Data lifecycle management method and system based on multi-dimensional AI evaluation model

Abstract

This application relates to the field of artificial intelligence technology, and in particular to a method and system for full lifecycle management based on a multi-dimensional AI evaluation model. The method includes: collecting industrial data from intelligent manufacturing equipment and preprocessing the industrial data; inputting the preprocessed industrial data into a multi-dimensional AI evaluation model to obtain multi-dimensional scores, and fusing the multi-dimensional scores to obtain a comprehensive data trust score; implementing flow control of industrial data throughout its entire lifecycle based on the comprehensive data trust score; collecting feedback indicators of the implementation effect of industrial data at each lifecycle stage, and adaptively iteratively updating the multi-dimensional AI evaluation model based on these feedback indicators. This application addresses the issues of single-dimensional industrial data evaluation, lack of value-driven approaches in full lifecycle management, and the lack of dynamic adaptive capabilities in the model.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method and system for full lifecycle management based on a multi-dimensional AI evaluation model. Background Technology

[0002] With the deep implementation of Industry 4.0 and smart manufacturing technologies, the total amount of industrial data generated by smart manufacturing enterprises during the production process is growing exponentially. This industrial data covers the operating parameters, process indicators, quality inspection results, and supply chain information of various smart manufacturing equipment such as CNC machine tools, industrial robots, vision inspection, and environmental sensors. It is characterized by multi-source heterogeneity, different sampling frequencies, dispersed time series, and uneven value density of individual industrial data. Massive amounts of industrial data are the core foundation for production line process optimization, fault prediction, and production decision-making. Therefore, refined management and control of industrial data throughout its entire lifecycle has become the core key to the implementation of smart manufacturing. However, traditional industrial data management solutions are difficult to adapt to the complex application scenarios of smart manufacturing, have multi-dimensional technical shortcomings, and are unable to achieve intelligent, dynamic, and value-oriented data management and control.

[0003] Traditional industrial data quality assessment systems are one-sided in their dimensions and poorly adaptable to different scenarios. They only evaluate data based on basic and general dimensions such as completeness and accuracy, failing to build multi-dimensional and scenario-based assessment mechanisms that take into account the specific characteristics of intelligent manufacturing industrial scenarios. Especially in core production scenarios such as CNC machine tool condition monitoring, the real-time nature and temporal consistency of data directly determine the accuracy of production status assessment and process control. Their importance far exceeds the absolute accuracy of the data. Traditional, singular assessment models cannot effectively identify the actual usability of industrial data, which can easily lead to low-quality data flowing into the production analysis process and affecting the accuracy of production decisions.

[0004] Meanwhile, the existing data management system suffers from a disconnect between data lifecycle management and value assessment, resulting in low overall data resource utilization. The industry has yet to establish a value-driven full lifecycle management mechanism, with data collection, storage, archiving, and maintenance processes being decoupled from data value assessment. This leads to a severe imbalance in storage resource allocation, with a large amount of low-value, redundant, and invalid data accumulating in the hot storage area, occupying core storage and access resources. Meanwhile, high-value data archived to the cold storage area lacks an active mining mechanism, making efficient retrieval and reuse impossible. This results in idle high-quality data resources, serious resource waste, and a significant reduction in the overall utilization efficiency of industrial data.

[0005] In addition, traditional evaluation models have a static and fixed structure and lack dynamic adaptability. Industrial production processes have dynamic uncertainties. The continuous adjustment of equipment operating status, production cycle, and process parameters will lead to real-time dynamic changes in the distribution characteristics and data patterns of industrial data. Static evaluation models with fixed parameters cannot adapt to the dynamic changes in data distribution, cannot update evaluation rules and evaluation logic in sync, and will continue to have problems such as evaluation deviation and inaccurate judgment. They are difficult to adapt to the dynamic and high-precision data evaluation and management needs of intelligent manufacturing. Summary of the Invention

[0006] This application provides a method and system for full lifecycle management based on a multi-dimensional AI evaluation model, in order to solve the problems of single evaluation dimensions of industrial data, lack of value-driven full lifecycle management, and lack of dynamic adaptive capabilities of the model.

[0007] To solve the above-mentioned technical problems, this application provides the following technical solution: A full lifecycle management method based on a multi-dimensional AI evaluation model includes the following steps: Step S110: Collect industrial data from intelligent manufacturing equipment and preprocess the industrial data; Step S120: Input the preprocessed industrial data into the multi-dimensional AI evaluation model to obtain multi-dimensional scores, and fuse the multi-dimensional scores to obtain a comprehensive data trust score; Step S130: Based on the comprehensive data trust score, implement flow control of industrial data throughout its entire lifecycle; Step S140: Collect feedback indicators of the implementation effect of industrial data at each lifecycle stage, and adaptively iterate and update the multi-dimensional AI evaluation model based on the implementation effect feedback indicators.

[0008] The above-described full lifecycle management method based on a multi-dimensional AI evaluation model preferably involves performing missing value completion, outlier identification and removal, and time-series resampling and alignment processing on industrial data according to a predetermined sequential process to generate time-series aligned data. The time-series aligned data is then subjected to adaptive normalization processing to generate dimensionless data, thus completing basic preprocessing. Based on the dimensionless data, time-frequency joint feature extraction is performed to construct data feature vectors, thus completing advanced preprocessing.

[0009] The above-described full lifecycle management method based on a multi-dimensional AI evaluation model preferably involves: comparing the data mean, standard deviation, and root mean square value in the data feature vector with the benchmark time-domain statistical features to obtain a time-domain feature value score; measuring the distance between the spectral amplitude of each frequency component in the data feature vector and the benchmark spectral amplitude to obtain a frequency-domain feature value score; fusing the time-frequency value scores using an adaptive weighting method; and then outputting a value trend score through linear mapping and Sigmoid nonlinear activation. The time-series aligned data is initially scored in three sub-dimensions: completeness, accuracy, and consistency. These initial scores are then fused using a linear weighting method to obtain a comprehensive quality score. Each dimensionless data point is individually verified for a safe value range, and the deviation of each dimensionless data point from the safe value range is calculated. The average deviation of all dimensionless data points is then calculated to obtain a risk score. The generation timestamp of each time-series aligned data point is extracted, and a timeliness score is calculated based on the timeliness decay law. Finally, the scores from each dimension are fused using an adaptive weighting method to obtain a comprehensive data trust score.

[0010] The above-described full lifecycle management method based on a multi-dimensional AI evaluation model preferably incorporates a comprehensive trust score based on data. When the upper limit of the risk safety range is reached and the data value trend is upward, industrial data is maintained or migrated to the hot data stage and stored in the high-performance storage area of ​​SSDs; when the lower limit of the risk safety range is reached... Data-driven trust scoring When the upper limit of the risk security range or the trend of data value change tends to stabilize, industrial data is migrated to the warm data stage and stored in the medium-performance storage area of ​​SSDs; when the overall data trust score... When the risk safety range lower limit threshold is reached, and the data value trend continues to decline, and the risk score is below the risk threshold, industrial data is migrated to the cold data stage and archived to the low-performance storage area of ​​SSDs; when the overall data trust score... When the risk score exceeds the lower limit of the risk safety range and the risk score exceeds the risk threshold, the data eviction process is triggered, and the corresponding industrial data is deleted from the SSD storage area.

