptive regulation and control of a reasoning path are realized, so that the model discrimination accuracy and robustness are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of edge artificial intelligence, and particularly relates to an edge artificial intelligence computing method for big data analysis. BACKGROUND

[0002] The edge artificial intelligence computing for big data analysis refers to a computing method for local real-time processing and intelligent analysis of large-scale heterogeneous data collected from workshop equipment, sensors, industrial control systems, etc. in an industrial production environment, relying on intelligent computing nodes (such as industrial gateways, embedded controllers, sensor-end AI modules, etc.) deployed on the edge side. This method realizes data preprocessing, feature extraction, pattern recognition and prediction judgment functions by integrating lightweight machine learning or deep learning models, effectively avoiding the high delay, network congestion and data leakage risks brought by transmitting all data to the cloud. In the industrial scene, this technology can be widely applied to device fault prediction, quality anomaly detection, energy optimization, production line scheduling, intelligent security, etc. Through local computing, the decision-making efficiency and system response speed are improved, thereby ensuring production continuity and intelligent level, and having high timeliness, high reliability and strong edge intelligent adaptive ability.

[0003] The existing edge artificial intelligence computing technology for big data analysis mainly relies on combining artificial intelligence algorithms with distributed computing architecture to complete local processing and analysis of data on the edge computing nodes in the industrial field. Its implementation usually includes multiple key links. First, in the data acquisition link, real-time data from sensors, PLC controllers, production execution systems (MES), etc. are acquired through industrial Internet of Things devices; second, in the data preprocessing stage, the original data is cleaned, denoised, normalized and time-aligned to ensure the accuracy of subsequent analysis; third, in the model inference link, lightweight AI models (such as compressed convolutional neural networks or decision tree models) deployed in edge devices are used to extract features, identify states and predict trends from data; then, through local decision-making mechanisms, the identification results are quickly responded, such as issuing device warnings or dynamically adjusting parameters; finally, model updating and coordination mechanisms are included, such as using edge-to-edge or edge-to-cloud communication for model incremental learning, federated learning synchronization or strategy updating, so as to adapt to changes and heterogeneity of the industrial field data environment. This distributed, multi-level, real-time processing mode is the core path for existing technologies to realize edge artificial intelligence computing.

[0004] The existing technology has the following deficiencies: In industrial applications, edge artificial intelligence computing methods usually need to process large-scale analog data (such as current, voltage, vibration, etc.) from device sensors locally and quickly to meet the demand for low latency and high efficiency analysis. In this process, in order to reduce the computing pressure and storage burden of the edge device, the original data is usually preprocessed by truncating floating-point to fixed-point. When part of the collected data values are near the truncation boundary, multiple physical quantities that originally have slight differences will be truncated to the same value, thereby forming local folding in the data distribution. Since the edge model is generally modeled based on the assumption of continuous distribution of feature values during the training stage, this kind of distribution folding will destroy the model's true distinguishing ability for working conditions, causing multiple different actual operating states to be incorrectly merged as the same feature at the input layer of the model, thereby triggering false classification judgments. However, existing edge artificial intelligence computing techniques for big data analysis cannot determine whether the current input data has lost sufficient distinguishing degree according to the distribution folding trend caused by boundary aggregation due to data truncation, and accordingly adaptively correct the reasoning process, which may lead to the model incorrectly judging multiple different industrial conditions as the same state, causing false positives and false negatives, and thus affecting the discrimination accuracy and operation reliability of the edge intelligent system in the industrial field.

[0005] The above information disclosed in the background section is only intended to enhance the understanding of the background of the present disclosure, and thus it can include information that does not constitute the prior art known to those of ordinary skill in the art. SUMMARY

[0006] The purpose of the present application is to provide an edge artificial intelligence computing method for big data analysis to solve the problems in the background art.

[0007] To achieve the above-mentioned purpose, the present application provides the following technical solution: an edge artificial intelligence computing method for big data analysis, specifically comprising the following steps: S1, constructing a value distribution map for continuously collected industrial analog data, and identifying the boundary aggregation situation caused by truncation behavior by analyzing the concentration, density gradient, and time series variation trend of the data in the boundary interval formed by setting a symmetric expansion range on the value axis with the value corresponding to the truncation boundary as the center; S2, after the boundary aggregation situation is identified, constructing a feature space overlap rate model based on the value mapping coincidence degree before and after truncation, the sample clustering overlap rate of the boundary interval and its adjacent sections above and below, and the data distribution compression ratio, for identifying whether there is a distribution folding trend; S3, based on the identification results of the boundary aggregation situation and the distribution folding trend, constructing a credible distinguishing degree evaluation function, generating a credible distinguishing degree evaluation coefficient, and dividing the current input data into a high credible region, a medium credible region, or a low credible region; S4, determining an inference path according to the result of the trustworthiness degree evaluation, processing high-trustworthiness region data by using a standard model inference path, processing medium-trustworthiness region data by using a parameter self-adaptive adjustment path, and processing low-trustworthiness region data by using a substitute inference path or a rejection processing mechanism; S5, based on the result of the trustworthiness interval division within a continuous time window, constructing a trustworthiness trend fluctuation sequence, and dynamically adjusting a truncation processing strategy, including a setting mode of a truncation step, a shift mode of a truncation boundary, and a configuration mode of a truncation mapping range, according to the trustworthiness trend fluctuation sequence.

[0008] Preferably, S1 specifically comprises the following steps: Based on a time axis, a fixed-width sliding sampling window is constructed, and industrial analog quantity data continuously collected in each window is divided into continuous numerical sections, and a numerical distribution atlas with numerical sections as a horizontal axis and sample frequencies as a vertical axis is generated in each window; In each numerical distribution atlas, a boundary interval is determined, which is formed by symmetrically expanding a boundary corresponding to a numerical value as a center in the upward and downward directions according to a set numerical span, the number of samples in the boundary interval is counted, and the change value of the frequency of the boundary interval in the current window and in a plurality of historical windows is calculated, which is used to obtain the data concentration in the boundary interval; The sample frequency in the boundary interval and in the adjacent equal-width sections above and below the boundary interval are compared, the density gradient difference value is calculated, and the change sequence of the difference value is extracted on a plurality of continuous windows to construct a time sequence change trend of the density gradient; When the data concentration in the boundary interval exceeds the sample mean multiple threshold value, the density gradient difference value is higher than the preset critical value, and the time sequence change trend meets the stable or enhanced change condition, it is identified that the boundary interval has a boundary aggregation condition caused by the truncation behavior.

[0009] Preferably, S2 specifically comprises the following steps: After the boundary aggregation condition is identified, the historical floating-point original data in the boundary interval and the corresponding fixed-point data after truncation are extracted, the coincidence rate of the numerical landing points before and after truncation is calculated based on the numerical mapping relationship between the original data and the truncated data, and the numerical mapping coincidence degree is taken as the numerical mapping coincidence degree; Based on the sample set in the boundary interval and in the adjacent equal-width sections in the upward and downward directions, an unsupervised clustering operation is performed in the original feature space, the cross-proportion of the samples in the boundary interval and the adjacent sections in the clustering labels is calculated, and the sample clustering overlap rate is taken as the sample clustering overlap rate; The sample data of the boundary interval and the adjacent sections before and after the truncation processing are extracted, the original physical quantity parameter sequences constituting the feature space are taken as the feature vector sets, and the mean change rate of the Euclidean distance between the samples in each set is calculated, which is taken as the data distribution compression ratio for measuring the space contraction effect; The numerical mapping coincidence degree, sample cluster overlap rate and data distribution compression ratio are normalized and input into a weighted fusion function to construct a feature space overlap rate model, and an integrated overlap score value is output; The integrated overlap score value is compared with a preset distribution folding identification threshold value, and when the score value is higher than the threshold value, it is judged that there is significant overlap in the feature space for the current boundary interval, and it is identified that there is a distribution folding trend.

[0010] Preferably, S3 specifically comprises the following steps: The data concentration, density gradient change trend in the boundary aggregation condition identification result, and the numerical mapping coincidence degree, sample cluster overlap rate and data distribution compression ratio in the distribution folding trend identification result are extracted as a set of trusted distinguishability evaluation factors; Each index in the set of trusted distinguishability evaluation factors is respectively standardized to have a zero mean and a unit variance, to obtain a standardized factor sequence with a unified numerical scale; A linear weighted fusion method is applied to combine the standardized factor sequence into a trusted distinguishability evaluation function according to a preset weight, and a single-valued trusted distinguishability evaluation coefficient is output, which is used to quantify the discriminability of the current input data in the feature space; According to the numerical interval of the trusted distinguishability evaluation coefficient, the current input data is divided into a high trusted area, a medium trusted area or a low trusted area, which is used to drive the hierarchical processing flow of the subsequent reasoning path.

