Asset information management method and system based on big data and artificial intelligence

Abstract

The invention relates to the technical field of information management, in particular to an asset information management method and system based on big data and artificial intelligence, and solves the problems that an existing asset management model is high in data dependence, weak in generalization ability, unstable in performance and the like. The method comprises the steps that asset full-life-cycle data are collected in a multi-source mode and preprocessed to obtain a real data set, and synthetic data are generated through the AIGC technology to construct a mixed training set; adjusting the pre-training model through deep transfer transfer learning to obtain an asset intelligent analysis model; newly added data are collected in real time based on Flink, and feature drift is detected by using EvidentlyAI; and judging model suitability according to a detection result, if the model suitability is adapted, outputting a multi-dimensional score and decision suggestion, and if the model suitability is not adapted, automatically retraining the model to realize intelligent dynamic management of asset information. The system is matched with visual monitoring, hierarchical decision pushing and machine account management execution mechanisms. According to the method, the asset management intelligence level and the model generalization ability are improved, data dependence is reduced, stable system performance is guaranteed, and accurate and efficient support is provided for asset decision making.

Description

Technical Field

[0001] This invention relates to the field of information management technology, specifically to an asset information management method and system based on big data and artificial intelligence. Background Technology

[0002] As enterprises expand their asset scale and deepen their digital transformation, asset information management has covered the entire lifecycle of procurement, warehousing, operation, maintenance, and disposal. Its level of intelligence directly affects the enterprise's operational efficiency and cost control. The application of big data and artificial intelligence technologies in this field is gradually becoming more widespread, but there are still many problems in actual application.

[0003] Small and medium-sized enterprises (SMEs) struggle to accumulate sufficient labeled data due to their limited asset types, short operating cycles, and inadequate monitoring equipment. Large enterprises also face data scarcity issues with niche assets and newly acquired assets, particularly the lack of critical labeled data such as fault records for new equipment and operational data for assets in specific scenarios. This results in insufficient training of deep learning-based intelligent asset analysis models (such as fault prediction and asset classification models), weak generalization capabilities, and accuracy rates generally below 80%, failing to meet actual business needs and driving up model adaptation costs. This data sparsity problem further exposes the immaturity of model adaptation technologies in small-sample scenarios. Existing transfer learning lacks specificity in asset management scenarios, often directly reusing general model parameters without designing fine-tuning strategies that consider the temporal and correlational characteristics of asset data. Limited real-world data is insufficient to effectively activate the scenario adaptability of pre-trained models, and the lack of standardized processes leads to long training cycles, high resource consumption, and limited accuracy improvements. Even when some technologies attempt to supplement the sample size with synthetic data, the resulting synthetic data often fails to deeply extract the core features of the asset data, leading to poor consistency between the synthetic and real data characteristics and even introducing noise. Furthermore, the lack of a scientifically sound mixed training ratio and data validation mechanism results in low-quality mixed training sets, which not only fail to improve model performance but also reduce model reliability. More importantly, the distribution of asset data dynamically changes with factors such as equipment aging, business adjustments, and environmental changes, resulting in data distribution drift and concept drift. Existing technologies often rely on periodic manual retraining, which cannot monitor drift in real time, and full retraining wastes resources and causes response delays, leading to a continuous decline in model performance and further affecting the accuracy of fault prediction and asset classification.

[0004] Furthermore, key aspects of traditional asset management, such as data preprocessing, model iteration, and drift correction, all rely heavily on manual intervention. From data labeling and parameter adjustment to retraining triggers, this process is not only inefficient but also prone to errors introduced by human intervention. Ultimately, this leads to untimely asset failure warnings and delayed maintenance decisions, increasing operational costs and risks for enterprises. Overall, the integration of the full-chain data processing capabilities of big data with the intelligent optimization capabilities of artificial intelligence has not been effectively achieved, making it difficult to meet the actual needs of enterprises, especially SMEs, for intelligent asset management. Summary of the Invention

[0005] In view of the shortcomings of existing technologies, the purpose of this invention is to provide an asset information management method based on big data and artificial intelligence that effectively integrates the end-to-end data processing capabilities of big data with the intelligent optimization capabilities of artificial intelligence, so as to meet the needs of enterprises, especially small and medium-sized enterprises, for intelligent asset management.

[0006] To achieve the above objectives, the present invention implements an asset information management method based on big data and artificial intelligence through the following technology, the steps of which are as follows: Step S1: Collect and preprocess asset lifecycle data to obtain a real asset dataset. Specifically: Step S11: Obtain historical inventory and initial real-time asset data through multi-source acquisition terminals (including RFID readers, IoT sensors, etc.); Step S12: Clean, deduplicate, and standardize the acquired data; Step S13: Split the dataset into a training subset and a benchmark subset in a 7:3 ratio.

[0007] Step S2: Generate synthetic asset data based on real asset data using AIGC technology. Specifically: Step S21 extracts core features from the training subset; Step S22 trains the model using the core features as input to generate initial synthetic asset data; Step S23 compares the data using a feature consistency verification algorithm to remove outlier data.

[0008] Step S3: Mix real and synthetic asset data in a preset ratio (20% - 30% for training subset and 70% - 80% for synthetic asset data) to construct a hybrid training dataset.

[0009] Step S4: Using deep transfer learning technology, transfer the pre-trained model (CNN-LSTM hybrid model or lightweight Transformer model) of the same asset scene to the target scene, adjust it with the hybrid training dataset, and obtain the asset intelligent analysis model.

[0010] Step S5: Collect new asset data based on the Flink framework. First, deploy a Flink cluster, then perform real-time data preprocessing, including format standardization, outlier removal, duplicate data filtering, and feature vector calculation. Using a baseline subset as a reference, use the EvidentlyAI tool to detect feature drift in the new data.

[0011] Step S6: Determine the model fit based on the feature drift results. If it fits, directly input the new data. If it doesn't fit, retrain until it fits before inputting the new data.

[0012] Step S7: The asset intelligent analysis model outputs multi-dimensional scores and decision suggestions based on the input data, realizing intelligent and dynamic management of asset information. Line feature drift detection process: Step S521: Encapsulate EvidentlyAI as a FlinkUDF and embed it into Flink Stream Processing's ProcessWindowFunction to support integrated stream and batch detection.

[0013] Step S522: Select the core features extracted in step S21 as the drift detection object, and use a sliding window mechanism to match detection indicators according to data type.

[0014] Step S523: Use the PSI value as the core threshold and combine it with other test results for comprehensive judgment. PSI formula: PSI = ∑(i = 1 to n) (Actuali - Expectedi) × ln(Actuali / Expectedi), where Actuali is the proportion of the i-th bin in the real-time data and Expectedi is the proportion of the i-th bin in the baseline data.

[0015] Step S54: Output a detection report once for each sliding window, and store the report in Elasticsearch in JSON format.

[0016] Furthermore, in step S6, the suitability of the asset intelligent analysis model is determined based on the feature drift detection results: using the core feature group stability PSI value as the primary criterion, and Kolmogorov-Smirnov tests and chi-square tests as auxiliary methods, feature drift is classified into three levels: no drift, slight drift, and severe drift. No drift indicates model suitability; slight drift requires model performance sampling verification before further assessment; severe drift indicates model unsuitability and automatic retraining until suitability is achieved.