[0011] The above-described full lifecycle management method based on a multi-dimensional AI evaluation model preferably involves: collecting feedback indicators of the implementation effect of industrial data at each stage of the lifecycle; normalizing and weighting these indicators with a time-series forgetting decay coefficient to obtain a comprehensive cost loss value; performing gradient optimization on each component based on the comprehensive cost loss value, calculating the gradient of the comprehensive cost loss value with respect to the adaptive weight coefficients, calculating the correction step size of each adaptive weight coefficient by combining the learning rate and the comprehensive cost loss value, obtaining the pre-updated value of the adaptive weight coefficient by subtracting the correction step size from the original adaptive weight coefficient, locking the pre-updated value of the adaptive weight coefficient within the 0-1 range through amplitude constraints, and obtaining the updated adaptive weight coefficient; replacing the original adaptive weight coefficient in the multi-dimensional AI evaluation model with the updated adaptive weight coefficient to achieve adaptive iterative updates of the multi-dimensional AI evaluation model.

[0012] A full lifecycle management system based on a multi-dimensional AI evaluation model includes: a data acquisition and preprocessing module, a multi-dimensional AI evaluation model, a data management and control module, and a model iteration and update module. The data acquisition and preprocessing module collects industrial data from intelligent manufacturing equipment and preprocesses the industrial data. The multi-dimensional AI evaluation model scores the preprocessed industrial data from multiple dimensions and merges the scores to obtain a comprehensive data trust score. The data management and control module manages the flow of industrial data throughout its entire lifecycle based on the comprehensive data trust score. The model iteration and update module collects feedback indicators of the implementation effect of industrial data at each lifecycle stage and adaptively iterates and updates the multi-dimensional AI evaluation model based on these feedback indicators.

[0013] In the aforementioned full lifecycle management system based on a multi-dimensional AI evaluation model, preferably, the industrial data is processed step by step according to a predetermined sequential process, including missing value completion, outlier identification and removal, and time-series resampling and alignment, to generate time-series aligned data. The time-series aligned data is then subjected to adaptive normalization to generate dimensionless data, thus completing basic preprocessing. Based on the dimensionless data, time-frequency joint feature extraction is performed to construct data feature vectors, thus completing advanced preprocessing.

[0014] The lifecycle management system based on the multi-dimensional AI evaluation model described above preferably includes the following modules: a value dimension evaluation module, a quality dimension evaluation module, a risk dimension evaluation module, a timeliness dimension evaluation module, and a score fusion module. Specifically, the value dimension evaluation module compares the data mean, standard deviation, and root mean square value in the data feature vector with the benchmark time-domain statistical features to obtain a time-domain feature value score. It also measures the distance between the spectral amplitudes of each frequency component in the data feature vector and the benchmark spectral amplitude to obtain a frequency-domain feature value score. The time-frequency value scores are fused using an adaptive weighting method, and then, after linear mapping and Sigmoid nonlinear activation, the value trend is output. The scoring system is as follows: The quality dimension assessment module assigns raw scores to the time-series aligned data in three sub-dimensions: completeness, accuracy, and consistency. These raw scores are then fused using a linear weighting method to obtain a comprehensive quality score. The risk dimension assessment module verifies the safe value range for each dimensionless data point, calculates the deviation of each dimensionless data point from the safe value range, and averages the deviations of all dimensionless data points to obtain a risk score. The timeliness dimension assessment module extracts the generation timestamp of each time-series aligned data point and calculates the timeliness score based on the timeliness decay law. The score fusion module uses an adaptive weighted fusion method to fuse the scores from each dimension to obtain a comprehensive data trust score.

[0015] In the aforementioned full lifecycle management system based on a multi-dimensional AI evaluation model, preferably, when the data is comprehensively trusted... When the upper limit of the risk safety range is reached and the data value trend is upward, industrial data is maintained or migrated to the hot data stage and stored in the high-performance storage area of ​​SSDs; when the lower limit of the risk safety range is reached... Data-driven trust scoring When the upper limit of the risk security range or the trend of data value change tends to stabilize, industrial data is migrated to the warm data stage and stored in the medium-performance storage area of ​​SSDs; when the overall data trust score... When the risk safety range lower limit threshold is reached, and the data value trend continues to decline, and the risk score is below the risk threshold, industrial data is migrated to the cold data stage and archived to the low-performance storage area of ​​SSDs; when the overall data trust score... When the risk score exceeds the lower limit of the risk safety range and the risk score exceeds the risk threshold, the data eviction process is triggered, and the corresponding industrial data is deleted from the SSD storage area.

[0016] The aforementioned lifecycle management system based on a multi-dimensional AI evaluation model preferably involves: collecting feedback indicators of the implementation effect of industrial data at each stage of the lifecycle; normalizing and weighting these indicators with a time-series forgetting decay coefficient to obtain a comprehensive cost loss value; performing gradient optimization on each component based on the comprehensive cost loss value, calculating the gradient of the comprehensive cost loss value with respect to the adaptive weight coefficients, calculating the correction step size of each adaptive weight coefficient by combining the learning rate and the comprehensive cost loss value, obtaining the pre-updated value of the adaptive weight coefficient by subtracting the correction step size from the original adaptive weight coefficient, locking the pre-updated value of the adaptive weight coefficient within the 0-1 range through amplitude constraints, and obtaining the updated adaptive weight coefficient; replacing the original adaptive weight coefficient in the multi-dimensional AI evaluation model with the updated adaptive weight coefficient to achieve adaptive iterative updates of the multi-dimensional AI evaluation model.

[0017] Beneficial effects: 1. By adopting a scheme of global clock timing registration, adding positive disturbance factor adaptive normalization and time-frequency joint feature extraction, the preprocessing link of multi-source heterogeneous industrial data is improved, effectively eliminating data processing defects caused by different sampling frequencies and numerical steady state anomalies, unifying data specifications and improving feature representation capabilities.

[0018] 2. Establish a four-dimensional AI joint evaluation framework encompassing value, quality, risk, and timeliness. This framework breaks away from the traditional, one-sided evaluation model that relies solely on basic indicators. It combines production line operating conditions to dynamically calculate the scores of each item and adaptively weight and integrate them to accurately quantify the overall credibility level of industrial data.

[0019] 3. Based on the comprehensive trust score of data, link value trends and risk indicators, divide the life cycle into four categories: hot, warm, cold, and obsolete, and match them with tiered storage strategies to optimize storage resource scheduling and reduce resource consumption caused by invalid data crowding out high-performance storage.

[0020] 4. Aggregate feedback indicators such as storage cost, data reuse rate, and operational anomalies, construct a comprehensive cost loss function by combining it with the time-series forgetting factor, and complete weight self-tuning with gradient optimization to achieve closed-loop iterative optimization of the evaluation model following production conditions.

[0021] 5. By connecting the entire business chain of data collection, preprocessing, intelligent evaluation, hierarchical storage, and model iteration, a full lifecycle automated management and control system is implemented, significantly improving the utilization rate of industrial data assets and laying a solid high-quality data foundation for intelligent manufacturing process improvement and intelligent production decision-making. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0023] Figure 1 This is a flowchart of the full lifecycle management method based on a multi-dimensional AI evaluation model provided in this application; Figure 2 This is a schematic diagram of the full lifecycle management system based on a multi-dimensional AI evaluation model provided in this application. Detailed Implementation

[0024] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0025] Example 1 like Figure 1 As shown, this application provides a full lifecycle management method based on a multi-dimensional AI evaluation model, including the following steps: Step S110: Collect industrial data from intelligent manufacturing equipment and preprocess the industrial data; Taking an intelligent manufacturing plant for automotive parts as an example, this application scenario is illustrated. The plant is equipped with intelligent manufacturing equipment such as CNC machine tools, industrial robots, AGV transport vehicles, and vision inspection equipment. Lightweight data acquisition probes are deployed on the PLC controller and edge gateway of each intelligent manufacturing device. These probes collect multi-source heterogeneous industrial data in real time, including equipment operation data, production process data, quality inspection data, and environmental monitoring data, at a predetermined sampling frequency.