[0011] Preferably, the specific way of constructing the trusted distinguishability evaluation function comprises: The data concentration, density gradient change trend, numerical mapping coincidence degree, sample cluster overlap rate and data distribution compression ratio are respectively subjected to mean normalization to construct a set of standardized factors with a unified numerical scale; The importance between each standardized factor is weighted based on the historical working condition classification accuracy in the industrial field, and an iterative optimization algorithm is used to determine the fusion weight of each standardized factor; The standardized factors and their corresponding weights are combined using a linear weighted fusion method to construct a trusted distinguishability evaluation function with clear expression structure and calculation path, and a trusted distinguishability evaluation coefficient is output through the function, which is used to reflect the discriminability of the current input data in the feature space.

[0012] Preferably, according to the numerical interval of the trusted distinguishability evaluation coefficient, the current input data is divided into a high trusted area, a medium trusted area or a low trusted area, specifically: The upper threshold and lower threshold of the trusted distinguishability evaluation coefficient are set, and the threshold boundary is determined by the reasoning accuracy change trend of the training data set under different trusted distinguishability evaluation coefficients; When the trustworthiness differentiation evaluation coefficient is higher than the upper threshold, the current input data is divided into a high trustworthiness area; when the trustworthiness differentiation evaluation coefficient is lower than the lower threshold, the current input data is divided into a low trustworthiness area; when the trustworthiness differentiation evaluation coefficient is between the upper threshold and the lower threshold, the current input data is divided into a medium trustworthiness area; The division result is bound with the input data characteristics and stored in a state buffer as an input basis for subsequent reasoning path adaptive scheduling strategies.

[0013] Preferably, S4 specifically comprises: The trustworthiness differentiation evaluation coefficient corresponding to the current input data is received, and a corresponding reasoning path type is selected from a preset reasoning path set according to a trustworthiness interval label corresponding to the trustworthiness differentiation evaluation coefficient, as an entry of a reasoning processing flow of the current input data; If the current input data is divided into a high trustworthiness area, a standard model reasoning path is called, and the standard model reasoning path is executed to obtain an analysis result under the condition that the structure and parameter configuration of the edge artificial intelligence calculation model remain unchanged; If the current input data is divided into a medium trustworthiness area, a parameter adaptive adjustment path is called, and the parameter weight of the edge artificial intelligence calculation model is dynamically reset based on the statistical characteristic information of the current input data to form a parameter adaptive model for the characteristic of the current input data, and the parameter adaptive adjustment path is executed to obtain an analysis result; If the current input data is divided into a low trustworthiness area, a substitute reasoning path or a rejection processing mechanism is selected and called according to a system configuration strategy, wherein the substitute reasoning path calls an independently constructed redundant edge artificial intelligence calculation model for analysis and processing, and the rejection processing mechanism triggers an abnormal feedback flow and interrupts the current reasoning process to prevent misjudgment caused by too low trustworthiness differentiation; The reasoning path type and the reasoning result are bound and recorded in a reasoning execution log for subsequent reasoning flow evaluation and path scheduling strategy optimization module to perform incremental learning and strategy updating.

[0014] Preferably, S5 specifically comprises the following steps: A continuous time window sequence with a fixed length is constructed based on a time axis, the trustworthiness interval division result corresponding to the current input data is recorded in each time window, and a trustworthiness interval division track is formed by arranging the trustworthiness interval division result in time sequence; The division result of each time window in the trustworthiness interval division track is encoded and mapped to construct a discrete state sequence, and a sliding window statistical method is used to calculate the occurrence frequency variation trend of the high trustworthiness area, the medium trustworthiness area and the low trustworthiness area in the sequence to generate a trustworthiness trend fluctuation sequence; According to the trend change characteristics in the credible trend fluctuation sequence, the overall discriminant stability of the current stage system to the input data is determined, and when the frequency in the low credibility area continuously increases or the fluctuation is enhanced, it is judged that the current truncation processing strategy has a weakening effect on the feature space expression. When it is judged that the current truncation processing strategy has a weakening effect, a preset parameter optimization module is called to dynamically adjust the truncation step setting mode in the truncation processing strategy, and the numerical retention precision is improved by reducing the truncation step. The offset mode of the truncation boundary is adjusted synchronously, the coincidence trend of the truncation boundary and the high-density data area is broken by introducing offset disturbance, and the data mapping is uniformly reconfigured to alleviate the boundary aggregation and distribution folding problem.

[0015] In the above technical solution, the technical effects and advantages provided by the present application are as follows: 1. The present application innovatively structures the boundary aggregation and distribution folding problem caused by the edge truncation processing of industrial analog quantity data, quantifies and feedbacks the problem, and significantly enhances the adaptive discriminant ability of the edge artificial intelligence system under complex working conditions. The boundary aggregation caused by the truncation behavior is analyzed by constructing a numerical distribution map, and a feature space overlap rate model is established by combining the numerical mapping overlap degree, the sample clustering overlap rate and the feature space compression ratio, so that the distribution folding trend is accurately determined. The discriminant ability of the input data is quantified by a credible area discriminant function, and the reasoning path is dynamically selected accordingly to avoid misjudgment caused by truncation interference, so that a closed-loop processing mechanism of "recognition + control" is realized.

[0016] 2. The present application has the following technical advantages: without changing the original edge model structure, the distribution disturbance caused by truncation is perceived, quantified and intervened, and the model fault tolerance is improved; by introducing composite indexes such as standardization fusion, clustering overlap degree and gradient change trend, the recognition accuracy of implicit data distortion is improved; by means of multi-level reasoning path selection mechanism (standard model, parameter self-adaptation, alternative path or rejection mechanism), the data input of different credibility levels is dynamically adapted, and false positives or false controls are effectively avoided; a credible trend fluctuation sequence is constructed, and the truncation strategy is adjusted accordingly, so that the edge computing process has a self-closed loop adjustment ability of "input-reasoning-feedback-optimization", thereby significantly improving the stability, accuracy and engineering robustness of edge artificial intelligence computing in industrial scenes. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0018] Figure 1 A flowchart of an edge artificial intelligence computing method for big data analysis according to the present application. DETAILED DESCRIPTION

[0019] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations may, however, be implemented in many different forms and should not be construed as limited to the examples set forth herein; rather, these example implementations are provided so that this disclosure will be thorough and complete, and will fully convey the scope of example implementations to those skilled in the art.

[0020] The present application provides an edge artificial intelligence computing method for big data analysis as shown in Figure 1 The edge artificial intelligence computing method for big data analysis according to the present application specifically includes the following steps: S1, constructing a numerical distribution atlas for continuously collected industrial analog quantity data, and identifying boundary aggregation caused by truncation behavior by analyzing the concentration, density gradient and time series trend of data in a boundary interval formed by setting a symmetric expansion range on the numerical axis with the corresponding numerical value of the truncation boundary as the center; In this embodiment, S1 specifically includes the following steps: Based on the time axis, a sliding sampling window with a fixed width is constructed, the continuously collected industrial analog quantity data in each window is divided into consecutive numerical sections, and a numerical distribution atlas is generated in each window with the numerical section as the horizontal axis and the sample frequency as the vertical axis; In industrial edge artificial intelligence computing, based on the time axis, a sliding sampling window with a fixed width is constructed and a numerical distribution atlas is generated, which can be achieved in the following way: first, set the time length of the sampling window, for example, in units of 1 second or 500 milliseconds, use the edge collection device to collect industrial analog quantity data (such as current, voltage, vibration signal) at a high frequency and continuously, and slice the data according to the set window length; then in each sampling window, determine the minimum and maximum values of the data value range, divide the entire numerical range into multiple numerical sections with equal width, such as 0.05 per section; then count the frequency of each data point falling into each section in the current window to construct a histogram structure of the current window; finally, encode the histogram with the numerical section as the horizontal axis and the corresponding sample frequency as the vertical axis to form the numerical distribution atlas of the current window. This atlas can accurately reflect the dispersion characteristics and distribution changes of industrial analog quantity data in a specific time period, supporting subsequent aggregation identification and dynamic analysis of boundary behavior.