[0017] Furthermore, in step S7, the final intelligent asset analysis model outputs a score based on the newly added asset data and classifies it into four levels: A, B, C, and D. Level A maintains the current operation and maintenance strategy, with quarterly inspections; Level B optimizes the frequency of operation and maintenance, focusing on monitoring core indicators; Level C clarifies the verification targets, time limits, and responsible parties; and Level D adopts emergency response measures and loss mitigation plans.

[0018] An asset information management system based on big data and artificial intelligence, capable of executing the aforementioned methods, includes: a multi-source asset data acquisition module, which collects asset lifecycle data, parses heterogeneous data, and transmits it uniformly; an asset data preprocessing module, which cleans, deduplicates, and standardizes the raw data, splits it into training and baseline subsets in a 7:3 ratio, and stores them; an AIGC synthetic asset data generation module, which extracts core features from the training subset, generates synthetic asset data using a generative adversarial network, verifies and removes anomalies to obtain data with a feature distribution consistent with real data; and a hybrid training dataset construction module, which constructs a hybrid training dataset using a 20%-30% training subset and a 70%-... The system utilizes 80% synthetic asset data to construct a hybrid training dataset, which is then split into training and validation sets and stored. A deep transfer learning model training module loads models of similar asset scenarios, fine-tunes them using the hybrid training dataset to obtain an intelligent asset analysis model, and stores it. A Flink real-time data acquisition module collects newly added real-time asset data based on a Flink cluster, processes it to generate real-time feature vectors, and performs feature drift detection. A feature drift detection module encapsulates EvidentlyAI as FlinkUDF, uses a benchmark subset as a reference for detection, and outputs a JSON report. A model fit determination and automatic retraining module determines model fit based on drift detection results; if fit, data is pushed to the model; otherwise, a retraining dataset is built and the model is automatically retrained to fit. A multi-dimensional asset scoring and decision suggestion module receives model inference results, calculates multi-dimensional and comprehensive scores, classifies them into A to D levels, generates corresponding decision suggestions, and pushes them according to priority. A system management and visualization module implements data archiving, access control, decision suggestion tracking, system status monitoring, and visualization, supporting collaborative operation throughout the entire process.

[0019] Compared with the prior art, the present invention has the following advantages and beneficial effects: (1) This invention uses AIGC technology to generate synthetic data that is consistent with the feature distribution of real asset data. The quality is guaranteed by feature consistency verification. It combines real and synthetic data for mixed training to make up for the lack of labeled data for small and medium-sized enterprises. Its generalization ability is better than that of traditional models and it is suitable for data-scarce scenarios.

[0020] (2) The invention adopts deep transfer learning technology, reuses the pre-trained model parameters of similar asset scenarios, and through fine-tuning strategies, only more than 50 real labeled data of target scenarios are needed to complete model adaptation, avoiding high resource consumption. The learning rate decay strategy is adopted to ensure the effect, shorten the training cycle, and reduce the cost of enterprise technology use.

[0021] (3) Based on the Flink big data real-time computing framework and EvidentlyAI tool, this invention monitors changes in asset data distribution in real time and accurately identifies drift; it triggers model adaptive adjustment or retraining through incremental learning algorithm, updates only some parameters, avoids resource waste, shortens response and retraining cycle, prevents model performance decay, and ensures reliable results.

[0022] (4) This invention realizes full-chain automation of asset data collection, preprocessing, model training, drift repair and decision output without a lot of manual intervention. It automatically completes each stage of operation, outputs accurate results, reduces errors and difference rates, reduces compliance audit manpower costs, and improves the efficiency of enterprise asset management.

[0023] (5) The asset intelligent analysis model of the present invention covers multiple core functions and outputs decision suggestions to support the full life cycle management of assets; its lightweight architecture, flexible training ratio and efficient retraining mechanism are suitable for enterprises of different sizes, with a wide range of applications and good industrialization prospects. Attached Figure Description

[0024] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a framework diagram of the system in this invention; Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0025] 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.

[0026] To make the objectives, process conditions, and advantages of the present invention clearer, the present invention will be further described in detail with reference to the following embodiments. However, the embodiments of the present invention are not limited thereto. Various substitutions and modifications can be made based on common technical knowledge and conventional means in the art without departing from the above-described technical concept of the present invention, and all such substitutions and modifications should be included within the scope of the present invention. The specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0027] Example 1: This embodiment provides an asset information management system based on big data and artificial intelligence. The system adopts a distributed architecture, and its main structural framework is as follows: Figure 1 As shown, it includes: The multi-source asset data acquisition module collects historical stock data and initial real-time data throughout the entire life cycle of assets, and transmits the heterogeneous data in a unified manner after parsing it. The asset data preprocessing module cleans, deduplicates, and standardizes the collected raw asset data, splitting it into a training subset and a benchmark subset in a 7:3 ratio and storing them. The AIGC synthetic asset data generation module extracts core features from the training subset, generates initial synthetic asset data through a generative adversarial network, and removes anomalies through cosine similarity verification to obtain synthetic asset data with the same feature distribution as the real data. The mixed training dataset construction module constructs a mixed training dataset with a ratio of 20% to 30% training subset and 70% to 80% synthetic asset data, and stores it after splitting the training set and validation set; The deep transfer learning model training module loads a CNN-LSTM hybrid model or a lightweight Transformer pre-trained model for similar asset scenarios, fine-tunes it with the hybrid training dataset to obtain an asset intelligent analysis model and stores it. The Flink real-time data acquisition module collects new real-time asset data based on the Flink cluster, and generates real-time feature vectors through JSON standardization, 3σ outlier removal, deduplication, and core feature extraction. The feature drift detection module encapsulates EvidentlyAI into FlinkUDF. Using a benchmark subset as a reference, it performs drift classification detection by combining PSI values ​​with Kolmogorov-Smirnov tests, chi-square tests, etc., and outputs a JSON format detection report. The model fit determination and automatic retraining module determines the model fit based on the drift detection results. If the model fits, new asset data is pushed to the model. If the model does not fit, a retraining dataset is built by mixing 30% new data with 70% of the original dataset, and the model is automatically retrained until it fits. The multi-dimensional asset scoring and decision-making suggestion module receives model inference results, calculates multi-dimensional and comprehensive scores and classifies them into levels A to D, generates corresponding decision-making suggestions and pushes them according to priority. The system management and visualization module enables data archiving, access control, decision suggestion tracking, and system status monitoring and visualization, supporting collaborative operation across the entire process.

[0028] Among them, the system management and visualization module supports the operation, maintenance, monitoring, data archiving and visualization of the entire system, ensuring the stable operation of the system and the implementation of management decisions.

[0029] This is achieved through the following technologies: Visualization panel: Built on ECharts, it displays asset rating distribution, drift detection results, model retraining progress, and decision suggestion execution status; Data archiving: Score reports, detection reports, and model training logs are archived to HDFS by "asset ID + time" and retained for 3 years; Access control: Roles are defined based on the RBAC model (administrator / operations personnel / decision-making level), and different roles correspond to different operation permissions; Execution tracking: Establish a decision recommendation execution log to record the receipt time, execution progress, and result feedback. If the recommendation is not executed within the specified time, an alarm will be triggered. System monitoring: Monitors the running status (CPU / memory / storage) of Flink cluster, model repository, and database, and pushes alerts to management when anomalies occur. Input: All module runtime data and report data; Outputs: Visual monitoring panel, access control commands, and system alarm information.