[0026] As an example, for CNC machine tools: spindle load, feed rate, tool wear status, and vibration spectrum data are collected at a sampling frequency of 1kHz; for industrial robots: joint angle, torque feedback, and position error data are collected at a sampling frequency of 500Hz; for vision inspection equipment: product images, defect recognition results, and inspection confidence data are collected through triggered acquisition; for environmental sensors: temperature, humidity, and cleanliness data are collected at a sampling frequency of 1Hz.

[0027] The collected industrial data undergoes sequential processing according to a predetermined sequence, including missing value completion, outlier identification and removal, and time-series resampling and alignment, generating time-series aligned data. This time-series aligned data is then adaptively normalized to generate dimensionless data, completing basic preprocessing. Next, time-frequency joint feature extraction is performed on this dimensionless data to construct a data feature vector, completing advanced preprocessing. The time-series aligned data and dimensionless data generated during basic preprocessing, along with the data feature vector constructed during advanced preprocessing, are output together for use in subsequent multi-dimensional AI evaluation models. By limiting the preprocessing to a sequential, step-by-step process—forcing the output of the previous step as the input of the next—the processing chain of industrial data becomes controllable and traceable, adaptable to the management of the entire lifecycle of industrial data.

[0028] Since multi-source heterogeneous industrial data is collected at different sampling frequencies, a standard timestamp granularity grid can be constructed using a global industrial clock as a unified time reference. For example, the grid time step of the standard timestamp granularity grid is 100ms / 1s. All industrial data originally collected at different sampling frequencies such as 1kHz, 500Hz, 1Hz, and triggered sampling are interpolated and mapped onto the standard timestamp granularity grid to generate time-aligned data. This completes the time-series resampling alignment of multi-source heterogeneous industrial data, achieving time-series synchronization of multi-source heterogeneous industrial data and avoiding the pain points of time-series misalignment and inability to jointly model industrial data. Among these, the global industrial clock is the reference clock source for unified time synchronization of all acquisition devices, edge gateways, and controllers within the intelligent manufacturing production line, serving as the sole time reference for time-series alignment of multi-source heterogeneous industrial data.

[0029] Furthermore, the expression for timing resampling alignment is as follows: ; in, In the first Temporal alignment data generated at each target difference time point; For the first The original sampling time points are not uniform across different intelligent manufacturing devices; For the first One original sampling time point; For the first time stamp granularity grid One target interpolation time; In the first Industrial data collected at each original sampling time point; In the first Industrial data collected at each of the original sampling time points.

[0030] In steady-state mass production industrial data, the maximum and minimum values ​​of industrial data are often equal, which can easily lead to failure due to the denominator being zero during normalization. Therefore, a minimal positive perturbation factor is introduced to uniformly map industrial data of different dimensions and amplitudes to the [0,1] interval, thereby achieving adaptive normalization processing of industrial data. By introducing a minimal positive perturbation factor as a protection factor for the denominator during normalization, the problem of failure due to the denominator being zero during normalization is solved, thus enabling the full lifecycle management of industrial data to adapt to steady-state mass production scenarios.

[0031] Furthermore, the expression for adaptive normalization is as follows: ; in, The data are dimensionless after normalization; For input timing-aligned data; This represents the maximum value in this batch of industrial datasets; This is the minimum value in this batch of industrial datasets; It is a minimal positive perturbation factor. For example, the value is 0.0001.

[0032] The operating status of intelligent manufacturing equipment simultaneously contains both time-domain steady-state variation patterns and frequency-domain vibration harmonic characteristics. A single domain feature cannot fully characterize the operating conditions of intelligent manufacturing equipment and the hidden value of industrial data. Therefore, time-domain statistical features and frequency-domain spectral features can be extracted simultaneously based on dimensionless data. Then, the time-domain statistical features and frequency-domain spectral features are concatenated to generate a data feature vector, thereby providing regular, synchronous, and highly representative input data for subsequent multi-dimensional AI evaluation models. This ensures the continuity of industrial data processing and avoids information gaps.

[0033] Furthermore, the expression for the time-domain statistical characteristics is as follows: ; in, This represents the mean of the data within the sliding window. The standard deviation of the data within the sliding window; The root mean square value of the data within the sliding window; This represents the total number of data samples within the current sliding window. For the first Dimensionless data.

[0034] Furthermore, the expression for the frequency domain spectral characteristics is as follows: ; in, For the first Spectral amplitude of first-order frequency components; This refers to the sequence number of the data sample; For the first A dimensionless data point; This represents the total number of data samples within the current sliding window. These are the frequency domain harmonic component indices, corresponding to different sampling frequencies; It is the imaginary unit.

[0035] Furthermore, the expression for generating the data feature vector by concatenating time-domain statistical features and frequency-domain spectral features is as follows: ; in, The resulting data feature vector is used as input to a multi-dimensional AI evaluation model. The amplitude of the first-order frequency component spectrum. The amplitude of the second-order frequency component spectrum. For the first Spectral amplitude of first-order frequency components; This represents the number of frequency component spectral amplitudes.

[0036] By relying on the global industrial clock for time-series resampling alignment, adaptive normalization with positive disturbance factor protection, and a three-level preprocessing logic of time-frequency joint feature extraction, this approach aligns with the constraints of complex intelligent manufacturing conditions and is suitable for processing multi-source heterogeneous industrial data under differentiated sampling frequencies. It fundamentally avoids defects such as time sequence misalignment, anomaly rejection, inconsistent dimensions, and one-sided feature representation. This provides a standardized input for subsequent multi-dimensional AI evaluation models to assess quality, value, risk, and timeliness, effectively ensuring the accuracy and reliability of all-dimensional scoring calculations.

[0037] Step S120: Input the preprocessed industrial data into the multi-dimensional AI evaluation model to obtain multi-dimensional scores, and fuse the multi-dimensional scores to obtain a comprehensive data trust score. A multi-dimensional AI evaluation model is constructed and trained using historical industrial data. This model includes modules for value evaluation, quality evaluation, risk evaluation, timeliness evaluation, and a score fusion module. The preprocessed industrial data from step S110 is input into the trained multi-dimensional AI evaluation model, and evaluations are performed through the different modules to obtain a multi-dimensional score.

[0038] The value dimension assessment module employs a nonlinear value assessment model, using the data feature vector output in step S110. As input, the data feature vector The mean of the data in Data standard deviation Root mean square value of data By comparing the time-domain statistical characteristics with those of intelligent manufacturing equipment under normal steady-state operating conditions, a time-domain feature value score is obtained, and then the data feature vector is... The distance between the spectral amplitude of each frequency component in the data and the reference spectral amplitude under normal steady-state operation of the intelligent manufacturing equipment is measured to obtain the frequency domain feature value score. Then, the time-frequency value score is fused by the adaptive weighting method, and then the value trend score is output by linear mapping and Sigmoid nonlinear activation.

[0039] Furthermore, the formula for calculating the time-domain feature value is as follows: ; in, Score the value of time-domain features; This represents the average baseline data for normal steady-state operation of intelligent manufacturing equipment. This serves as the standard deviation of the baseline data for the normal steady-state operating conditions of intelligent manufacturing equipment. The root mean square value is the baseline data for the normal steady-state operating conditions of intelligent manufacturing equipment. and The baseline time-domain statistical characteristics of intelligent manufacturing equipment under normal steady-state operating conditions; Weighted by the data mean, Weights are the standard deviations of the data. The weights are the root mean square values ​​of the data. .

[0040] Furthermore, the formula for calculating the frequency domain feature value score is as follows: ; in, Score the value of frequency domain features; For the first Spectral amplitude of first-order frequency components; For the first The first-order reference spectrum amplitude; This represents the Euclidean distance between the current spectral amplitude and the reference vector.