[0021] In each numerical distribution atlas, determine the boundary interval formed by symmetrically expanding in the upward and downward directions according to the set numerical span with the corresponding numerical value of the truncation boundary as the center, count the number of samples in the boundary interval, and calculate the frequency change value in the current window and in multiple historical windows to obtain the concentration of data in the boundary interval; In each numerical distribution map, the data concentration recognition of the boundary interval can be achieved in the following way: first, according to the truncation rule adopted in the current industrial edge device, the numerical value corresponding to the truncation boundary is extracted (for example, in fixed-point truncation, the boundary value is 2.0, 2.1, 2.2, etc. if one decimal place is retained), then a certain truncation boundary value is taken as the center, and a symmetrical upper and lower expansion span (for example, ±0.05) is set, so as to form a boundary interval (for example, taking 2.1 as the center, an interval of [2.05, 2.15] is formed); then the total number of samples in the numerical segment covered by the boundary interval is counted in the current numerical distribution map, and the frequency in the current window is obtained; at the same time, the sample frequency of the same interval is counted in the previous continuous several time windows, and the mean value and standard deviation are calculated; the frequency in the current window is compared with the historical mean value, if the concentration degree is much higher than the average level, the concentration degree can be quantified by the formula "concentration degree = current frequency / historical mean value" or "Z-score = (current frequency - mean value) / standard deviation"; for example, if the average number of samples of a certain boundary interval in the last 5 windows is 8, and the current window is 24, then the concentration degree is 3, which indicates that there is a significant data accumulation in the interval, indicating the emergence of boundary aggregation trend.

[0022] The sample frequency in the boundary interval and its adjacent equal-width segments above and below are compared, the density gradient difference value is calculated, and the change sequence of the difference value is extracted on continuous multiple windows to construct the time sequence change trend of the density gradient; When calculating the density gradient difference value of the boundary interval and constructing the time sequence change trend, the following method can be used: first, after determining the boundary interval, select one equal-width adjacent segment above and below it on the numerical axis as the comparison reference, for example, the boundary interval is [2.05, 2.15], then the adjacent [1.95, 2.05] and [2.15, 2.25] can be selected as the comparison segments; the sample frequencies in the three segments are counted respectively, the frequency difference between the boundary interval and the adjacent segments is calculated, and the density value difference, i.e. the density gradient difference value, is obtained by dividing the segment width; specifically, the calculation can be performed by "density gradient difference value = boundary interval frequency / interval width - adjacent segment frequency / interval width"; then the above calculation is repeated in continuous multiple sampling windows, and the density gradient difference value corresponding to each window is recorded in time sequence to form a time sequence; by calculating the change trend of the time sequence (such as first-order difference, moving average slope or regression fitting trend direction), whether the density gradient of the boundary interval presents rising, falling or fluctuating enhancement characteristics can be judged; for example, in five continuous windows, if the density of the boundary interval continuously increases from 1.5, 2.1, 2.8, 3.4 to 4.0, it indicates that the density gradient presents an enhancement trend, which can assist in judging that the boundary aggregation effect is being formed.

[0023] When the data concentration within the boundary interval exceeds a threshold multiple of the sample mean, the density gradient difference is higher than a preset critical value, and the time series change trend meets the conditions of stable or enhanced change, the boundary interval is identified as having a boundary clustering situation caused by truncation behavior.

[0024] To identify boundary clustering caused by truncation, a strategy based on a triple threshold judgment mechanism can be adopted. First, the sample frequency within the boundary interval is compared with the mean sample frequency of the corresponding interval across multiple historical windows. If the current frequency exceeds a multiple threshold of the historical mean (e.g., set to 2 or 3 times), a significant clustering phenomenon is considered to have occurred in that interval. This multiple threshold can be dynamically set based on statistical analysis of a large amount of historical industrial data, preventing overreaction to short-term fluctuations while promptly capturing abnormal clustering trends. Second, the density gradient difference must be higher than a preset critical value, which is used to determine the density jump magnitude between the boundary interval and its adjacent segments. The critical value can be obtained by multiplying the historical standard deviation of sample density by an empirical coefficient (e.g., 1.5 or 2.0) to ensure that the gradient change in the boundary interval is statistically significant. Third, the time series of the density gradient must show a stable or enhancing trend, i.e., the gradient value increases steadily over multiple consecutive time windows, with small fluctuations or a positive slope, indicating that the clustering effect is persistent rather than a transient disturbance.

[0025] The design of jointly judging these three indicators aims to comprehensively capture the true boundary clustering phenomenon caused by truncation from three dimensions: intensity (concentration), amplitude (gradient difference), and evolutionary trend (time series changes), thereby effectively suppressing isolated misjudgments or misidentifications caused by random perturbations. This combined judgment method avoids the problem of excessive model sensitivity caused by relying on a single feature and also enhances the robustness and universality of the method under complex industrial conditions. Through this approach, not only can boundary clustering areas be identified efficiently, but a clear foundation can also be provided for subsequent credibility assessment and adaptive inference path selection.

[0026] S2. After the boundary clustering situation is identified, a feature space overlap rate model is constructed based on the degree of overlap of numerical mapping before and after truncation, the sample clustering overlap rate of the boundary interval and its adjacent segments above and below, and the data distribution compression ratio, in order to identify whether there is a distribution folding trend. In this embodiment, S2 specifically includes the following steps: After the boundary clustering situation is identified, the historical floating-point raw data and the corresponding truncated fixed-point data within the boundary interval are extracted. Based on the numerical mapping relationship between the raw data and the truncated data, the overlap ratio of the numerical landing points before and after truncation is calculated as the degree of numerical mapping overlap. In edge artificial intelligence computing, to calculate the degree of coincidence of value mapping in the boundary interval, first, the original floating-point analog quantity data recorded in the same sampling time period and the fixed-point data after truncation processing are obtained, and one-to-one correspondence is performed according to the numerical order. The implementation can be achieved by retaining the original and truncated versions of the sampling cache at the same time, or using a backtracking mechanism to recover the floating-point data from the pre-compression cache area. Subsequently, according to the corresponding fixed-point value of each original floating-point number under the truncation rule, a value mapping comparison table is established. On this basis, each group of data in the boundary interval is traversed, the number of original values falling on the same fixed-point value is counted, and the ratio operation is performed with the total number of samples in the original value interval covered by the fixed-point value, to obtain the value mapping coincidence rate of the interval. For example, if the fixed-point value of a certain interval is 3.10, its corresponding original floating-point mapping interval is [3.10, 3.15), and there are actually 98 original samples truncated to 3.10, then the mapping coincidence rate is 98 / 100 = 0.98. When the rate is close to 1, it indicates that the truncation has caused a large number of similar but not identical original values to be combined into the same truncated value, revealing the "pseudo-consistency" problem caused by boundary aggregation, which constitutes a potential inducement of distribution folding. This index quantitatively depicts the degree of damage to the resolution of the value space caused by truncation, providing a criterion basis for subsequent overlap analysis.

[0027] Based on the boundary interval and the sample set in the adjacent equal-width segments above and below the boundary interval, unsupervised clustering operation is performed in the original feature space, and the cross-proportion of the clustering labels of the samples in the boundary interval and the adjacent segment samples is calculated as the sample clustering overlap rate. To achieve the calculation of the sample clustering overlap rate, first, based on the boundary interval and its upper and lower adjacent two equal number segments, the original floating point data samples in the corresponding time window are extracted respectively, and the original feature space containing multiple physical feature dimensions (such as voltage, current, vibration frequency, etc.) is constructed. Then, in the joint feature space, an unsupervised clustering algorithm such as K-means, DBSCAN or Gaussian Mixture Model is used to uniformly cluster all samples, outputting the clustering label to which each sample belongs. In the clustering result, the distribution proportion of the samples in the boundary interval and the upper and lower adjacent segments in each cluster is counted respectively to identify the label overlap. The specific calculation method is as follows: let the total number of boundary interval samples be A, of which a samples share the same clustering label with the upper segment samples, and b samples share the same label with the lower segment samples, then the sample clustering overlap rate can be defined as (a+b) / A. For example, if there are 100 samples in the boundary interval, of which 30 are in the same class as the upper segment and 25 are in the same class as the lower segment, then the overlap rate is 55%. This index reflects the fuzzy attribution degree of the boundary interval samples in the clustering structure. If the overlap rate is high, it means that the feature overlap caused by truncation compression makes it difficult for the model to distinguish different physical states, thereby misleading the model decision.