[0030] The specific methods and processes for asset information management based on the above system are as follows: Figure 2 As shown, it includes the following steps: Step S1: Collect data throughout the entire asset lifecycle and preprocess the collected asset data to obtain a real asset dataset; Step S2: Based on real asset data, use AIGC technology to generate synthetic asset data with the same feature distribution as the real asset data; Step S3: Mix real asset data and synthetic asset data according to a preset ratio to construct a hybrid training dataset; Step S4: Using deep transfer learning technology, the pre-trained model trained in the same asset scenario is transferred to the target asset management scenario. Then, the pre-trained model after transfer is adjusted through the constructed hybrid training dataset to obtain the asset intelligent analysis model. Similar asset scenarios are those with consistent core asset attributes, similar management processes, and the same data type. The specific criteria are: a matching degree of ≥85% for core asset attributes, including asset category, intended use, and value. The matching degree is calculated using the following formula: ; The overlap rate of management processes is ≥70%, and the overlapping processes include procurement, use, maintenance, troubleshooting, and handling. Data type homogeneity rate ≥ 90%, meaning that the coverage ratio of static attributes, dynamic operation, business process, and historical fault data is consistent. Step S5: Based on the big data real-time computing framework Flink, continuously collect new asset data generated in real time in the target asset management scenario. Using the real asset data collected in Step S1 as a reference standard, call the EvidentlyAI tool to perform feature drift detection on the collected new real-time data. Step S6: Based on the feature drift detection results, determine the suitability of the obtained asset intelligent analysis model; if the asset intelligent analysis model meets the suitability requirements, directly input the newly added asset data into the asset intelligent analysis model; if the asset intelligent analysis model does not meet the suitability requirements, automatically retrain the asset intelligent analysis model until it meets the suitability requirements, and then input the newly added asset data into the asset intelligent analysis model. Step S7: The final intelligent asset analysis model outputs scores for the assets in different dimensions based on the input new asset data. Each score corresponds to a decision suggestion, thereby realizing intelligent and dynamic management of asset information.

[0031] Example 2: Based on the above examples, this example further defines the specific content of the multi-source asset data acquisition module, the asset data preprocessing module, and step S1.

[0032] (I) The multi-source asset data acquisition module enables unified acquisition of asset lifecycle data across multiple terminals and protocols, covering historical inventory and initial real-time data, providing raw data for subsequent processing. The underlying technologies are as follows: hardware access integrates RFID readers, IoT sensors, OCR scanning equipment, and business system interfaces; data reception is deployed via an access gateway, which parses protocols and receives data uniformly to Kafka, with topics partitioned by asset type; data is cached using Redis for 5 minutes. Input consists of various asset data types, and output is the raw heterogeneous asset dataset pushed to Kafka.

[0033] (II) The asset data preprocessing module cleans, deduplicates, and standardizes the raw data, splitting it into training and benchmark subsets. Implementation techniques: Data cleaning uses Spark to remove missing core fields; deduplication uses hash values ​​generated based on ID and timestamp to filter duplicate data; standardization encodes different data types; the dataset is split into training and benchmark subsets in a 7:3 ratio and stored in HBase. The input is the raw dataset from Kafka, and the output is the cleaned dataset stored in HBase.

[0034] (III) Specific Process of Step S1: Step S11 acquires full lifecycle asset data through multi-source acquisition terminals, including historical inventory and initial real-time data. Multi-source acquisition terminals include RFID readers, IoT sensors, OCR scanning devices, and business system interfaces. RFID readers collect location information at a frequency of 5-10 minutes, with an accuracy error ≤1 meter; IoT sensors collect dynamic operation data at a frequency of 1-3 seconds; OCR scanning devices collect relevant records with an accuracy rate ≥98% and support batch processing; the business system interface collects multiple data sets, with a synchronization frequency of once every 30 minutes and a success rate ≥99.9%. Historical inventory data includes static attributes, historical business processes, and historical fault data. Asset data includes historical static data, historical business process data, historical fault data, and real-time newly added asset data. Historical static data includes technical parameters and specifications, storage area number, and responsible person information; historical business process data includes acceptance, calibration, and other records; historical fault data includes the scope of fault impact, etc.; real-time newly added asset data includes dynamic operation data (such as equipment operating environment humidity, noise, etc.) and real-time business process data (such as inspection, fault reporting, and other records).

[0035] Step S12: Clean, deduplicate, and standardize the acquired asset data to obtain the real asset dataset. Numerical data is processed using min-max normalization: first, continuous numerical data is filtered, outliers are removed using the 3σ principle, and missing values ​​are filled in using the median, according to the formula... Normalization is performed, and the filtering range is adjusted if the variance is less than 0.01. Non-numerical data is processed using encoding conversion: it is divided into three categories, binary variables are encoded with 0-1, multi-category variables are encoded with unique codes, text is encoded with TF-IDF, and date data is converted to Unix timestamps.

[0036] Step S13: Split the real asset dataset into a training subset and a baseline subset in a 7:3 ratio.

[0037] Example 3: Further define the AIGC synthetic asset data generation module and step S2 process.

[0038] (i) The AIGC synthetic asset data generation module generates synthetic data with consistent feature distribution based on a training subset. This is achieved through core feature extraction (using the Scikit-learn algorithm to extract core features such as operating temperature), GAN model training (using an improved WGAN-GP network), feature consistency verification (calculating cosine similarity to remove outliers), and data storage (storing to HDFS). The input is the training subset, and the output is the synthetic asset dataset.

[0039] (ii) Step S2: Step S21: Extract core features from the training subset, including core features of numerical (basic statistical features, etc.), time-series (time-series correlation features), categorical (classification distribution features), and textual (semantic features) data. Before extraction, preprocess the training subset, such as data cleaning to remove dirty data. Workflow: Identify and handle outliers (using the 3σ principle), duplicate values ​​(based on asset ID + data timestamp for deduplication), and missing values ​​(median for numerical data, mode for categorical data, and "no description" for textual data); standardize formats: convert date data to Unix timestamps, unify encoding for categorical data, and use UTF-8 encoding for textual data while removing invalid characters; split the dataset into sub-datasets by numerical, categorical, textual, and time-series data.

[0040] Step S212: Extraction of core features from numerical data. ① Basic statistical features, calculation of mean. variance, median Quartiles ② Distribution morphological characteristics, using the Shapiro-Wilk detection method. Verify and determine the distribution type, and record. Key parameters. ③ Cross-item association features, using Pearson correlation. Calculate the correlation degree and retain feature pairs with |r|≥0.3. ④ Time series data characteristics Evidence, calculate time series trends and autocorrelation Features, retaining |ACF(1)|≥0.7 features. ⑤ Categorical data features, calculating the category proportion For comparison and recording of core categories, when associating numerical data, use analysis of variance to retain features with p < 0.05. ⑥ For textual data features, use TF-IDF encoding after text preprocessing. Dimensional reduction.

[0041] Step S216: Feature Integration and Standardization. The core features are concatenated into an initial vector in a specific order, and numerical features are normalized using min-max normalization. The dimension of the feature vector is controlled to be 10≤k≤20.

[0042] Step S217: Redundant feature screening (optimizing feature quality): Variance screening removes features with variance < 0.01; Correlation screening removes one feature if the Pearson correlation coefficient |r| ≥ 0.9; Importance screening uses random forest to calculate feature importance and removes features with importance < 5%.

[0043] Step S22: An optimized Conditional Generative Adversarial Network (CGAN) model is trained using the extracted core features to generate initial synthetic asset data. This model, tailored to the characteristics of asset data, introduces the asset data type as a conditional variable into the basic GAN architecture, addressing a common GAN problem.

[0044] CGAN consists of a generator (G) and a discriminator (D).