[0041] Furthermore, the expression for the fusion of time-frequency value scores is as follows: ; in, This is the base value score after the fusion of time-frequency value scores; To enable adaptive fusion weights, adjustments can be made dynamically based on the type of intelligent manufacturing equipment and production conditions.

[0042] Furthermore, the expression for the value trend score is as follows: ; in, Score the value trend. ; These are the output weight coefficients of the linear mapping; It is a sigmoid nonlinear activation function that compresses the values ​​to [0,1].

[0043] The quality dimension assessment module employs a multi-dimensional adaptive weighted quality fusion model. Using the time-series aligned data output from step S110 as input, it performs raw scoring on the time-series aligned data across three sub-dimensions: completeness, accuracy, and consistency. The raw scores of these three sub-dimensions are then fused using a linear weighting method to obtain a comprehensive quality score. This linear weighting method facilitates the adjustment of weight coefficients and allows for scenario adaptation.

[0044] Furthermore, the integrity score is obtained by calculating the ratio of the amount of statistically aligned time-series data to the theoretically expected amount of industrial data on the standard timestamp granularity grid. The formula for calculating the integrity score is as follows: ; in, To score for completeness, ; This refers to the amount of time-aligned data. This represents the theoretical amount of industrial data that should exist on a standard timestamp granularity grid.

[0045] Furthermore, the normalized dimensionless data is compared with the normalized standard true value of the intelligent manufacturing equipment under normal steady-state operating conditions, and then the accuracy score is obtained by inversely calculating the average deviation. The formula for calculating the accuracy score is as follows: ; in, For accuracy scoring, ; For the first A dimensionless data point; For the first The normalized standard true value is a dimensionless standard reference data obtained by intelligent manufacturing equipment under normal steady-state operating conditions and processed by the same set of adaptive normalization methods. This represents the total number of data samples within the current sliding window.

[0046] Furthermore, using the normalized dimensionless data, the degree of data fluctuation between adjacent time points is calculated to obtain a consistency score. The formula for calculating the consistency score is as follows: ; in, For consistency scoring, ; For the first Dimensionless data.

[0047] Arrange the original scores of the above three sub-dimensions in sequence to form the original score vector. , the original score vector The original scores of the three sub-dimensions are fused using a linear weighting method to obtain the comprehensive quality score. The expression for the comprehensive quality score is as follows: ; in, For the overall quality score, ; These are the weight coefficients for the three sub-dimensions, and .

[0048] Furthermore, based on the emphasis requirements of each sub-dimensional of the comprehensive quality score in different production scenarios of intelligent manufacturing, the weight coefficients of each sub-dimensional can be dynamically adjusted to achieve intelligent quality assessment that adapts to different scenarios.

[0049] The risk dimension assessment module adopts the interval deviation quantification risk assessment algorithm. It takes the dimensionless data after adaptive normalization processing output in step S110 as input, verifies the safe value range for each dimensionless data, calculates the deviation of each dimensionless data relative to the safe value range, and then calculates the average of the deviations of all dimensionless data to obtain the risk score.

[0050] Furthermore, the formula for calculating the deviation is as follows: ; in, For the first Deviation of dimensionless data; This is the lower limit of the safe value range; This represents the upper limit of the safe value range.

[0051] Furthermore, the expression for the risk score is as follows: ; in, To score the risk, , The higher the value, the more significantly the operating parameters of the intelligent manufacturing equipment deviate from the normal safe range, and the higher the risk of equipment operation.

[0052] The timeliness dimension assessment module adopts the timeliness decay model. It takes the time-series aligned data output in step S110 as input, extracts the generation timestamp of each time-series aligned data, and calculates the timeliness score by combining the timeliness decay law.

[0053] Furthermore, the expression for the timeliness score is as follows: ; in, Rate the timeliness. The closer the value is to 1, the more timely the data is; the closer the value is to 0, the less timely the data is. The initial timeliness score is usually set to 1; For different stages of time-related decay, and ; For the generation timestamp of time-series aligned data; The time thresholds are set manually based on industrial scenarios such as the business cycle of intelligent manufacturing equipment and the effective preservation time of data; This is the current scoring moment.

[0054] By relying on a multi-dimensional AI evaluation model with a modular architecture and combining it with the normal steady-state operating conditions of intelligent manufacturing equipment, the system dynamically calculates various scores from four dimensions: value, quality, risk, and timeliness, providing reliable sub-item data support for the subsequent integrated calculation of comprehensive trust scores.

[0055] Value trend score to be obtained Overall quality score Risk Score and timeliness rating Then, the scoring fusion module uses an adaptive weight fusion method to combine the scores from each dimension (value trend score). Overall quality score Risk Score and timeliness rating The data is then integrated and calculated to obtain a comprehensive trust score.

[0056] Furthermore, the formula for fusion calculation is as follows: ; in, A comprehensive trust score is assigned based on the data. For adaptive weighting coefficients, satisfying The adaptive weighting coefficient can be dynamically adjusted based on feedback indicators of implementation effectiveness.

[0057] As an example, the initial value of the adaptive weight coefficient is set to... , , , .

[0058] Taking the vibration data of a certain batch of CNC machine tool spindles as an example, the evaluation results are as follows: Value Trend Score , Overall quality score , Risk Score , Timeliness rating , Overall Trust Score ; As another example, the evaluation results for spare parts inventory data of a certain batch of discontinued models are as follows: Value Trend Score , Overall quality score , Risk Score , Timeliness rating , Overall Trust Score , Step S130: Based on the comprehensive trust score of the data, implement the flow control of industrial data throughout its entire life cycle. The data lifecycle is divided into hot data, warm data, and cold data stages. In addition, an elimination stage can be added, based on a comprehensive trust score. It performs flow control of industrial data throughout its entire lifecycle (hot data stage, warm data stage, cold data stage, and obsolescence stage), automatically assigns storage locations and storage strategies to industrial data, and completes the full lifecycle management of industrial data in the hot data stage, warm data stage, cold data stage, and obsolescence stage.

[0059] when When (and the data value trend is upward), industrial data remains or migrates to the hot data stage and is stored in the high-performance storage area of ​​SSDs; when When (or the trend of data value change tends to stabilize), industrial data migrates to the warm data stage and is stored in the medium-performance storage area of ​​SSDs; when (And the data value trend continues to decline, risk score) When the data is below the risk threshold, industrial data is migrated to the cold data stage and archived to the low-performance storage area of ​​SSDs; when... (and risk score) When the risk threshold is exceeded, the data eviction process is triggered, and the corresponding industrial data is deleted from the SSD storage area.

[0060] Furthermore, by comparing the current value trend score with the historical value trend scores, the trend of data value change is determined. The formula for calculating the data value change trend is as follows: ; in, For the changing trends of data value; Rate the current value trend; Rate the historical value trend; The number of scores given to historical value trends.

[0061] As an example, a preset upper limit threshold for the risk safety range is provided. Risk safety range lower limit threshold , The data was determined to be stored in the high-performance storage area of ​​an SSD, and multiple index copies were created to ensure access efficiency.

[0062] As yet another example The system determined that the batch of industrial data should be moved from the high-performance storage area of ​​the SSD to the medium-performance storage area of ​​the SSD and stored using columnar compression. This would reduce storage space and costs while retaining basic query functions such as field retrieval and condition filtering.

[0063] By relying on comprehensive trust scores to divide the data storage lifecycle into four categories—hot data stage, warm data stage, cold data stage, and obsolescence stage—and combining the data value change trend with risk level to determine the flow direction of industrial data, we can achieve hierarchical storage scheduling of industrial data and optimize storage space allocation.