[0028] For the sample data of the boundary interval and its upper and lower adjacent segments before and after truncation, the original physical quantity parameter sequences constituting the feature space are extracted as feature vector sets, and the mean change rate of the Euclidean distance between samples in each set is calculated as the data distribution compression ratio to measure the space contraction effect; To calculate the data distribution compression ratio, first, the original physical quantity parameter sequences of each sample in the boundary interval and its upper and lower adjacent segments before and after truncation, such as current, voltage, temperature, vibration frequency, etc., are extracted to form corresponding multi-dimensional feature vector sets. Then, for the two sets of feature vectors before and after truncation, the pairwise Euclidean distances between all samples in the same feature space are calculated to obtain the distance matrix of each set. Then, the mean of the distance matrix of each set is calculated, denoted as D1 (before truncation) and D2 (after truncation) respectively. The data distribution compression ratio can be represented as D2 / D1, reflecting the compression degree of the feature space scale by truncation. For example, if the average distance between samples before truncation is 0.62, and after truncation it decreases to 0.28, then the compression ratio is 0.28 / 0.62≈0.45, indicating that the truncation behavior significantly reduces the difference between features. This index can quantify the damage to the recognition ability of the original feature space caused by truncation, and is a key basis for measuring the feature folding trend. When the compression ratio is lower than the set threshold, it means that the distinguishability between samples has been greatly reduced due to truncation, which may cause the model to confuse the recognition of different working conditions and affect the reasoning accuracy.

[0029] "Extracting the original physical quantity parameter sequence constituting the feature space as a feature vector set" refers to constructing a feature vector containing multiple physical dimensions for each sampling point before calculating the distribution change between the boundary interval and its adjacent sections. The implementation is as follows: Extract the original floating-point data corresponding to each time point from the edge collection system in the industrial field, including current, voltage, temperature, pressure, vibration frequency, and other key physical quantities, and combine these numerical values of different dimensions into a fixed-length numerical vector according to the preset order. For example, if the data collected at a certain time is current = 3.14A, voltage = 220.5V, vibration = 0.28g, temperature = 68.2℃, then the corresponding four-dimensional feature vector is [3.14, 220.5, 0.28, 68.2]. Converting the data of all samples in the continuous time window into such multi-dimensional vectors forms a sample set, which constitutes the basis for representing the feature space. The key to this step is to ensure consistency in dimensions, order, and units for all samples, thereby ensuring the physical comparability and computational stability of the Euclidean distance, which is a prerequisite for subsequent calculation of the space compression ratio.

[0030] The numerical mapping coincidence degree, sample clustering overlap rate, and data distribution compression ratio are normalized and input into a weighted fusion function to construct a feature space overlap rate model, and output a comprehensive overlap score value; To construct the feature space overlap rate model, the numerical mapping coincidence degree, sample clustering overlap rate, and data distribution compression ratio are first normalized. Specifically, for each index, set the minimum and maximum value intervals, and convert the original value to a dimensionless value between 0 and 1 according to the formula x' = (x - min) / (max - min) to eliminate the interference of dimension differences on subsequent fusion calculations. After normalization, the three indicators are input into the pre-defined weighted fusion function to construct a comprehensive score model. The weighted fusion function can use a linear weighting form S = w_1A + w_2B + w_3C, where A is the normalized coincidence degree, B is the overlap rate, C is the compression ratio, and w_1, w_2, w_3 are weight coefficients obtained through experience or training optimization. The output S is the comprehensive overlap score value, which is used to measure the overlap degree caused by truncation in the feature space. For example, if A = 0.85, B = 0.62, and C = 0.71 in a certain window, and the weights are 0.4, 0.3, and 0.3 respectively, then S = 0.4 x 0.85 + 0.3 x 0.62 + 0.3 x 0.71 = 0.748. When S exceeds the set threshold (such as 0.7), it can be preliminarily judged that there is a significant distribution folding risk, which serves as the basis for subsequent identification.

[0031] Normalization is to convert values with different dimensions or value ranges into a unified dimensionless interval, which is often used in data fusion preprocessing to avoid the dominance of a certain indicator in the fusion due to value differences. Common methods include min-max normalization and Z-score standardization. Weighted fusion function is to assign different weight coefficients to multiple normalized indicators according to their relative importance, and calculate a comprehensive value representing the overall trend of the characteristics through weighted summation. This function can be set by expert experience or adjusted adaptively through machine learning algorithm, and is widely used in multi-index synthesis judgment.

[0032] The comprehensive overlap score value is compared with a preset distribution folding recognition threshold. When the score value is higher than the threshold, it is judged that the current boundary interval has significant overlap in the feature space, and it is recognized as having a distribution folding trend.

[0033] In the execution of distribution folding trend recognition, the following method can be used to achieve this step: first, the system obtains the comprehensive overlap score value calculated by the current window, that is, a dimensionless indicator between 0 and 1 output by the weighted fusion function. Then, the score value is compared with the preset distribution folding recognition threshold one by one. If the score value is higher than the threshold, it means that the current boundary interval shows obvious numerical mapping overlap, clustering overlap and data compression effect in the feature space, which are highly correlated when all three are enhanced at the same time, constituting a typical "distribution folding" feature, that is, multiple original states are misclassified as a single class in the feature input layer, so it can be recognized as having a distribution folding trend. For example, when the score value is 0.78 and the threshold is set to 0.72, the system can directly judge that the current data region has a strong folding trend and mark it as a high-risk input interval for dynamic adjustment of the subsequent processing path.

[0034] As for the "preset distribution folding recognition threshold", the value comes from the statistical modeling process of a large amount of industrial scene historical data. The specific method is as follows: in multiple typical working conditions, mark the data segments known to have distribution folding and those known not to have distribution folding, and calculate their comprehensive overlap score value interval distribution. Through ROC curve or minimum false positive rate strategy, select a threshold that is the best balance between recognition accuracy and sensitivity as the judgment boundary, for example, select the critical score value when the accuracy rate reaches 90% and the recall rate exceeds 85% as the standard, and finally set the threshold to 0.72. This threshold can be fine-tuned as needed and supports automatic dynamic optimization of the subsequent online learning mechanism to adapt to the differences brought by different industrial equipment and working conditions.

[0035] S3, based on the recognition results of the boundary aggregation situation and the distribution folding trend, a reliable distinguishability evaluation function is constructed, a reliable distinguishability evaluation coefficient is generated, and the current input data is divided into a high reliable area, a medium reliable area or a low reliable area; In this embodiment, S3 specifically includes the following steps: Extract the data concentration, density gradient change trend in the boundary aggregation situation recognition result, and the value mapping coincidence degree, sample clustering overlap rate and data distribution compression ratio in the distribution folding trend recognition result as the trusted distinguishability evaluation factor set; The way to realize the extraction of the data concentration, density gradient change trend, value mapping coincidence degree, sample clustering overlap rate and data distribution compression ratio as the trusted distinguishability evaluation factor set can be performed in the following steps: first, after completing the boundary aggregation situation recognition, the sample distribution indicators calculated in the boundary interval are called to directly extract the data concentration value obtained by the sample frequency statistics in the sliding window; at the same time, the gradient change curve between the sample densities of the symmetric segments above and below the boundary interval is obtained, and the change slope is extracted as the density gradient change trend. Secondly, in the distribution folding trend recognition process, the proportion of the same landing points before and after truncation is calculated as the value mapping coincidence degree by using the corresponding relationship between the original data before and after truncation and the fixed-point data mapping. Subsequently, the unsupervised clustering algorithm such as DBSCAN or KMeans is executed in the original feature space, the cross proportion of the samples between the boundary interval and the adjacent segments in the clustering label is counted, and the sample clustering overlap rate is generated. Finally, based on the comparison of the mean values of the Euclidean distances of the original data and the truncated data in the feature space, the feature vector contraction degree is calculated, and the data distribution compression ratio is obtained. The above five results are uniformly collected to form the trusted distinguishability evaluation factor set, which is used for the subsequent standardization and evaluation function construction process.

[0036] Each index in the trusted distinguishability evaluation factor set is standardized by zero mean and unit variance to obtain a standardized factor sequence with uniform numerical scale; The implementation of the zero mean and unit variance standardization processing of each index in the trusted distinguishability evaluation factor set can be completed by the following way: first, for each evaluation factor (such as data concentration, density gradient change trend, etc.), the full-sample mean and standard deviation in the historical training data are counted. Then, for each factor value input, the standardization conversion is performed, and the calculation formula is: standardized value = (original value - mean) ÷ standard deviation. The core idea of this processing method is to convert the original distribution of different factors into a standard normal distribution with mean of 0 and standard deviation of 1, so that each index has consistency in numerical scale, avoids the bias problem in subsequent weighted fusion caused by different dimensions or numerical magnitude differences, and improves the fairness and overall stability of the evaluation function in the fusion of different factors. This standardization processing method is widely used in machine learning and data fusion scenarios, and is especially suitable for multi-index integrated analysis, which is a key step for building a trusted distinguishability evaluation function.