[0045] The generator (G) is responsible for generating the initial synthetic asset data: the input layer takes z+c (random noise vector + condition variable vector) as input; there are three hidden layers with 256, 128 and 64 neurons respectively, and the activation function is LeakyReLU (negative slope = 0.2) for all of them; the output layer has a k-dimensional output dimension, the activation function is Tanh, and the output is the core feature vector Vsyn.

[0046] The discriminator (D) is responsible for distinguishing between real and synthetic data: the input layer takes x+c as input, with an input dimension of 2k+3; there are two hidden layers with 128 and 64 neurons respectively, and the activation function is LeakyReLU (negative slope = 0.2); the output layer has an output dimension of 1, the activation function is Sigmoid, and the output data is the confidence level of realism.

[0047] Optimizer and Learning Rate: The generator (G) uses the Adam optimizer with parameters β1=0.5, β2=0.999, and a learning rate lr. G =0.0002; the discriminator (D) uses the Adam optimizer with parameters β1=0.5, β2=0.999, and a learning rate of 0.0002. lrD =0.0001.

[0048] Training batch and iteration count: Batch size 32, maximum number of iterations 2000 rounds, and the generation effect is verified every 100 rounds.

[0049] Loss function: The training objective of CGAN is for the generator to deceive the discriminator, and the discriminator to distinguish between true and false. The loss function consists of two parts: (1) Discriminator loss (L D ): Maximize the distinction between true and false data Where: m is the batch sample size (m = 32); Let be the feature vector of the i-th real asset data; Let i be the i-th random noise vector; Let be the condition variable vector for the i-th sample; D is the i-th synthetic data feature vector output by the generator; D(·) is the confidence output of the discriminator. (2) Generator loss (L G Minimize the discriminant's distinguishing ability The core objective is to increase the discriminator's confidence in the synthetic data. →1, making it impossible for them to distinguish between true and false.

[0050] To adapt to the diversity of asset data, adjustments were made to the basic CGAN architecture: 1. Temporal data: A GAN-LSTM hybrid architecture is adopted, with one LSTM layer (64 neurons) added after the generator hidden layer to capture temporal correlation features and ensure that the generated data conforms to the temporal continuity.

[0051] 2. Categorized data: Embedded category constraints strengthen the weight of category labels in condition variables (multiplied by a coefficient of 1.2), and a Softmax layer is added after the generator output layer to ensure that the category distribution is consistent with the real data.

[0052] 3. Text-based data: A lightweight fusion of CGAN and LLM is adopted. The text semantic feature vector is added to the generator input, and the discriminator adds a text similarity verification branch (cosine similarity ≥ 0.85) to make the generated text semantics match the real data.

[0053] During training, the mean of the loss function and the initial similarity to the generated data are calculated every 100 rounds. The iteration stops when convergence is achieved if either of the following conditions is met: ① Loss function convergence: (This condition is met in two consecutive validation rounds.) ② Preliminary similarity convergence: the mean cosine similarity between the generated data and the real data is ≥0.8; ③ The maximum number of iterations reaches 2000 rounds.

[0054] After the model converges, initial synthetic data is generated according to the following rules: the data volume is 1-2 times that of the training subset; it is generated according to the data type distribution of the assets in the training subset; the generated feature vector V is then used to generate the initial synthetic data. syn The original data format is restored through denormalization and type decoding.

[0055] Step S23: Use the feature consistency verification algorithm to compare the feature similarity between the initial synthetic asset data and the training subset data, calculate the similarity using the cosine similarity formula, remove abnormal data, and obtain synthetic asset data with the feature distribution consistent with the training subset.

[0056] Step S231: Denormalize and reconstruct the feature vector of the initial synthetic asset data to obtain the synthetic feature vector V. sy n. First, restore the initial synthesized data to its original data format, then perform inverse normalization. The formula is (where is the normalized value of the synthesized feature vector, To train the subset corresponding to the maximum / minimum features); then remove extreme outliers (using the 3σ primitive). but, , (To obtain the mean and variance of the corresponding features for the training subset), finally, according to the rule in step S21, the core features are extracted again to construct the synthetic feature vector. (10≤k≤20), ensuring that the true feature vectors of the training subset are consistent. V real Dimensions are consistent.

[0057] Step S232: Based on the core features of the training subset extracted in step S21, calculate the average feature vector as the verification benchmark. The formula is... for( M The number of samples in the training subset. For training subset number i Core feature vectors of each sample; (Based on the true feature vector).

[0058] Step S233: For each preprocessed initial synthetic data, calculate its similarity to the benchmark true feature vector using the cosine similarity formula. similarity Sim ,formula For (j) The index of the core feature components (1≤j≤k); The j-th component of the baseline true feature vector; This represents the j-th component of the synthesized eigenvector; "·" denotes the vector dot product. (represents the L2 norm of a vector).

[0059] Step S234: Perform similarity screening and outlier data removal: Set a similarity threshold θ=0.9 (at this threshold, the consistency of the synthetic and real data feature distribution is ≥90%), screen out the initial synthetic data with Sim≥θ as candidate valid synthetic data, and remove those with Sim<θ and mark them as feature deviation outliers.

[0060] Step S235: Perform secondary verification of valid data. Calculate the percentage of valid composite data. ,male If the number of candidate valid synthetic data is 0%, the data is determined to be sufficient, and this batch of data is output as the final synthetic asset data. If the success rate is less than 80%, the effectiveness is deemed insufficient. Return to step S22, increase the CGAN model's training iterations by 500 rounds (maximum not exceeding 3000 rounds), regenerate the initial synthetic data, and then execute the verification process in step S23. Up to ≥80%.

[0061] Example 4: Based on the above, the content of the mixed training dataset construction module and step S3 is limited.

[0062] (I) Hybrid Training Dataset Construction Module: A hybrid dataset is constructed by mixing a subset of real training data and synthetic asset data in a preset ratio. Technical Implementation: Data ratio control: 20%-30% real training data and 70%-80% synthetic data are extracted, with configurable adjustment supported (default 2:8); Data fusion: The feature dimensions of the hybrid data are validated to ensure consistency with the core feature dimensions, generating a TFRecord file; Dataset partitioning: The dataset is split into a training set and a validation set at an 8:2 ratio, with stratified sampling to ensure that the core data distribution deviation is ≤5%. Inputs are a subset of real asset training data and a synthetic asset dataset; output is a hybrid training dataset (TFRecord format, stored in HDFS).

[0063] (ii) In the mixed training dataset constructed in step S3, the training subset accounts for 20% - 30%, the synthetic asset data accounts for 70% - 80%, and it contains fault prediction and asset classification task annotation information.

[0064] Example 5: Based on the above, the training module of the deep transfer learning model and the content of step S4 are limited.

[0065] (I) The Deep Transfer Learning Model Training Module transfers pre-trained models for similar asset scenarios to the target scenario, and fine-tunes them using a hybrid training dataset to generate an intelligent asset analysis model. Technical Implementation: Pre-trained model loading: Select a CNN-LSTM hybrid model or a lightweight Transformer model based on the asset type, and load pre-trained weights for similar scenarios; Transfer learning strategy: Freeze 70% of the general feature extraction layers and unfreeze 30% of the scenario adaptation layers; Model fine-tuning: Use the AdamW optimizer, setting the learning rate so that the scenario adaptation layer > general feature extraction layer, and iteratively train until the validation set performance meets the requirements; Model storage: Serialize the qualified model and store it in a model repository, managing it by version number. Input: Hybrid training dataset and pre-trained models for similar scenarios; Output: Intelligent asset analysis model (stored in MLflow).