[0064] Step S140: Collect feedback indicators of the implementation effect of industrial data at each stage of the life cycle, and adaptively iterate and update the multi-dimensional AI evaluation model based on the implementation effect feedback indicators.

[0065] Collect feedback indicators of the implementation effect of industrial data at each stage of its life cycle, such as access frequency, reuse rate, storage cost, anomaly risk, and audit non-compliance. Combine these implementation effect feedback indicators with the time-series forgetting decay coefficient for normalization and time-series weighted aggregation to obtain the comprehensive cost loss value.

[0066] Furthermore, the formula for calculating the overall cost loss value is as follows: ; in, This represents the total cost and loss value. For the first Weighting coefficients for implementation effect feedback indicators. ; The number of types of indicators for feedback on implementation effectiveness; For the first Normalized results of implementation effect feedback indicators; This is the temporal forgetting decay coefficient, which is a constant. For example, it can be 0.0012. If rapid forgetting is required, it can be 0.01. This represents the time interval between the normalized result and the current statistic; As a time-series forgetting factor, the weight of old indicators decays exponentially over time; For the first The first of the feedback indicators for implementation effectiveness One implementation effect feedback indicator; For the first The number of metrics for feedback on the implementation results; For the first The minimum value among the feedback indicators for the implementation effect; For the first The maximum value among the feedback metrics for the implementation effect; It is a minimal positive perturbation factor. For example, the value is 0.0001.

[0067] Based on the overall cost loss value, perform gradient optimization for each component, and solve for the overall cost loss value with respect to the adaptive weight coefficients. The gradient is then used to calculate the adaptive weight coefficients by combining the learning rate and the overall cost loss value. The correction step size is determined, and then the pre-update value of the adaptive weight coefficient is obtained by subtracting the correction step size from the original adaptive weight coefficient. Then, the pre-update value of the adaptive weight coefficient is locked in the range of 0 to 1 by the amplitude limit constraint, so as to obtain the updated adaptive weight coefficient.

[0068] The update formula for the adaptive weight coefficients is as follows: ; in, For the updated adaptive weight coefficients ; The adaptive weight coefficients before the update ; Corresponding to adaptive weight coefficients ; The learning rate is used for iteration. This represents the total cost and loss value. The total cost loss value For the adaptive weight coefficients respectively Find the gradient; For the first Item dimension weights; For the clipping function, Limit the amplitude to between 0 and 1, and satisfy the following conditions: Constraints.

[0069] The updated adaptive weight coefficients to be calculated Then, the updated adaptive weight coefficients will be... Replace the adaptive weight coefficients in the multi-dimensional AI evaluation model before the update This enables adaptive iterative updates to the multi-dimensional AI evaluation model.

[0070] Example 2 like Figure 2 As shown, this application provides a full lifecycle management system 200 based on a multi-dimensional AI evaluation model, including: a data acquisition and preprocessing module 210, a multi-dimensional AI evaluation model 220, a data management and control module 230, and a model iteration and update module 240.

[0071] The data acquisition and preprocessing module 210 acquires industrial data from intelligent manufacturing equipment and preprocesses the industrial data.

[0072] Taking an intelligent manufacturing plant for automotive parts as an example, this application scenario is illustrated. The plant is equipped with intelligent manufacturing equipment such as CNC machine tools, industrial robots, AGV transport vehicles, and vision inspection equipment. Lightweight data acquisition probes are deployed on the PLC controller and edge gateway of each intelligent manufacturing device. These probes collect multi-source heterogeneous industrial data in real time, including equipment operation data, production process data, quality inspection data, and environmental monitoring data, at a predetermined sampling frequency.

[0073] As an example, for CNC machine tools: spindle load, feed rate, tool wear status, and vibration spectrum data are collected at a sampling frequency of 1kHz; for industrial robots: joint angle, torque feedback, and position error data are collected at a sampling frequency of 500Hz; for vision inspection equipment: product images, defect recognition results, and inspection confidence data are collected through triggered acquisition; for environmental sensors: temperature, humidity, and cleanliness data are collected at a sampling frequency of 1Hz.

[0074] The collected industrial data undergoes sequential processing according to a predetermined sequence, including missing value completion, outlier identification and removal, and time-series resampling and alignment, generating time-series aligned data. This time-series aligned data is then adaptively normalized to generate dimensionless data, completing basic preprocessing. Next, time-frequency joint feature extraction is performed on this dimensionless data to construct a data feature vector, completing advanced preprocessing. The time-series aligned data and dimensionless data generated during basic preprocessing, along with the data feature vector constructed during advanced preprocessing, are output together for use in subsequent multi-dimensional AI evaluation models. By limiting the preprocessing to a sequential, step-by-step process—forcing the output of the previous step as the input of the next—the processing chain of industrial data becomes controllable and traceable, adaptable to the management of the entire lifecycle of industrial data.

[0075] Since multi-source heterogeneous industrial data is collected at different sampling frequencies, a standard timestamp granularity grid can be constructed using a global industrial clock as a unified time reference. For example, the grid time step of the standard timestamp granularity grid is 100ms / 1s. All industrial data originally collected at different sampling frequencies such as 1kHz, 500Hz, 1Hz, and triggered sampling are interpolated and mapped onto the standard timestamp granularity grid to generate time-aligned data. This completes the time-series resampling alignment of multi-source heterogeneous industrial data, achieving time-series synchronization of multi-source heterogeneous industrial data and avoiding the pain points of time-series misalignment and inability to jointly model industrial data. Among these, the global industrial clock is the reference clock source for unified time synchronization of all acquisition devices, edge gateways, and controllers within the intelligent manufacturing production line, serving as the sole time reference for time-series alignment of multi-source heterogeneous industrial data.

[0076] Furthermore, the expression for timing resampling alignment is as follows: ; in, In the first Temporal alignment data generated at each target difference time point; For the first The original sampling time points are not uniform across different intelligent manufacturing devices; For the first One original sampling time point; For the first time stamp granularity grid One target interpolation time; In the first Industrial data collected at each original sampling time point; In the first Industrial data collected at each of the original sampling time points.

[0077] In steady-state mass production industrial data, the maximum and minimum values ​​of industrial data are often equal, which can easily lead to failure due to the denominator being zero during normalization. Therefore, a minimal positive perturbation factor is introduced to uniformly map industrial data of different dimensions and amplitudes to the [0,1] interval, thereby achieving adaptive normalization processing of industrial data. By introducing a minimal positive perturbation factor as a protection factor for the denominator during normalization, the problem of failure due to the denominator being zero during normalization is solved, thus enabling the full lifecycle management of industrial data to adapt to steady-state mass production scenarios.

[0078] Furthermore, the expression for adaptive normalization is as follows: ; in, The data are dimensionless after normalization; For input timing-aligned data; This represents the maximum value in this batch of industrial datasets; This is the minimum value in this batch of industrial datasets; It is a minimal positive perturbation factor. For example, the value is 0.0001.

[0079] The operating status of intelligent manufacturing equipment simultaneously contains both time-domain steady-state variation patterns and frequency-domain vibration harmonic characteristics. A single domain feature cannot fully characterize the operating conditions of intelligent manufacturing equipment and the hidden value of industrial data. Therefore, time-domain statistical features and frequency-domain spectral features can be extracted simultaneously based on dimensionless data. Then, the time-domain statistical features and frequency-domain spectral features are concatenated to generate a data feature vector, thereby providing regular, synchronous, and highly representative input data for subsequent multi-dimensional AI evaluation models. This ensures the continuity of industrial data processing and avoids information gaps.