[0037] The linear weighted fusion method is applied to combine the standardized factor sequences according to preset weights to construct a reliable distinguishability evaluation function, and a single-valued reliable distinguishability evaluation coefficient is output, which is used to quantify the distinguishability of the current input data in the feature space. The linear weighted fusion method is applied to combine the standardized factor sequences according to preset weights to construct a reliable distinguishability evaluation function, and a single-valued reliable distinguishability evaluation coefficient is output, which is used to quantify the distinguishability of the current input data in the feature space.

[0038] According to the numerical interval of the reliable distinguishability evaluation coefficient, the current input data is divided into a high reliable area, a medium reliable area or a low reliable area, which is used to drive the hierarchical processing flow of the subsequent inference path.

[0039] In this embodiment, the specific way of constructing the reliable distinguishability evaluation function includes: The mean normalization processing is performed on the data concentration, the density gradient change trend, the numerical mapping overlap degree, the sample clustering overlap rate and the data distribution compression ratio, respectively, to construct a standardized factor set with a unified numerical scale. The mean normalization processing is respectively performed on the data concentration, the density gradient change trend, the numerical mapping coincidence degree, the sample clustering overlap rate and the data distribution compression ratio. The mean normalization processing can be implemented in the following manner: first, a large number of samples of each index are extracted from historical industrial operation data, and the global mean and range of each index are calculated. Subsequently, for each current index value, the formula “(current value - mean) / range” is used for normalization operation, so that all index values are mapped to similar numerical scale intervals, usually between -1 and 1 or 0 and 1. The core purpose of the mean normalization processing is to eliminate the influence of different indicators due to the difference in dimension and value range, so that the indicators have equal comparison basis in the subsequent fusion process, and avoid that some high-level indicators have a dominant effect on the evaluation function. In addition, this processing method facilitates the sensitive identification of abnormal deviation behavior by the algorithm, helps to improve the stability and interpretability of the subsequent credible distinguishability evaluation function, and ensures that the influence of each evaluation factor in the feature space is balanced.

[0040] The importance of each standardized factor is weighted based on the classification accuracy of historical working conditions in the industrial field, and an iterative optimization algorithm is used to determine the fusion weight of each standardized factor. The importance of each standardized factor is weighted based on the classification accuracy of historical working conditions in the industrial field, and a supervised factor contribution evaluation method can be used to achieve this. Specifically, first, a large number of historical industrial working condition data sets containing known classification labels are constructed, and in this data set, each sample corresponds to a set of standardized factor values and its true classification result. Subsequently, a classification model (such as logistic regression or gradient boosting tree) is used to train the standardized factor set, and the classification accuracy changes of the model under different factor combinations are recorded. By comparing the marginal contribution of each factor in the model prediction performance through cross-validation, the importance index of each factor can be quantified as the basis for subsequent weighting. This process can effectively capture the influence of different evaluation factors on the final credible distinguishability discrimination result, realize objective weight learning based on working condition discrimination accuracy, and improve the response ability of the evaluation function to key difference features.

[0041] The fusion weight of each normalization factor is determined by an iterative optimization algorithm, which can be implemented by using a linear regression minimization scheme based on gradient descent. The specific process is as follows: initially, the weight values of each factor are randomly set, a reliable distinguishability evaluation function is constructed, and the predicted reliability coefficients of all training samples are calculated, and the error is calculated with the historical working condition classification label; then, the gradient of each weight is calculated based on the error, and the weight is updated according to the learning rate; repeat the process until the loss function converges or the maximum number of iterations is reached. Common algorithms used in this process include batch gradient descent (BGD), stochastic gradient descent (SGD), or adaptive moment estimation (Adam), etc. The advantage of the iterative optimization algorithm is that it can find the optimal weight configuration in the high-dimensional evaluation factor space to improve the discriminant efficiency and generalization ability of the evaluation function.

[0042] The linear weighted fusion method is used to combine the normalization factors and their corresponding weights to construct a reliable distinguishability evaluation function with clear expression structure and clear calculation path, and the reliable distinguishability evaluation coefficient is output through the function, which is used to reflect the distinguishability of the current input data in the feature space.

[0043] The implementation of using linear weighted fusion method to combine the normalization factors and their corresponding weights can be completed based on vector dot product operation. Specifically, first, all normalized reliable distinguishability evaluation factors are combined into a factor vector, and the weight value corresponding to each factor is combined into a weight vector, then the two vectors are weighted and summed, and a single numerical value is output as the reliable distinguishability evaluation coefficient of the current input data. This method is simple and clear, and has good interpretability and computability. The reason for choosing linear weighted fusion is that it can directly reflect the contribution of each factor to the final reliable evaluation result, effectively reduce the model complexity, facilitate subsequent real-time deployment and fast calculation in edge computing scenarios, and ensure that the overall evaluation mechanism has high controllability and engineering feasibility.

[0044] The reliable distinguishability evaluation function with clear expression structure and clear calculation path can be formalized as a linear combination function, that is, the reliable distinguishability evaluation function F(x) = w1·x1+ w2·x2+... + w n ·x n , where x1 to x n are the normalized factors, w1 to w nTo train the weight coefficients. In the implementation process, first, the factors are unified in scale through standardization processing, and then the preset or trained weights are embedded into the function template to form a complete evaluation function structure through linear weighting. In actual operation, after receiving new input data each time, the corresponding standardized factor value is automatically extracted and substituted into the above function structure for weighted operation, so that the reliable distinguishability evaluation coefficient can be quickly obtained. The numerical size of the evaluation coefficient is used to quantify the distinguishability of the input data in the feature space, providing a basis for subsequent reasoning path selection.

[0045] Linear weighted fusion is a function construction method that combines multiple input factors with the same dimension or after unified normalization processing according to the specified proportion. The core idea is to assign a numerical weight to each input factor, and use these weights to reflect the importance of the factors, and then add the weighted factors to obtain an output value that can be used for decision-making. This method is widely used in multi-index evaluation and has the advantages of high computational efficiency, clear structure, and strong model interpretability. In the reliable distinguishability evaluation scenario, linear weighted fusion not only accurately represents the quantitative contribution of each factor to the reliability, but also facilitates dynamic optimization of model performance through weight adjustment mechanism, adapting to different operating environments and working conditions in industrial sites.

[0046] In this embodiment, according to the numerical interval of the reliable distinguishability evaluation coefficient, the current input data is divided into high reliability area, medium reliability area or low reliability area, specifically: The upper threshold and lower threshold of the reliable distinguishability evaluation coefficient are set, and the threshold boundary is determined by the inference accuracy change trend of the training data set under different reliable distinguishability evaluation coefficients; The upper threshold and lower threshold of the reliable distinguishability evaluation coefficient can be achieved by statistical modeling and performance regression analysis on industrial historical data set. The specific method includes: first, based on the labeled complete training data set, calculate the corresponding reliable distinguishability evaluation coefficient of each sample data, and record the actual inference accuracy of the sample in the edge artificial intelligence model; then, segment the evaluation coefficients and inference accuracy of all samples by interval, draw the relationship curve between the evaluation coefficient and the inference accuracy; then determine the critical influence range of the evaluation coefficient on the inference effect through fitting analysis, and identify the boundary point where the model performance appears significant decline or tends to be stable. Finally, set the evaluation coefficient at which the inference accuracy begins to rapidly decline as the lower threshold, and set the evaluation coefficient at which the inference accuracy tends to be saturated as the upper threshold. Through this precision response trend curve method based on actual inference effect, the reliable interval division threshold can be scientifically and objectively set, ensuring that the subsequent distinguishability division is both discriminative and adaptive, improving the decision reliability of the entire edge intelligent system in industrial scenarios.

[0047] When the trustworthiness differentiation evaluation coefficient is higher than the upper threshold, the current input data is divided into a high trustworthiness area; when the trustworthiness differentiation evaluation coefficient is lower than the lower threshold, the current input data is divided into a low trustworthiness area; and when the trustworthiness differentiation evaluation coefficient is between the upper threshold and the lower threshold, the current input data is divided into a medium trustworthiness area. The current input data is divided into a high trustworthiness area, a medium trustworthiness area or a low trustworthiness area based on the trustworthiness differentiation evaluation coefficient, which can be achieved by a segmented threshold mapping function. Specifically, the numerical values of the upper threshold and the lower threshold are first fixed in the system, and the corresponding trustworthiness differentiation evaluation coefficient is calculated in real time after the input data stream passes through the trustworthiness differentiation evaluation function. Then, the evaluation coefficient is compared with the two thresholds logically: if the value is greater than the upper threshold, it is determined that the current input data has significant feature distinguishability, and is classified into the high trustworthiness area; if the value is lower than the lower threshold, it is determined that the current data has a serious limitation in the expression ability in the feature space, and is classified into the low trustworthiness area; and if the evaluation coefficient is between the upper threshold and the lower threshold, it is determined that the current data has certain distinguishing potential but is not stable, and is classified into the medium trustworthiness area. This division method can ensure that the subsequent model executes a differentiated processing path according to the trustworthiness level, thereby enhancing the robust response capability of the system to abnormal feature folding or fuzzy data in the industrial edge environment.