[0066] (ii) Step S4: The pre-trained model is either a CNN-LSTM hybrid model or a lightweight Transformer model. For the target task of asset status assessment and fault early warning, primarily using numerical / time-series data, the CNN-LSTM hybrid model is selected. For the target task of asset value analysis and business process optimization, primarily using text / relational data, the lightweight Transformer model is selected. The pre-trained model must achieve an accuracy ≥ 88%, recall ≥ 85%, and F1 score ≥ 86% on a test set of similar asset scenarios, ensuring its transferable feature extraction capability.

[0067] Step S4: Detailed process: Step S41: Use deep transfer learning to transfer the pre-trained model trained in the same asset scenario to the target asset management scenario.

[0068] Step S411: Divide the pre-trained model into a general feature extraction layer and a scene adaptation layer. The general feature extraction layer accounts for 70% (such as the first 2 layers of CNN-LSTM + 1 layer of LSTM, or the first 4 layers of encoder of Transformer), which extracts general features of assets; the scene adaptation layer accounts for 30% (such as fully connected layers and output layers), which adapts to the specific scene task requirements.

[0069] Step S412: Implement the transfer strategy. Freeze the training of the general feature extraction layer and fix the parameters; unfreeze and fine-tune the scene adaptation layer and initialize the parameters. After transfer, calculate the feature extraction consistency of the pre-trained model on a small sample dataset (accounting for 10% of the mixed training dataset) of the target scene. Consist ≥ 0.8 is required.

[0070] Step S42: Adjust the transferred pre-trained model using the mixed training dataset.

[0071] Step S421: Preprocess the hybrid training dataset. The hybrid training dataset consists of the real asset training subset from step S3 and the synthetic asset data in a 2:8 ratio.

[0072] Step S422: For time-series data, perform timeline shifting and noise injection enhancement; for text-based data, perform synonym replacement and word order adjustment enhancement, increasing the data volume by 1.5 times. Split the dataset into training and validation sets in an 8:2 ratio, and use stratified sampling to ensure that the core data distribution deviation is ≤5%.

[0073] Step S422: Adjust model training parameters. Use the AdamW optimizer, with a weight decay coefficient λ = 0.001, a learning rate of scene adaptation layer > general feature extraction layer (after unfreezing), a batch size of 32 (temporal type) / 64 (text / numerical type), an initial iteration of 1000 rounds, and performance verification every 200 rounds.

[0074] Step 423: Construct the loss function. The cross-entropy loss function is used for classification tasks. The regression task uses the mean squared error loss function. , Task Adjust weights according to importance.

[0075] Step S424: Model adjustment strategy. For the first 500 rounds, only the scene adaptation layer is trained; after 500 rounds, if the performance improvement rate on the validation set is <3%, unfreeze 30% of the general feature extraction layer and train it jointly; if the validation set loss increases or the performance improvement rate is <0.5% for 3 consecutive rounds, stop training and retain the optimal parameters.

[0076] After obtaining the asset intelligent analysis model, performance verification and optimization are carried out: ①Performance validation metric: Accuracy for classification tasks Recall rate F1 value The regression task is the mean absolute error. Coefficient of determination Common metrics include inference latency (≤200ms) and number of model parameters (≤10 million).

[0077] ② Validation criteria: For classification tasks, Acc ≥ 90%, Recall ≥ 88%, F1 ≥ 89%; for regression tasks, MAE ≤ 5% of the true value range, R... 2 ≥0.85. If the target is not met, return to the adjustment steps, add 200 more rounds of training iterations, or adjust the proportion of synthetic data in the mixed training dataset (increasing the proportion of real data to 6.0%), and retrain until the target is met.

[0078] Example 6: This embodiment further defines the specific content of the Flink real-time data acquisition module, the feature drift detection module, and step S5.

[0079] (i) The Flink real-time data acquisition module collects new real-time asset data from the target scene based on the Flink cluster, preprocesses and generates feature vectors to provide data for drift detection and model inference.

[0080] Technical Implementation: Flink cluster deployment: 1 master and 3 slave distributed architecture, JobManager memory 8GB, TaskManager memory per node 16GB, parallelism 32, and RocksDB is used for the state backend (checkpoint interval 5 minutes). Real-time data access: Access to Kafka and HTTP interfaces via FlinkSource, with data reception latency ≤1 second; Preprocessing flow: Heterogeneous data is converted to JSON format, and fields are named according to rules; outliers are removed using the 3σ principle; duplicate data is filtered; statistical and time-series features are calculated, and a real-time feature vector V is generated. real-time ; Data forwarding: Push the preprocessed feature vectors to Flink stream processing operators for invocation.

[0081] Input: New real-time asset data for the target scenario (Kafka / HTTP); Output: Real-time feature vectors in standardized JSON format (Flink streaming data).

[0082] (ii) The feature drift detection module uses the baseline subset of step S1 as a reference, calls EvidentlyAI to detect feature drift, and outputs a graded detection report.

[0083] Technical Implementation: EvidentlyAI Integration: Encapsulated as a FlinkUDF and embedded in ProcessWindowFunction, supporting integrated stream and batch detection; Detection configuration: sliding window size 5 minutes, step size 1 minute, minimum number of samples ≥ 100; different detection methods are used for different feature types; Drift determination: Based on the PSI value classification and combined with the results of auxiliary tests, a comprehensive determination is made. Report output: Generates a JSON detection report stored in Elasticsearch, which supports retrieval.

[0084] Input: Real-time feature vectors (Flink stream), real asset benchmark subset (HBase); Output: Feature drift detection report (Elasticsearch).

[0085] (III) Step S5: The process of continuously collecting new asset data based on Flink: Step S511: Deploy the FLink cluster; Cluster architecture: "1 master and 3 slaves" distributed architecture, JobManager memory ≥ 8GB, TaskManager memory per node ≥ 16GB, CPU cores ≥ 8, and supports dynamic adjustment of parallelism (maximum 32). State backend: Use RocksDB, configure an incremental checkpoint mechanism with an interval of 5 minutes, and retain the state for 24 hours to ensure that data is not lost and that failures are recoverable; Data receiving end: Adapts to multiple input sources via FlinkSource, supporting IoT sensors, real-time data from business systems, and RFID / OCR scanning data, each with different access methods and latency requirements. Time = 10 seconds; Step S512: Perform real-time data preprocessing, converting real-time data from different data sources into a unified JSON format, with field naming consistent with step S1; use the 3σ principle to remove outliers and filter duplicate data through FlinkProcessFunction; calculate newly added data features in real-time according to the rules of step S21, generating a real-time feature vector V. real - time .

[0086] Example 7: Based on the above, this example limits the model adaptability determination and automatic retraining module and the content of step S6.

[0087] (I) The model fit determination and automatic retraining module determines the fit based on the drift detection results. If the model fits, it inputs data for inference; otherwise, it retrains. The technical implementation is as follows: Adaptability determination: If there is no drift, the model is adapted and new data is pushed to the model; if there is slight drift, sampling verification is performed, and if the performance degradation is ≤3%, the model is adapted, otherwise it is not adapted; if there is severe drift, the model is not adapted and retraining is triggered.

[0088] Automatic retraining process: Construct a retraining set using 30% new real-time data and 70% of the original mixed training dataset, with a total sample size ≥ 1.2 times that of the original dataset; incremental training follows the original architecture with a learning rate of 50% of the original; freeze the general layer for the first 300 rounds, and unfreeze 30% of the joint training if the target is not met; after the target is met, switch to grayscale, and replace the entire dataset if there are no abnormalities.