[0080] Furthermore, the expression for the time-domain statistical characteristics is as follows: ; in, This represents the mean of the data within the sliding window. The standard deviation of the data within the sliding window; The root mean square value of the data within the sliding window; This represents the total number of data samples within the current sliding window. For the first Dimensionless data.

[0081] Furthermore, the expression for the frequency domain spectral characteristics is as follows: ; in, For the first Spectral amplitude of first-order frequency components; This refers to the sequence number of the data sample; For the first A dimensionless data point; This represents the total number of data samples within the current sliding window. These are the frequency domain harmonic component indices, corresponding to different sampling frequencies; It is the imaginary unit.

[0082] Furthermore, the expression for generating the data feature vector by concatenating time-domain statistical features and frequency-domain spectral features is as follows: ; in, The resulting data feature vector is used as input to a multi-dimensional AI evaluation model. The amplitude of the first-order frequency component spectrum. The amplitude of the second-order frequency component spectrum. For the first Spectral amplitude of first-order frequency components; This represents the number of frequency component spectral amplitudes.

[0083] By relying on the global industrial clock for time-series resampling alignment, adaptive normalization with positive disturbance factor protection, and a three-level preprocessing logic of time-frequency joint feature extraction, this approach aligns with the constraints of complex intelligent manufacturing conditions and is suitable for processing multi-source heterogeneous industrial data under differentiated sampling frequencies. It fundamentally avoids defects such as time sequence misalignment, anomaly rejection, inconsistent dimensions, and one-sided feature representation. This provides a standardized input for subsequent multi-dimensional AI evaluation models to assess quality, value, risk, and timeliness, effectively ensuring the accuracy and reliability of all-dimensional scoring calculations.

[0084] The multi-dimensional AI evaluation model 220 scores the pre-processed industrial data from multiple dimensions and integrates these scores to obtain a comprehensive data trust score.

[0085] A multi-dimensional AI evaluation model is constructed and trained using historical industrial data. The multi-dimensional AI evaluation model 220 includes: a value dimension evaluation module 221, a quality dimension evaluation module 222, a risk dimension evaluation module 223, a timeliness dimension evaluation module 224, and a score fusion module 225. Preprocessed industrial data is input into the trained multi-dimensional AI evaluation model, and evaluation is performed through the different modules of the model to obtain a multi-dimensional score.

[0086] Among them, the value dimension assessment module 221 adopts a non-linear value assessment model, based on data feature vectors. As input, the data feature vector The mean of the data in Data standard deviation Root mean square value of data By comparing the time-domain statistical characteristics with those of intelligent manufacturing equipment under normal steady-state operating conditions, a time-domain feature value score is obtained, and then the data feature vector is... The distance between the spectral amplitude of each frequency component in the data and the reference spectral amplitude under normal steady-state operation of the intelligent manufacturing equipment is measured to obtain the frequency domain feature value score. Then, the time-frequency value score is fused by the adaptive weighting method, and then the value trend score is output by linear mapping and Sigmoid nonlinear activation.

[0087] Furthermore, the formula for calculating the time-domain feature value is as follows: ; in, Score the value of time-domain features; This represents the average baseline data for normal steady-state operation of intelligent manufacturing equipment. This serves as the standard deviation of the baseline data for the normal steady-state operating conditions of intelligent manufacturing equipment. The root mean square value is the baseline data for the normal steady-state operating conditions of intelligent manufacturing equipment. and The baseline time-domain statistical characteristics of intelligent manufacturing equipment under normal steady-state operating conditions; Weighted by the data mean, Weights are the standard deviations of the data. The weights are the root mean square values ​​of the data. .

[0088] Furthermore, the formula for calculating the frequency domain feature value score is as follows: ; in, Score the value of frequency domain features; For the first Spectral amplitude of first-order frequency components; For the first The first-order reference spectrum amplitude; This represents the Euclidean distance between the current spectral amplitude and the reference vector.

[0089] Furthermore, the expression for the fusion of time-frequency value scores is as follows: ; in, This is the base value score after the fusion of time-frequency value scores; To enable adaptive fusion weights, adjustments can be made dynamically based on the type of intelligent manufacturing equipment and production conditions.

[0090] Furthermore, the expression for the value trend score is as follows: ; in, Score the value trend. ; These are the output weight coefficients of the linear mapping; It is a sigmoid nonlinear activation function that compresses the values ​​to [0,1].

[0091] The quality dimension assessment module 222 employs a multi-dimensional adaptive weighted quality fusion model. Using time-series aligned data as input, it assigns raw scores to the data across three sub-dimensions: completeness, accuracy, and consistency. These raw scores are then fused using a linear weighting method to obtain a comprehensive quality score. This linear weighting approach facilitates the adjustment of weighting coefficients and allows for scenario adaptation.

[0092] Furthermore, the integrity score is obtained by calculating the ratio of the amount of statistically aligned time-series data to the theoretically expected amount of industrial data on the standard timestamp granularity grid. The formula for calculating the integrity score is as follows: ; in, To score for completeness, ; This refers to the amount of time-aligned data. This represents the theoretical amount of industrial data that should exist on a standard timestamp granularity grid.

[0093] Furthermore, the normalized dimensionless data is compared with the normalized standard true value of the intelligent manufacturing equipment under normal steady-state operating conditions, and then the accuracy score is obtained by inversely calculating the average deviation. The formula for calculating the accuracy score is as follows: ; in, For accuracy scoring, ; For the first A dimensionless data point; For the first The normalized standard true value is a dimensionless standard reference data obtained by intelligent manufacturing equipment under normal steady-state operating conditions and processed by the same set of adaptive normalization methods. This represents the total number of data samples within the current sliding window.

[0094] Furthermore, using the normalized dimensionless data, the degree of data fluctuation between adjacent time points is calculated to obtain a consistency score. The formula for calculating the consistency score is as follows: ; in, For consistency scoring, ; For the first Dimensionless data.

[0095] Arrange the original scores of the above three sub-dimensions in sequence to form the original score vector. , the original score vector The original scores of the three sub-dimensions are fused using a linear weighting method to obtain the comprehensive quality score. The expression for the comprehensive quality score is as follows: ; in, For the overall quality score, ; These are the weight coefficients for the three sub-dimensions, and .

[0096] Furthermore, based on the emphasis requirements of each sub-dimensional of the comprehensive quality score in different production scenarios of intelligent manufacturing, the weight coefficients of each sub-dimensional can be dynamically adjusted to achieve intelligent quality assessment that adapts to different scenarios.

[0097] The risk dimension assessment module 223 adopts the interval deviation quantification risk assessment algorithm. It takes the dimensionless data after adaptive normalization as input, verifies the safe value range for each dimensionless data, calculates the deviation of each dimensionless data relative to the safe value range, and then calculates the average of the deviations of all dimensionless data to obtain the risk score.

[0098] Furthermore, the formula for calculating the deviation is as follows: ; in, For the first Deviation of dimensionless data; This is the lower limit of the safe value range; This represents the upper limit of the safe value range.

[0099] Furthermore, the expression for the risk score is as follows: ; in, To score the risk, , The higher the value, the more significantly the operating parameters of the intelligent manufacturing equipment deviate from the normal safe range, and the higher the risk of equipment operation.

[0100] The timeliness dimension assessment module 224 adopts the timeliness decay model. It takes time-series aligned data as input, extracts the generation timestamp of each time-series aligned data, and calculates the timeliness score by combining the timeliness decay law.

[0101] Furthermore, the expression for the timeliness score is as follows: ; in, Rate the timeliness. The closer the value is to 1, the more timely the data is; the closer the value is to 0, the less timely the data is. The initial timeliness score is usually set to 1; For different stages of time-related decay, and ; For the generation timestamp of time-series aligned data; The time thresholds are set manually based on industrial scenarios such as the business cycle of intelligent manufacturing equipment and the effective preservation time of data; This is the current scoring moment.