[0048] The division result is bound with the input data features and stored in the state buffer as the input basis for the adaptive scheduling strategy of the subsequent inference path, realizing a dynamic distinguishing process driven by trustworthiness.

[0049] The trustworthiness level division result is bound with the input data features and stored in the state buffer, which can be achieved by constructing a labeled data encapsulation structure. Specifically, after the input data completes the trustworthiness differentiation evaluation and completes the high, medium and low trustworthiness area division, a multi-field data object containing the original feature value, timestamp, sensor source and trustworthiness level label is generated; then the object is written into the local state buffer, which is a ring buffer structure managed in time sequence, supporting fast indexing and classification extraction of different trustworthiness level data. This binding process ensures that each input data retains its trustworthiness judgment context in the subsequent inference process, so that the inference scheduling strategy can dynamically select different model paths, parameter configurations or decision mechanisms according to the trustworthiness level of the data, thereby realizing accurate response to data quality fluctuations and enhancing distinguishing stability in industrial edge intelligent systems. This method avoids the misjudgment risk brought by the unified processing path, and improves the robustness and controllability of the overall system.

[0050] S4, determining an inference path according to the trustworthiness differentiation evaluation result, processing high trustworthiness area data by using a standard model inference path, processing medium trustworthiness area data by using a parameter self-adaptive adjustment path, and processing low trustworthiness area data by using an alternative inference path or a rejection processing mechanism. In this embodiment, S4 specifically includes: receiving a trustworthiness evaluation coefficient corresponding to the current input data, and selecting a corresponding reasoning path type in a preset reasoning path set as an inference processing flow entry of the current input data according to a trustworthiness interval label corresponding to the trustworthiness evaluation coefficient; This step can be implemented by constructing a label mapping table based on the trustworthiness evaluation coefficient and a path scheduling mapper. The specific implementation is as follows: first, a reasoning path set is preset in the edge artificial intelligence computing platform, which includes a standard model reasoning path, a parameter self-adaptive adjustment path, and a substitute reasoning path or a rejection processing mechanism, each path corresponding to a specific trustworthiness interval label. After the system receives the trustworthiness evaluation coefficient of the current input data, the evaluation coefficient is mapped to the label of the "high trustworthiness region", "medium trustworthiness region" or "low trustworthiness region" according to the preset trustworthiness interval threshold. Then, the trustworthiness interval label is mapped to the matching reasoning path type through the table lookup operation, and the reasoning path is taken as the inference processing flow entry of the current input data. The preset reasoning path set refers to a plurality of reasoning execution path sets defined and encapsulated according to different trustworthiness intervals when the edge artificial intelligence computing system is deployed, which are used to realize the differential processing of different trustworthiness level data and ensure efficient data processing and avoid misjudgment risk caused by uncertain data in the industrial field.

[0051] If the current input data is classified into the high trustworthiness region, the standard model reasoning path is called, and the analysis result is obtained by executing the standard model reasoning path without changing the edge artificial intelligence computing model structure and parameter configuration; If the current input data is classified into the high trustworthiness region, the standard model reasoning path can be called to realize the inference processing through the static binding method. The specific implementation is as follows: the system fixes the structure and parameter configuration of the standard model as a basic model version, and preloads the version in the edge computing device to construct the standard reasoning path entry. After receiving the input data marked as the high trustworthiness region, the standard reasoning path branch is triggered through the trustworthiness interval label, so that the data bypasses any model parameter or structure adjustment step and is directly input into the fixed structure and parameter edge artificial intelligence computing model to execute the inference operation. This method avoids the model uncertainty caused by dynamic adjustment and ensures the inference efficiency and judgment accuracy of high trustworthiness data under optimal conditions. The reason for adopting this method is that the high trustworthiness region data has sufficient discriminability, strong stability, and complete feature expression. Directly using the standard model reasoning can maximize the use of existing training results, ensure the stability and interpretability of the inference result in the high-reliability environment of the industrial field, and avoid errors caused by excessive intervention.

[0052] If the current input data is classified into the medium trust region, the parameter adaptive adjustment path is called, and based on the statistical characteristic information of the current input data, the parameter weight of the edge artificial intelligence calculation model is dynamically reset to form a parameter adaptive model for the characteristics of the current input data, and the analysis result is obtained by executing the parameter adaptive adjustment path; If the current input data is classified into the medium trust region, the parameter adaptive adjustment path can be used to implement the inference processing. The specific implementation is as follows: the system first extracts the statistical characteristic information of the current input data, including mean, variance, skewness and kurtosis, and constructs a feature-driven parameter adjustment vector based on the similarity distribution state in the historical data; then, according to the adjustment vector, the key weight parameters in the edge artificial intelligence calculation model are dynamically fine-tuned, such as adjusting the activation threshold, weight normalization coefficient or regularization strength of a specific layer, to form a parameter adaptive model that adapts to the current input feature structure; finally, the inference process is executed using the adjusted model. This method can adapt to the case where the input feature exists certain deviation or compression while keeping the model structure stable, thereby enhancing the discrimination ability of the model in the boundary fuzzy region. The reason for using this method is that the medium trust region data has certain discrimination ability, but the feature distribution is uncertain, and the standard model inference is prone to misjudgment, while the adaptive adjustment at the parameter level not only improves the robustness of the inference, but also avoids the information loss caused by completely replacing the path.

[0053] If the current input data is classified into the low trust region, according to the system configuration strategy, the alternative inference path or the rejection processing mechanism is selected to be called, wherein the alternative inference path calls an independently constructed redundant edge artificial intelligence calculation model for analysis and processing, and the rejection processing mechanism triggers an abnormal feedback process and interrupts the current inference process to prevent misjudgment caused by too low trust region division; If the current input data is classified as low-confidence region, the system can automatically select trigger alternative inference path or rejection processing mechanism by configuring decision logic to improve the robustness of the system under low discriminative input. The specific implementation is as follows: first, the system triggers the low-confidence region label based on the confidence degree evaluation coefficient and the preset threshold mapping table; then, by configuring the strategy to judge the demand of the current industrial task for misjudgment tolerance, if the precision reduction fault tolerance is allowed, the alternative inference path is activated, and the redundant edge artificial intelligence computing model independently constructed is called for analysis and processing. The redundant model uses a more powerful generalization architecture and is specifically used for processing data input with serious feature overlap; if the system strategy is for high-security strict control scene, the rejection processing mechanism is activated, the current inference process is interrupted, and the input data is marked as abnormal, and an abnormal feedback signal is output to the industrial operation and control platform, prompting the need for manual intervention or resampling. The reason for using this way is that low-confidence region data has serious feature distribution distortion, and it is easy to cause misjudgment by continuing to use the standard model or adaptive model. Through redundant inference or rejection mechanism, high-risk input can be effectively isolated to ensure the operation reliability and discriminative safety of edge artificial intelligence computing system in industrial field.

[0054] The inference path type is bound to the inference result and recorded in the inference execution log for subsequent inference process evaluation and path scheduling strategy optimization module for incremental learning and strategy updating to realize dynamic inference path adaptive control based on confidence degree evaluation results.

[0055] In order to realize dynamic inference path adaptive control based on confidence degree evaluation results, the following methods can be used: after completing the inference processing of each input data, the system binds the inference path type currently used to the result of this inference output one by one, and writes it into the inference execution log database in a structured format. The log content includes: input data feature summary, confidence degree evaluation coefficient, inference path type, inference output result and corresponding confidence index. Then, the path scheduling strategy optimization module in the system reads the log data periodically, and uses an adaptive optimization algorithm based on incremental learning to identify the dynamic mapping relationship between different confidence intervals and actual inference accuracy, and adjust the selection strategy of inference path under each confidence interval, such as fine-tuning the regularization parameter of the model used for medium-confidence region data or the response sensitivity of the alternative inference trigger mechanism. The purpose of this mechanism is to record and feedback the linkage effect between inference path and result continuously, so as to gradually optimize the response strategy of the system to the confidence information, so that the edge artificial intelligence computing process can adapt to data changes in complex industrial field environment, and realize more stable and accurate intelligent analysis capability.