[0089] Model invocation: The adapted model receives data for inference and outputs feature vectors to the scoring module.

[0090] Input: Feature drift detection report, new real-time data, asset intelligent analysis model; Output: Adapted model inference feature vector, retrained new model.

[0091] (ii) In step S5, using the baseline subset from step S1 as a reference, the EvidentlyAI tool is used to detect the drift process of newly added real-time data features: Step S521: Encapsulate EvidentlyAI as a FlinkUDF and embed it into Flink stream processing to support integrated stream and batch detection.

[0092] Step S522: Select the core features from step S21 as the detection object, and use a sliding window mechanism to assign indicators according to data type. Window size = 5 minutes (adaptively adjusted, minimum number of samples ≥ 100), step size = 1 minute.

[0093] Step S523: Use PSI value as the core threshold for comprehensive judgment Certainly.

[0094] Step S54: Output a report for each window and save it to Elasticsearch. The report supports retrieval and includes information such as asset type.

[0095] Added real-time data acquisition and drift detection anomaly handling: Data acquisition anomaly: An alarm was triggered for no data input from a certain data source for 30 seconds, and a backup source was activated; data integrity and timeliness anomaly flags were invalid, and the data was stored separately for review and supplementation.

[0096] Drift detection anomalies: Record feature change trends and generate hourly reports. 10 consecutive windows of slight drift escalate to "Pay Attention" status; severe drift triggers a Level 3 alarm, suspends inference, and retains a data snapshot. When the data source returns to normal and meets quality standards, or severe drift is manually confirmed to be normal (e.g., due to changes in asset conditions), the drift detection status is reset, model inference resumes, and abnormal data snapshots are retained for 90 days. The suitability of the asset intelligent analysis model is determined based on feature drift detection results: The group stability PSI value of core features is used as the primary detection condition, with Kolmogorov-Smirnov test, chi-square test, Anderson-Darling test, and augmented Dickey-Fuller test as auxiliary detection conditions. Feature drift levels are classified as no drift, slight drift, and severe drift. No drift indicates model suitability; slight drift requires model performance sampling verification before final determination; severe drift indicates model incompatibility, and the model is automatically retrained until it becomes suitable.

[0097] ① The PSI value quantifies the distribution difference between newly added and benchmark asset data on the same feature, reflecting the overall stability of the data distribution. It is a core indicator for feature drift detection, applicable to all core features, determines the drift level, and provides a core basis for model fit assessment. Standardized calculation formula. Involves feature binning, etc. Parameters. The PSI value is calculated only for core features. The drift classification standard is no drift PSI < 0.1, slight drift 0.1 ≤ PSI < 0.2, and severe drift PSI ≥ 0.2. The comprehensive judgment of multiple features requires all core features to meet the corresponding level requirements.

[0098] ② The Kolmogorov-Smirnov test (KS test) is a nonparametric test method used to verify whether the distribution pattern of newly added and baseline numerical asset data is consistent, assisting in the detection of numerical feature drift, and is used in conjunction with PSI. Standardized calculation formula Cumulative distribution function Etc. Applicable to all numerical core features. The test is performed by calculating the KS statistic and the p-value. P ≥ 0.05 is used to verify no drift, and P < 0.05 is used in conjunction with the PSI.

[0099] ③ The Anderson-Darling test (AD test) is an improved version of the KS test, more sensitive to differences in the tails of the data distribution. It is used to verify whether the distribution of newly added numerical asset data is the same as that of the baseline data, supplementing the KS test and suitable for in-depth verification of numerical features. Standardized calculation formula Number of samples involved Quantities, etc. Applicable to numerical core features with a clearly defined baseline distribution type. During testing, compare the AD statistic with the critical value. An AD statistic ≤ the critical value serves as a strong validation condition for drift-free behavior; an AD statistic > the critical value requires marking as "to be of interest".

[0100] ④ Augmented Dickey-Fuller Test (ADF test is a time series test method used to verify the stationarity of newly added time series asset data. It is also suitable for verifying the stability of time series features and avoiding model inference failure. Standardized calculation formula) It involves multiple parameters. It is applicable to time-series core features, and the null hypothesis is tested. H 0 ,pass t Test and judgment Significance: If P < 0.05, reject. H 0 The data is judged to be stationary, which serves as a specific verification condition for drift-free time series features; if P≥0.05, the time series features are judged to be drifting, triggering model fit verification, which can solve the problem of trend / seasonal drift detection and avoid increasing prediction error.

[0101] Chi-square test is a nonparametric test method used to verify whether the category distributions of newly added categorized asset data and baseline categorized data are independent. It is used in conjunction with PSI to verify the consistency of categorized feature distributions. Standardized calculation formula. Involving relevant parameters Applicable to all categorical core features, null hypothesis during testing H 0 (The two data categories are independently distributed), calculate After taking the statistics, look up the P-value in the table. If P ≥ 0.05, the category distribution is considered consistent, serving as an auxiliary verification condition for no drift. If P < 0.05, combine it with the PSI value to determine whether it has been upgraded to a slight drift, which can compensate for the PSI's insufficient capture of the detailed distribution of categorical features.

[0102]

[0103] The specific details of the drift level correlation determination are as follows: (1) No drift determination, direct fit: The core condition is PSI < 0.1; the auxiliary test standards are: for numerical features, KS test P ≥ 0.05, AD test statistic ≤ 95% confidence level critical value; for categorical features, chi-square test P ≥ 0.05, category proportion offset ≤ 3%; for time series features, ADF test P < 0.05, autocorrelation coefficient offset ≤ 0.05. If all core features meet the conditions, the model fit is directly determined without additional verification, generating a drift-free, direct fit conclusion.

[0104] (2) Slight drift determination requires performance verification: The core condition is that there are 1-3 core features that satisfy 0.1≤PSI<0.2, the remaining features have PSI<0.1, and only one auxiliary test fails (non-critical feature); the constraint is that the drifting feature is not a core decision feature of the model, otherwise it is upgraded to severe drift. The determination result requires the initiation of model performance sampling verification, and the suitability is determined based on the results.

[0105] Model performance sampling and validation process: The sampling rule is stratified random sampling, with a total sample size of 100 (if the amount of newly added real-time data is less than 100, it will be accumulated later). This covers 100% of the core asset types and various data categories in the target scenario. The sample percentage of each category deviates from the category distribution of the newly added real-time data by ≤5%. Each category includes ≥30 samples directly related to the drift characteristics. Two independent validation sets are formed by consecutive sampling twice. The validation metric is the same as in step S4, with accuracy for the classification task. The regression tasks include recall, F1 score, mean absolute error, and coefficient of determination. . (mean of the true values); If the difference in the same index between two independent validation sets is ≤1%, then resampling and validation should be performed.

[0106] Adaptation determination threshold and logic: Formula for calculating the magnitude of the decline: in: The optimal performance index value for model compliance; This represents the performance metric value for the current sampling verification. If all indicators of the two validation sets If the percentage is ≤3% and the performance stability verification passes, the system is considered compatible; if any metric in any validation set... If the percentage is >3%, or the performance stability check fails, the product is deemed unsuitable; the critical case is <2.5%. ≤3%: Add 1 new sampling validation set, all 3 validation sets now meet the requirement. A score of ≤3% is required to be considered compatible.