[0102] By relying on a multi-dimensional AI evaluation model with a modular architecture and combining it with the normal steady-state operating conditions of intelligent manufacturing equipment, the system dynamically calculates various scores from four dimensions: value, quality, risk, and timeliness, providing reliable sub-item data support for the subsequent integrated calculation of comprehensive trust scores.

[0103] Value trend score to be obtained Overall quality score Risk Score and timeliness rating Subsequently, the scoring fusion module 225 uses an adaptive weight fusion method to combine the scores of each dimension (value trend score). Overall quality score Risk Score and timeliness rating The data is then integrated and calculated to obtain a comprehensive trust score.

[0104] Furthermore, the formula for fusion calculation is as follows: ; in, A comprehensive trust score is assigned based on the data. For adaptive weighting coefficients, satisfying The adaptive weighting coefficient can be dynamically adjusted based on feedback indicators of implementation effectiveness.

[0105] As an example, the initial value of the adaptive weight coefficient is set to... .

[0106] Taking the vibration data of a certain batch of CNC machine tool spindles as an example, the evaluation results are as follows: Value Trend Score , Overall quality score , Risk Score , Timeliness rating , Overall Trust Score ; As another example, the evaluation results for spare parts inventory data of a certain batch of discontinued models are as follows: Value Trend Score , Overall quality score , Risk Score , Timeliness rating , Overall Trust Score , The data management module 230 performs flow management and control of industrial data throughout its entire lifecycle based on comprehensive data trust scoring.

[0107] The data lifecycle is divided into hot data, warm data, and cold data stages. In addition, an elimination stage can be added, based on a comprehensive trust score. It performs flow control of industrial data throughout its entire lifecycle (hot data stage, warm data stage, cold data stage, and obsolescence stage), automatically assigns storage locations and storage strategies to industrial data, and completes the full lifecycle management of industrial data in the hot data stage, warm data stage, cold data stage, and obsolescence stage.

[0108] when When (and the data value trend is upward), industrial data remains or migrates to the hot data stage and is stored in the high-performance storage area of ​​SSDs; when When (or the trend of data value change tends to stabilize), industrial data migrates to the warm data stage and is stored in the medium-performance storage area of ​​SSDs; when (And the data value trend continues to decline, risk score) When the data is below the risk threshold, industrial data is migrated to the cold data stage and archived to the low-performance storage area of ​​SSDs; when... (and risk score) When the risk threshold is exceeded, the data eviction process is triggered, and the corresponding industrial data is deleted from the SSD storage area.

[0109] Furthermore, by comparing the current value trend score with the historical value trend scores, the trend of data value change is determined. The formula for calculating the data value change trend is as follows: ; in, For the changing trends of data value; Rate the current value trend; Rate the historical value trend; The number of scores given to historical value trends.

[0110] As an example, a preset upper limit threshold for the risk safety range is provided. Risk safety range lower limit threshold , The data was determined to be stored in the high-performance storage area of ​​an SSD, and multiple index copies were created to ensure access efficiency.

[0111] As yet another example The system determined that the batch of industrial data should be moved from the high-performance storage area of ​​the SSD to the medium-performance storage area of ​​the SSD and stored using columnar compression. This would reduce storage space and costs while retaining basic query functions such as field retrieval and condition filtering.

[0112] By relying on comprehensive trust scores to divide the data storage lifecycle into four categories—hot data stage, warm data stage, cold data stage, and obsolescence stage—and combining the data value change trend with risk level to determine the flow direction of industrial data, we can achieve hierarchical storage scheduling of industrial data and optimize storage space allocation.

[0113] The model iteration and update module 240 collects feedback indicators of the implementation effect of industrial data at each stage of the life cycle, and adaptively iterates and updates the multi-dimensional AI evaluation model based on the implementation effect feedback indicators.

[0114] Collect feedback indicators of the implementation effect of industrial data at each stage of its life cycle, such as access frequency, reuse rate, storage cost, anomaly risk, and audit non-compliance. Combine these implementation effect feedback indicators with the time-series forgetting decay coefficient for normalization and time-series weighted aggregation to obtain the comprehensive cost loss value.

[0115] Furthermore, the formula for calculating the overall cost loss value is as follows: ; in, This represents the total cost and loss value. For the first Weighting coefficients for implementation effect feedback indicators. ; The number of types of indicators for feedback on implementation effectiveness; For the first Normalized results of implementation effect feedback indicators; This is the temporal forgetting decay coefficient, which is a constant. For example, it can be 0.0012. If rapid forgetting is required, it can be 0.01. This represents the time interval between the normalized result and the current statistic; As a time-series forgetting factor, the weight of old indicators decays exponentially over time; For the first The first of the feedback indicators for implementation effectiveness One implementation effect feedback indicator; For the first The number of metrics for feedback on the implementation results; For the first The minimum value among the feedback indicators for the implementation effect; For the first The maximum value among the feedback metrics for the implementation effect; It is a minimal positive perturbation factor. For example, the value is 0.0001.

[0116] Based on the overall cost loss value, perform gradient optimization for each component, and solve for the overall cost loss value with respect to the adaptive weight coefficients. The gradient is then used to calculate the adaptive weight coefficients by combining the learning rate and the overall cost loss value. The correction step size is determined, and then the pre-update value of the adaptive weight coefficient is obtained by subtracting the correction step size from the original adaptive weight coefficient. Then, the pre-update value of the adaptive weight coefficient is locked in the range of 0 to 1 by the amplitude limit constraint, so as to obtain the updated adaptive weight coefficient.

[0117] The update formula for the adaptive weight coefficients is as follows: ; in, For the updated adaptive weight coefficients ; The adaptive weight coefficients before the update ; Corresponding to adaptive weight coefficients ; The learning rate is used for iteration. This represents the total cost and loss value. The total cost loss value For the adaptive weight coefficients respectively Find the gradient; For the first Item dimension weights; For the clipping function, Limit the amplitude to between 0 and 1, and satisfy the following conditions: Constraints.

[0118] The updated adaptive weight coefficients to be calculated Then, the updated adaptive weight coefficients will be... Replace the adaptive weight coefficients in the multi-dimensional AI evaluation model before the update This enables adaptive iterative updates to the multi-dimensional AI evaluation model.

[0119] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0120] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A full lifecycle management method based on a multi-dimensional AI evaluation model, characterized in that, Includes the following steps: Step S110: Collect industrial data from intelligent manufacturing equipment and preprocess the industrial data; Step S120: Input the preprocessed industrial data into the multi-dimensional AI evaluation model to obtain multi-dimensional scores, and fuse the multi-dimensional scores to obtain a comprehensive data trust score. Step S130: Based on the comprehensive trust score of the data, implement the flow control of industrial data throughout its entire life cycle. Step S140: Collect feedback indicators of the implementation effect of industrial data at each stage of the life cycle, and adaptively iterate and update the multi-dimensional AI evaluation model based on the implementation effect feedback indicators.

2. The full lifecycle management method based on a multi-dimensional AI evaluation model according to claim 1, characterized in that, The industrial data is processed step by step according to a predetermined serial process, including missing value completion, outlier identification and removal, and time-series resampling and alignment, to generate time-series aligned data. The time-series aligned data is then subjected to adaptive normalization to generate dimensionless data, thus completing the basic preprocessing. Time-frequency joint feature extraction is performed based on dimensionless data to construct data feature vectors and complete high-order preprocessing.