[0056] S5, based on the result of the trusted interval division in the continuous time window, a trusted trend fluctuation sequence is constructed, and the truncation processing strategy is dynamically adjusted based on the same, including the setting mode of the truncation step, the offset mode of the truncation boundary and the configuration mode of the truncation mapping range.

[0057] In this embodiment, S5 specifically includes the following steps: A continuous time window sequence with a fixed length is constructed based on a time axis, and the result of the trusted interval division corresponding to the current input data is recorded in each time window, and is arranged in time sequence to form a trusted interval division track; The implementation of constructing a continuous time window sequence with a fixed length based on a time axis and recording the result of the trusted interval division of the current input data can divide the data stream into time windows with equal time span by using a sliding window mechanism, for example, constructing windows in units of seconds, minutes or other sampling periods. In each time window, the system receives the input data stream from the edge device, and according to the pre-generated trusted division evaluation coefficient, all or representative input samples in this time period are marked as high-trust area, medium-trust area or low-trust area. Subsequently, the final trusted interval division label formed in each window is bound with its time stamp, and is arranged in time sequence to form a trusted interval division track. The track serves as a basis for representing the stability of the model in discriminating input data over time, and provides continuity and time sequence basis for subsequent trend fluctuation analysis and dynamic regulation. In this way, the change trend of the discrimination state in the short term can be captured, the potential stability decline signal can be identified, and the system robustness and response accuracy can be improved by intervening and adjusting the processing strategy in advance.

[0058] The division results of each time window in the trusted interval division track are encoded and mapped to construct a discrete state sequence, and the sliding window statistical method is used to calculate the change trend of the occurrence frequency of high-trust area, medium-trust area and low-trust area in the sequence to generate a trusted trend fluctuation sequence; The encoding mapping of the division results of each time window in the trusted interval division trajectory and the construction of the trusted trend fluctuation sequence can be achieved by mapping the high trusted area, the medium trusted area and the low trusted area to discrete state values, for example, assigning values of 2, 1 and 0 respectively. The trusted labels of each time window are converted into state values according to the encoding rule, and are arranged in time sequence to form a discrete state sequence. Then, a sliding statistical window (such as a step of 3 or 5 windows) is used to slide on the state sequence to calculate the frequency of each type of state, thereby generating the change trend sequence of the high trusted area, the medium trusted area and the low trusted area in the time dimension. The frequency change trend is used to reveal the dynamic evolution behavior of the input data being divided into each trusted level by the model in the continuous period, thereby constructing the trusted trend fluctuation sequence. In the above manner, the volatility and phased trend of the system's trusted degree discrimination result on the input data can be captured, which provides data support for subsequent judgment of whether the current truncation strategy causes a decrease in discrimination stability, and enhances the system's perception ability of model degradation or data distortion problems.

[0059] According to the trend change characteristics in the trusted trend fluctuation sequence, the overall discrimination stability of the system to the input data in the current stage is determined. When the frequency of the low trusted area continuously increases or the fluctuation is enhanced, it is judged that the current truncation processing strategy has a weakening effect on the feature space expression. To evaluate the discrimination stability of the system in the current stage, the change trend characteristics of each trusted level (high trusted area, medium trusted area and low trusted area) in the trusted trend fluctuation sequence can be extracted, including the frequency rise rate, fluctuation amplitude and trend duration. Specifically, the change slope of the low trusted area frequency in the continuous time window can be calculated by the difference analysis method, the fluctuation intensity can be evaluated by combining the sliding standard deviation index, and the threshold rule can be used to identify the continuously rising or enhanced trend. When the frequency of the low trusted area continuously increases in multiple consecutive time windows, or the fluctuation amplitude exceeds a set proportion of the historical mean, it can be judged that the trusted discrimination ability of the system to the input data is decreasing in the current stage. This trend usually reflects that the truncation strategy leads to feature compression or expression distortion under the current sampling data structure, that is, the discriminability of the feature space is weakened. Therefore, the adaptability and effectiveness of the current truncation processing strategy can be determined by the above method, which provides a reliable criterion for subsequent dynamic strategy adjustment and avoids the accumulation of model misjudgment, which leads to false alarms or control risks in the industrial field.

[0060] When it is judged that the current truncation processing strategy has a weakening effect, a preset parameter optimization module is called to dynamically adjust the truncation step setting method in the truncation processing strategy, and the numerical retention precision is improved by reducing the truncation step; After judging that the current truncation processing strategy exists the effect of weakening feature expression, a preset parameter optimization module can be called to dynamically adjust the truncation processing strategy. The parameter optimization module is a strategy selection and parameter reconfiguration system constructed based on prior working condition knowledge and historical performance data, and has a rule mapping table and adaptive adjustment logic for truncation parameter adjustment integrated inside. In specific implementation, the module receives the stability index of the current trusted trend fluctuation sequence feedback, and dynamically reduces the truncation step size setting value in combination with the trade-off constraints of numerical accuracy and calculation efficiency in industrial scenes, thereby improving the numerical resolution of the retained data and slowing down the information loss caused by feature space compression. The process can realize fine update of the truncation step size through an autoregressive control function or a strategy matching method based on a state transition graph. In this way, the accuracy level can be managed in detail without changing the overall data truncation framework, and the discriminant stability and robustness of the edge artificial intelligence model in the complex data environment of the industrial field can be improved.

[0061] The offset mode of the truncation boundary is adjusted synchronously, the coincidence trend of the truncation boundary and the high-density data area is broken by introducing an offset disturbance, and the truncation mapping range is reconfigured to make the data mapping uniform, thereby relieving the boundary aggregation and distribution folding problem.

[0062] To further relieve the boundary aggregation and feature distribution folding problem caused by truncation, the offset mode of the truncation boundary can be adjusted synchronously on the basis of adjusting the truncation step size. In specific implementation, based on the boundary area abnormal density reflected by the current trusted trend fluctuation sequence, a small-amplitude truncation boundary offset disturbance is introduced to adjust the starting point of the truncation boundary in a non-fixed position, thereby breaking the structural relationship of the continuous overlap of the original truncation boundary and the high-density data area. The disturbance can adopt a Gaussian noise disturbance model or an offset strategy driven by the minimum overlap interval in a sliding window, so that the truncation starting point dynamically changes in multiple time windows. At the same time, the truncation mapping range needs to be reconfigured to adjust the boundary coverage relationship and distribution density mapping relationship of the data mapping interval, so that the feature values after truncation are more uniformly distributed in space, and the input ambiguity caused by the data accumulation effect is reduced. The combined strategy realizes flexible shaping of the feature distribution structure by local disturbance and range reconfiguration without changing the overall numerical truncation framework, effectively improving the recognition accuracy and reliability of the edge artificial intelligence model for industrial complex working conditions.

[0063] The above-described embodiments can be implemented in part or in whole through software, hardware, firmware or any combination thereof. When implemented in software, the above-described embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When loaded and executed by a computer, the computer instructions or computer programs cause the computer to perform all or part of the processes or functions described in the embodiments of the present application. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable apparatus. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center through a wired or wireless (for example, infrared, wireless, microwave, etc.) manner. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. containing one or more available medium collections. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state disk.

[0064] It should be understood that the size of the sequence number of each process described above in various embodiments of the present application does not mean the order of execution, and the execution order of each process should be determined by its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0065] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be realized in electronic hardware, or a combination of computer software and electronic hardware. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0066] In several embodiments provided in the present application, it should be understood that the disclosed system and method can be implemented in other ways. For example, the above-described embodiments are only illustrative, for example, the division of units is only a logical function division, and actual implementation can have another division manner, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed units can be indirect coupling or communication connection through some interfaces, devices or units, and can be electrical, mechanical or other forms.

[0067] The units described as separate components may or may not be physically separate, and the components displayed as units may or may not be physical units, that is, may be located in one place, or may be distributed to multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0068] In addition, each functional unit in each embodiment of the present application can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit.