[0107] Special scenario adaptation judgment supplement For high-value assets (single asset value ≥ 1 million yuan) or key business assets (such as core equipment of the production line) in the target scenario, the matching judgment threshold will be strictly adjusted, i.e. The performance degradation threshold under slight drift is ≤2%; The sample size of the validation set has been increased to 200 records, covering 100% of core asset types and data categories; If the drift characteristic is a core decision characteristic of a critical asset, even if the PSI is 0.15 (slight drift), performance verification will be triggered directly, and if the verification fails, retraining will be started immediately.

[0108] (3) Severe drift determination, directly unsuitable: Core condition: There exists any core feature with PSI ≥ 0.2, or ≥ 4 core features satisfying 0.1 ≤ PSI < 0.2; Auxiliary test triggering conditions: ≥2 auxiliary tests fail to meet the standard, or the core decision feature shows drift (PSI≥0.1); Judgment result: The model is directly determined to be unsuitable, and the automatic retraining process is immediately triggered, while the drift trigger source is recorded.

[0109] Example 8: Based on the above embodiments, this embodiment further defines the specific content of the multi-dimensional asset scoring and decision suggestion module and the corresponding step S7.

[0110] (i) The multi-dimensional asset scoring and decision-making suggestion module outputs multi-dimensional asset scores based on the model reasoning results, classifies them into levels and generates corresponding decision-making suggestions to achieve intelligent dynamic management.

[0111] This is achieved through the following technologies: Multi-dimensional scoring calculation: Asset health status (40%), economic value (30%), operational risk (20%), and operational efficiency (10%), supporting weight adjustment by asset type (e.g., 50% weight for production equipment health dimension); positive indicators (compliance rate) are calculated according to... Calculation, reverse calculation of negative indicators (probability of failure), ; Grading: Overall score 90-100 (Grade A), 80-89 (Grade B), 60-79 (Grade C), 0-59 (Grade D); Decision recommendations are generated as follows: Level A: Maintain the current operation and maintenance strategy and conduct quarterly inspections; Level B: Optimize the frequency of operation and maintenance and monitor key indicators; Level C: Clarify the verification objects / time limits / responsible parties; Level D: Emergency response and loss mitigation plans. Recommended push notifications: Push notifications to different channels based on the level (Level A / B: Email; Level C: WeChat Work, Email; Level D: SMS, Phone, Pop-up).

[0112] Input: Feature vectors after model inference (from the adaptation module); Output: Asset rating report (including dimension rating, level, and decision recommendations), stored in PostgreSQL, and simultaneously pushed to the asset management visualization panel.

[0113] (ii) In step S7, the final asset intelligent analysis model outputs scores for the assets across different dimensions based on the input new asset data. The assets are then classified into four levels—A, B, C, and D—based on their scores from highest to lowest. The decision recommendations for each level are as follows: The recommended course of action for Level A is to maintain the current operational strategy and conduct routine inspections on a quarterly basis, without any additional intervention. The recommended strategy for Level B is to optimize the frequency of operations and maintenance, and to focus on monitoring fluctuations in key performance indicators. The recommended decision for Level C is to clearly define the objects to be verified, the time limit for verification, and the responsible party. The decision recommendations for Level D are emergency response measures and loss mitigation plans.

[0114] The specific scoring rules are as follows: (1) Definition of core scoring dimensions: Basic dimensions: Asset health status (weight 40%), asset economic value (weight 30%), asset operation and maintenance risk (weight 20%), and asset operation and maintenance efficiency (weight 10%). Dimensional adaptation adjustment: Adapt to different asset types.

[0115] (2) The evaluation indicators and data sources for each dimension are shown in Table 2 below: Table 2. Evaluation Indicators and Data Sources for Each Dimension

[0116] (4) Standardization requirements for the scoring system: Scoring range: All dimensions and the overall score are scored on a 100-point scale (0~100 points). The higher the score, the better the asset status. Indicator quantification rules: Each evaluation indicator must be converted into a quantitative value of 0 to 100 points; Weight determination criteria: Calculated using the analytic hierarchy process (AHP) and calibrated by scores from at least three industry experts, with a weight deviation of ≤5%.

[0117] (5) The scoring calculation adopts a three-level calculation rule of indicator quantification, dimension weighting, and comprehensive scoring, specifically including: Level 1: Quantitative Calculation of Evaluation Indicators Positive indicators, the higher the value, the better. The specific quantitative formula is as follows: in: This is the current indicator value; The minimum value of the indicator in the benchmark data; The maximum value of the indicator in the baseline data; Negative indicators are better the lower the value; the specific quantitative formula is as follows: ; Critical handling: If < (Positive indicator) or > (Negative indicator) =0; if > (Positive indicator) or < (Negative indicator) =100.

[0118] Level 2: Weighted Calculation of Dimensional Scores The specific calculation formula is as follows: ; Where: m is the number of evaluation indicators for this dimension; For the first i The weight of each indicator, i.e. ; Final score for a single dimension Level 3: Overall Score The calculation formula is: ; Where: n is the number of rating dimensions. For the first j The weights of each dimension, i.e. =1.

[0119] (6) The scoring levels are divided as shown in Table 3 below. Table 3. Rating Level Classification

[0120] Decision recommendations are tailored to asset type and drift status.

[0121] Output specifications for decision recommendations: Standardized JSON format, including "Asset ID, scoring dimension, dimension score, scoring level, decision recommendation content, recommendation execution time limit, responsible party, and generation timestamp"; when multiple dimension scoring levels are inconsistent, a comprehensive recommendation is generated according to the principle of "highest priority level prevails"; Push notifications based on priority: Low priority (A / B level): Push to the asset management platform message center and simultaneously send an email to the operations and maintenance manager; Medium priority (C level): Pushed to asset management platform pop-up, email, and WeChat work notification; if not read within 2 hours, a second reminder will be triggered. High priority (Level D): Push notifications to the asset management platform pop-up, WeChat / SMS, and telephone voice reminders. If no response is received within 15 minutes, an escalation alarm will be triggered.

[0122] The specific process of intelligent and dynamic management of asset information: The asset management platform constructs a visual monitoring panel to display the distribution of asset scores, the progress of high-priority recommendations, and a list of assets with abnormal scores related to feature drift in real time; the scores of A / B level assets are updated every hour, and the scores of C / D level assets are updated every minute to ensure that abnormal states are captured in real time. Establish a suggestion implementation log to record the suggestion receipt time, execution start time, execution completion time, execution result (meeting / not meeting the standard), and reason for not meeting the standard; if a C-level suggestion is not completed within the verification period, it will be automatically upgraded to a high-priority reminder; if a D-level suggestion does not receive a response within 30 minutes, it will be pushed to the company's top management.

[0123] The changes in asset scores after the recommendations are implemented will be used as the basis for feedback; the weights of the scoring indicators will be adjusted quarterly based on the feedback data, with an adjustment range of ≤10%; if the feedback finds that the scoring deviation is caused by model inference error, the model retraining process will be automatically triggered to optimize the inference accuracy. All asset rating records, decision recommendations, and execution ledgers are archived to a distributed file system (HDFS) by asset ID and time dimension and retained for 3 years; it supports multi-dimensional retrieval by rating level, time range, asset type, and recommended execution results, with a retrieval latency of ≤100ms, meeting the needs of auditing and management traceability.