3. The full lifecycle management method based on a multi-dimensional AI evaluation model according to claim 2, characterized in that, The mean, standard deviation, and root mean square value of the data in the data feature vector are compared with the benchmark time-domain statistical features to obtain the time-domain feature value score. The distance between the spectral amplitude of each frequency component in the data feature vector and the benchmark spectral amplitude is measured to obtain the frequency-domain feature value score. The time-frequency value scores are fused through an adaptive weighting method, and then the value trend score is output after linear mapping and Sigmoid nonlinear activation. The time-series aligned data is scored in three sub-dimensions: completeness, accuracy, and consistency. The scores of the three sub-dimensions are then fused using a linear weighting method to obtain a comprehensive quality score. For each dimensionless data point, the safe range of values ​​is verified one by one. The deviation of each dimensionless data point from the safe range of values ​​is calculated. The average deviation of all dimensionless data points is then calculated to obtain the risk score. Extract the generation timestamp of each time-series aligned data and calculate the timeliness score based on the timeliness decay law; The scores from each dimension are fused and calculated using an adaptive weight fusion method to obtain a comprehensive data trust score.

4. The full lifecycle management method based on a multi-dimensional AI evaluation model according to claim 1, characterized in that, When data is used to comprehensively assess trust scores When the risk safety range upper limit threshold is reached and the data value change trend is rising, industrial data is maintained or migrated to the hot data stage and stored in the high-performance storage area of ​​SSD. When the lower limit threshold of the risk safety range Data-driven trust scoring When the upper limit of the risk safety range or the trend of data value change tends to be stable, industrial data is migrated to the warm data stage and stored in the medium-performance storage area of ​​SSD. When data is used to comprehensively assess trust scores When the risk safety range lower limit threshold is reached, the data value trend continues to decline, and the risk score is lower than the risk threshold, industrial data is migrated to the cold data stage and archived to the low-performance storage area of ​​SSD. When data is used to comprehensively assess trust scores When the risk score exceeds the lower limit of the risk safety range and the risk score exceeds the risk threshold, the data eviction process is triggered, and the corresponding industrial data is deleted from the SSD storage area.

5. The full lifecycle management method based on a multi-dimensional AI evaluation model according to claim 1, characterized in that, The implementation effect feedback indicators of industrial data at each stage of the life cycle are collected, and the implementation effect feedback indicators are normalized and time-weightedly aggregated in combination with the time-series forgetting decay coefficient to obtain the comprehensive cost loss value. Based on the comprehensive cost loss value, perform gradient optimization for each item, solve the gradient of the comprehensive cost loss value with respect to the adaptive weight coefficients, calculate the correction step size of each adaptive weight coefficient by combining the learning rate and the comprehensive cost loss value, obtain the pre-update value of the adaptive weight coefficient by subtracting the correction step size from the original adaptive weight coefficient, and lock the pre-update value of the adaptive weight coefficient in the range of 0 to 1 by the amplitude constraint to obtain the updated adaptive weight coefficient. The updated adaptive weight coefficients replace the previous adaptive weight coefficients in the multi-dimensional AI evaluation model, thereby enabling adaptive iterative updates to the multi-dimensional AI evaluation model.

6. A full lifecycle management system based on a multi-dimensional AI evaluation model, characterized in that, include: The system includes a data acquisition and preprocessing module, a multi-dimensional AI evaluation model, a data management and control module, and a model iteration and update module. The data acquisition and preprocessing module collects industrial data from intelligent manufacturing equipment and performs preprocessing on the industrial data. A multi-dimensional AI evaluation model scores the pre-processed industrial data from multiple dimensions, and integrates the multi-dimensional scores to obtain a comprehensive data trust score. The data management and control module, based on comprehensive data trust scoring, implements flow management and control of industrial data throughout its entire lifecycle. The model iteration and update module collects feedback indicators of the implementation effect of industrial data at each stage of the life cycle, and adaptively iterates and updates the multi-dimensional AI evaluation model based on the implementation effect feedback indicators.

7. The full lifecycle management system based on a multi-dimensional AI evaluation model according to claim 6, characterized in that, The industrial data is processed step by step according to a predetermined serial process, including missing value completion, outlier identification and removal, and time-series resampling and alignment, to generate time-series aligned data. The time-series aligned data is then subjected to adaptive normalization to generate dimensionless data, thus completing the basic preprocessing. Time-frequency joint feature extraction is performed based on dimensionless data to construct data feature vectors and complete high-order preprocessing.

8. The full lifecycle management system based on a multi-dimensional AI evaluation model according to claim 7, characterized in that, The multi-dimensional AI assessment model includes: a value dimension assessment module, a quality dimension assessment module, a risk dimension assessment module, a timeliness dimension assessment module, and a scoring fusion module; Among them, the value dimension assessment module compares the mean, standard deviation, and root mean square value of the data in the data feature vector with the benchmark time-domain statistical features to obtain the time-domain feature value score. It also measures the distance between the spectral amplitude of each frequency component in the data feature vector and the benchmark spectral amplitude to obtain the frequency-domain feature value score. The time-frequency value score is fused through the adaptive weighting method, and then outputs the value trend score through linear mapping and Sigmoid nonlinear activation. The quality dimension assessment module performs raw scoring on time-series aligned data in three sub-dimensions: completeness, accuracy, and consistency. The raw scores of the three sub-dimensions are then fused using a linear weighting method to obtain a comprehensive quality score. The risk dimension assessment module verifies the safe value range for each dimensionless data point, calculates the deviation of each dimensionless data point from the safe value range, and calculates the average deviation of all dimensionless data points to obtain the risk score. The timeliness assessment module extracts the generation timestamp of each time-series aligned data and calculates the timeliness score by combining the timeliness decay law; The scoring fusion module uses an adaptive weight fusion method to merge and calculate the scores from each dimension to obtain a comprehensive data trust score.

9. The full lifecycle management system based on a multi-dimensional AI evaluation model according to claim 6, characterized in that, When data is used to comprehensively assess trust scores When the risk safety range upper limit threshold is reached and the data value change trend is rising, industrial data is maintained or migrated to the hot data stage and stored in the high-performance storage area of ​​SSD. When the lower limit threshold of the risk safety range Data-driven trust scoring When the upper limit of the risk safety range or the trend of data value change tends to be stable, industrial data is migrated to the warm data stage and stored in the medium-performance storage area of ​​SSD. When data is used to comprehensively assess trust scores When the risk safety range lower limit threshold is reached, the data value trend continues to decline, and the risk score is lower than the risk threshold, industrial data is migrated to the cold data stage and archived to the low-performance storage area of ​​SSD. When data is used to comprehensively assess trust scores When the risk score exceeds the lower limit of the risk safety range and the risk score exceeds the risk threshold, the data eviction process is triggered, and the corresponding industrial data is deleted from the SSD storage area.

10. The full lifecycle management system based on a multi-dimensional AI evaluation model according to claim 6, characterized in that, The implementation effect feedback indicators of industrial data at each stage of the life cycle are collected, and the implementation effect feedback indicators are normalized and time-weightedly aggregated in combination with the time-series forgetting decay coefficient to obtain the comprehensive cost loss value. Based on the comprehensive cost loss value, perform gradient optimization for each item, solve the gradient of the comprehensive cost loss value with respect to the adaptive weight coefficients, calculate the correction step size of each adaptive weight coefficient by combining the learning rate and the comprehensive cost loss value, obtain the pre-update value of the adaptive weight coefficient by subtracting the correction step size from the original adaptive weight coefficient, and lock the pre-update value of the adaptive weight coefficient in the range of 0 to 1 by the amplitude constraint to obtain the updated adaptive weight coefficient. The updated adaptive weight coefficients replace the previous adaptive weight coefficients in the multi-dimensional AI evaluation model, thereby enabling adaptive iterative updates to the multi-dimensional AI evaluation model.

Images