[0069] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An edge artificial intelligence computing method for big data analysis, characterized by, Specifically comprising the following steps: S1, constructing a numerical distribution atlas for continuously collected industrial analog quantity data, identifying boundary aggregation caused by truncation behavior by analyzing the concentration of data in the boundary interval formed by setting a symmetric expansion range on the numerical axis with the truncated boundary corresponding value as the center, the density gradient, and the time series trend; S2, after the boundary aggregation is identified, constructing a feature space overlap rate model based on the numerical mapping overlap degree before and after truncation, the sample clustering overlap rate of the boundary interval and its adjacent sections above and below, and the data distribution compression ratio, to identify whether there is a distribution folding trend; S3, based on the identification results of boundary aggregation and distribution folding trend, constructing a credible discrimination degree evaluation function, generating a credible discrimination degree evaluation coefficient, and dividing the current input data into high credible area, medium credible area or low credible area; S4, determining the reasoning path according to the credible discrimination degree evaluation result, processing high credible area data with standard model reasoning path, processing medium credible area data with parameter self-adaptive adjustment path, and processing low credible area data with alternative reasoning path or rejection processing mechanism; S5, based on the credible interval division result in the continuous time window, constructing a credible trend fluctuation sequence, and dynamically adjusting the truncation processing strategy according to it, including the setting method of truncation step, the offset method of truncation boundary and the configuration method of truncation mapping range. 2.The edge artificial intelligence computing method for big data analytics of claim 1, wherein, S1 specifically comprises the following steps: Based on the time axis, a sliding sampling window with fixed width is constructed, the continuously collected industrial analog quantity data in each window is divided into continuous numerical sections, and a numerical distribution atlas is generated in each window with numerical section as horizontal axis and sample frequency as vertical axis; In each numerical distribution atlas, the boundary interval formed by symmetrically expanding in the upward and downward directions according to the set numerical span with the truncated boundary corresponding value as the center is determined, the number of samples in the boundary interval is counted, and the frequency change value in the current window and historical multiple windows is calculated to obtain the data concentration in the boundary interval; The sample frequency in the boundary interval and its adjacent equal-width sections above and below is compared, the density gradient difference value is calculated, and the change sequence of the difference value is extracted on the continuous multiple windows to construct the time series trend of the density gradient; When the data concentration in the boundary interval exceeds the sample mean multiple threshold, the density gradient difference value is higher than the preset critical value, and the time series trend meets the stable or enhanced change condition, it is identified that the boundary aggregation caused by truncation behavior exists in the boundary interval. 3.The edge artificial intelligence computing method for big data analytics of claim 1, wherein, S2 specifically comprises the following steps: After the boundary aggregation is identified, the historical floating point original data in the boundary interval and the corresponding fixed point data after truncation are extracted, the numerical mapping relationship between the original data and the truncated data is calculated, the coincidence rate of the numerical landing points before and after truncation is calculated as the numerical mapping overlap degree; Based on the sample set in the boundary interval and its adjacent equal-width sections in the upward and downward directions, unsupervised clustering operation is performed in the original feature space, the cross proportion of the samples in the boundary interval and the adjacent section samples in the clustering label is calculated as the sample clustering overlap rate; The sample data of the boundary interval and its adjacent sections before and after the truncation processing are extracted as the original physical quantity parameter sequences constituting the feature space as the feature vector set, and the mean value change rate of the Euclidean distance between the samples in each set is calculated as the data distribution compression ratio for measuring the space contraction effect; The numerical mapping coincidence degree, sample clustering overlap rate and data distribution compression ratio are normalized and input into a weighted fusion function to construct a feature space overlap rate model and output a comprehensive overlap score value; The comprehensive overlap score value is compared with a preset distribution folding recognition threshold value, and when the score value is higher than the threshold value, it is judged that the current boundary interval has significant overlap in the feature space, and it is recognized as having a distribution folding trend.

4. The edge artificial intelligence computing method for big data analytics of claim 1, wherein, S3 specifically includes the following steps: Extract the data concentration, density gradient change trend in the boundary aggregation condition recognition result, and the numerical mapping coincidence degree, sample clustering overlap rate and data distribution compression ratio in the distribution folding trend recognition result as a set of trusted distinguishability evaluation factors; Each index in the set of trusted distinguishability evaluation factors is standardized to have a zero mean and a unit variance to obtain a standardized factor sequence with a unified numerical scale; A linear weighted fusion method is applied to combine the standardized factor sequence according to the preset weight to construct a trusted distinguishability evaluation function, and output a single-valued trusted distinguishability evaluation coefficient for quantifying the discriminability of the current input data in the feature space; According to the numerical interval of the trusted distinguishability evaluation coefficient, the current input data is divided into a high-trust area, a medium-trust area or a low-trust area to drive the hierarchical processing flow of the subsequent reasoning path.

5. The edge artificial intelligence computing method for big data analytics of claim 4, wherein, The specific way to construct the trusted distinguishability evaluation function includes: The data concentration, density gradient change trend, numerical mapping coincidence degree, sample clustering overlap rate and data distribution compression ratio are subjected to mean normalization to construct a set of standardized factors with a unified numerical scale; The importance between each standardized factor is weighted based on the historical working condition classification accuracy in the industrial field, and an iterative optimization algorithm is used to determine the fusion weight of each standardized factor; The standardized factors and their corresponding weights are combined using a linear weighted fusion method to construct a trusted distinguishability evaluation function with clear expression structure and calculation path, and the trusted distinguishability evaluation coefficient is output through the function to reflect the discriminability of the current input data in the feature space.

6. The edge artificial intelligence computing method for big data analytics of claim 5, wherein, According to the numerical interval of the trusted distinguishability evaluation coefficient, the current input data is divided into a high-trust area, a medium-trust area or a low-trust area, specifically: The upper threshold and lower threshold of the trusted distinguishability evaluation coefficient are set, and the threshold boundary is determined by the reasoning accuracy change trend of the training data set under different trusted distinguishability evaluation coefficients; When the trusted distinguishability evaluation coefficient is higher than the upper threshold, the current input data is divided into a high-trust area; When the trusted distinguishability evaluation coefficient is lower than the lower threshold, it is divided into a low-trust area; When the trusted distinguishability evaluation coefficient is between the upper threshold and the lower threshold, it is divided into a medium-trust area; The division result is bound with the input data features and stored in a state buffer as an input basis for subsequent reasoning path adaptive scheduling strategies.

7. The edge artificial intelligence computing method for big data analytics of claim 1, wherein, S4 specifically includes: Receiving a trusted degree evaluation coefficient corresponding to the current input data, and selecting a corresponding reasoning path type in a preset reasoning path set as an entry of a reasoning processing flow of the current input data according to a trusted interval label corresponding to the trusted degree evaluation coefficient; If the current input data is divided into a high-trust region, a standard model reasoning path is called, and the analysis result is obtained by executing the standard model reasoning path under the condition that the structure and parameter configuration of the edge artificial intelligence calculation model remain unchanged; If the current input data is divided into a medium-trust region, a parameter adaptive adjustment path is called, and the parameter weight of the edge artificial intelligence calculation model is dynamically reset based on the statistical feature information of the current input data to form a parameter adaptive model for the characteristics of the current input data, and the analysis result is obtained by executing the parameter adaptive adjustment path; If the current input data is divided into a low-trust region, a substitute reasoning path or a rejection processing mechanism is selected according to the system configuration strategy, wherein the substitute reasoning path calls an independently constructed redundant edge artificial intelligence calculation model for analysis and processing, and the rejection processing mechanism triggers an abnormal feedback process and interrupts the current reasoning process to prevent misjudgment caused by too low trusted degree; The reasoning path type and the reasoning result are bound and recorded in a reasoning execution log for subsequent reasoning process evaluation and path scheduling strategy optimization module for incremental learning and strategy updating.

8. The edge artificial intelligence computing method for big data analytics of claim 1, wherein, S5 specifically includes the following steps: Based on a time axis, a fixed-length continuous time window sequence is constructed, the trusted interval division result corresponding to the current input data is recorded in each time window, and a trusted interval division trajectory is formed in chronological order; The division results of each time window in the trusted interval division trajectory are encoded and mapped to construct a discrete state sequence, and the occurrence frequency change trend of the high-trust region, the medium-trust region and the low-trust region in the sequence is calculated using a sliding window statistical method to generate a trusted trend fluctuation sequence; According to the trend change characteristics in the trusted trend fluctuation sequence, the overall discriminant stability of the system to the input data in the current stage is determined, and when the frequency of the low-trust region continuously increases or fluctuates intensively, it is judged that the current truncation processing strategy has a weakening effect on the feature space expression; When it is judged that the current truncation processing strategy has a weakening effect, a preset parameter optimization module is called to dynamically adjust the truncation step setting method in the truncation processing strategy, and the numerical retention precision is improved by reducing the truncation step; The offset mode of the truncation boundary is adjusted synchronously, the coincidence trend of the truncation boundary and the high-density data region is broken by introducing an offset disturbance, and the mapping range of the truncation is reconfigured to make the data mapping uniform, thereby alleviating the problems of boundary aggregation and distribution folding.

Images