[0124] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. An asset information management method based on big data and artificial intelligence, characterized in that, Includes the following steps: Step S1: Collect data throughout the entire asset lifecycle and preprocess the collected asset data to obtain a real asset dataset; Step S2: Based on real asset data, use AIGC technology to generate synthetic asset data with the same feature distribution as the real asset data; Step S3: Mix real asset data and synthetic asset data according to a preset ratio to construct a hybrid training dataset; Step S4: Using deep transfer learning technology, the pre-trained model trained in the same asset scenario is transferred to the target asset management scenario. Then, the pre-trained model after transfer is adjusted through the constructed hybrid training dataset to obtain the asset intelligent analysis model. Step S5: Based on the big data real-time computing framework Flink, continuously collect new asset data generated in real time in the target asset management scenario. Using the real asset data collected in Step S1 as a reference standard, call the EvidentlyAI tool to perform feature drift detection on the collected new real-time data. Step S6: Based on the feature drift detection results, determine the suitability of the obtained asset intelligent analysis model; if the asset intelligent analysis model meets the suitability requirements, directly input the newly added asset data into the asset intelligent analysis model. If the asset intelligent analysis model does not meet the adaptation requirements, the asset intelligent analysis model will be automatically retrained until it meets the adaptation requirements, and then the newly added asset data will be input into the asset intelligent analysis model. Step S7: The final intelligent asset analysis model outputs scores for the assets in different dimensions based on the input new asset data. Each dimension's score corresponds to a decision suggestion, thereby realizing intelligent and dynamic management of asset information.

2. The asset information management method based on big data and artificial intelligence according to claim 1, characterized in that, The specific process of step S1 includes the following steps: Step S11: Obtain asset data throughout its entire lifecycle through a multi-source acquisition terminal, specifically including historical existing asset data and initial real-time asset data; The multi-source acquisition terminal includes an RFID reader, an IoT sensor, an OCR scanning device, and a business system interface; Step S12: Clean, deduplicate, and standardize the acquired asset data to obtain a real asset dataset; Step S13: Split the real asset dataset into a training subset and a benchmark subset in a 7:3 ratio.

3. The asset information management method based on big data and artificial intelligence according to claim 2, characterized in that, The specific process of step S2 includes the following steps: Step S21: Extract the core features of real asset data from the training subset of the real asset dataset; Step S22: Use a generative adversarial network model, input the extracted core features for training, and generate initial synthetic asset data; Step S23: By using the feature consistency verification algorithm, compare the feature similarity between the initial synthetic asset data and the training subset data, calculate using the cosine similarity formula, and remove abnormal synthetic asset data, thus obtaining synthetic asset data that is consistent with the feature distribution of the training subset in the real asset data.

4. The asset information management method based on big data and artificial intelligence according to claim 3, characterized in that, In the hybrid training dataset constructed in step S3, the training subset accounts for 20% to 30% of the data, and the synthetic asset data accounts for 70% to 80% of the data.

5. The asset information management method based on big data and artificial intelligence according to claim 4, characterized in that, The pre-trained model in step S4 is a CNN-LSTM hybrid model or a lightweight Transformer model.

6. The asset information management method based on big data and artificial intelligence according to claim 5, characterized in that, In step S5, the specific process of continuously collecting newly generated asset data in the target asset management scenario based on the big data real-time computing framework Flink is as follows: Step S511: Deploy the FLink cluster; Step S512: Perform real-time data preprocessing: Convert real-time data from different data sources into JSON format, with field names consistent with the real asset data fields in Step S1; use the 3σ principle to remove outliers and filter duplicate data using FlinkProcessFunction; calculate the statistical and temporal features of the newly added data in real time according to the core feature extraction rules of Step S21, and generate a real-time feature vector V consistent with the baseline feature dimensions. real-time .

7. The asset information management method based on big data and artificial intelligence according to claim 6, characterized in that, In step S5, the specific process of using the baseline subset of real asset data collected in step S1 as a reference standard and calling the EvidentlyAI tool to perform feature drift detection on the newly collected real-time data is as follows: Step S521: Encapsulate EvidentlyAI as a FlinkUDF and embed it into the ProcessWindowFunction of FlinkStreamProcessing to support integrated stream and batch detection; Step S522: Select the core features extracted in step S21 as the drift detection object, and use a sliding window mechanism to configure detection indicators according to data type; Step S523: Using the PSI value as the core threshold, and combining it with other test results for comprehensive judgment, the PSI calculation formula is as follows: ； in, The proportion of the i-th bin in real-time data. The proportion of the i-th bin as the baseline data; Step S54: Output a detection report for each sliding window, and store the report in JSON format to Elasticsearch.

8. The asset information management method based on big data and artificial intelligence according to claim 7, characterized in that, In step S6, the specific details of determining the adaptability of the obtained asset intelligent analysis model based on the feature drift detection results are as follows: The feature drift detection results use the population stability PSI value of the core feature as the main detection condition, and the Kolmogorov-Smirnov test, chi-square test, Anderson-Darling test, and augmented Dickey-Fuller test as auxiliary detection conditions. The feature drift situation is divided into three levels: no drift, slight drift, and severe drift. Among them, no drift directly determines the compatibility of the obtained asset intelligent analysis model; For slight drift, it is necessary to initiate model performance sampling verification, and finally determine whether the obtained asset intelligent analysis model is suitable based on the verification results. If there is a severe drift, the obtained asset intelligent analysis model will be directly determined to be unsuitable, and the asset intelligent analysis model will be automatically retrained until the asset intelligent analysis model meets the suitability.

9. The asset information management method based on big data and artificial intelligence according to claim 8, characterized in that, In step S7, based on the input new asset data, the final asset intelligent analysis model outputs scores for the assets across different dimensions. The assets are then categorized into four levels—A, B, C, and D—based on these scores from highest to lowest. The corresponding decision recommendations for each level are as follows: The recommended course of action for Level A is to maintain the current operational strategy and conduct routine inspections on a quarterly basis, without any additional intervention. The recommended strategy for Level B is to optimize the frequency of operations and maintenance, and to focus on monitoring fluctuations in key performance indicators. The recommended decision for Level C is to clearly define the objects to be verified, the time limit for verification, and the responsible party. The decision recommendations for Level D are emergency response measures and loss mitigation plans.

10. An asset information management system based on big data and artificial intelligence, the system being capable of executing the asset information management method according to any one of claims 1 to 9, characterized in that, include: The multi-source asset data acquisition module collects historical stock data and initial real-time data throughout the entire life cycle of assets, and transmits the heterogeneous data in a unified manner after parsing it. The asset data preprocessing module cleans, deduplicatizes, and standardizes the collected raw asset data, splitting it into a training subset and a benchmark subset in a 7:3 ratio and storing them. The AIGC synthetic asset data generation module extracts core features from the training subset, generates initial synthetic asset data through a generative adversarial network, and removes anomalies through cosine similarity verification to obtain synthetic asset data with the same feature distribution as the real data. The mixed training dataset construction module constructs a mixed training dataset with a ratio of 20%~30% training subset and 70%~80% synthetic asset data, and stores it after splitting the training set and validation set; The deep transfer learning model training module loads a CNN-LSTM hybrid model or a lightweight Transformer pre-trained model for similar asset scenarios, fine-tunes it with the hybrid training dataset to obtain an asset intelligent analysis model and stores it. The Flink real-time data acquisition module collects new real-time asset data based on the Flink cluster, and generates real-time feature vectors through JSON standardization, 3σ outlier removal, deduplication, and core feature extraction. The feature drift detection module encapsulates EvidentlyAI into FlinkUDF. Using a benchmark subset as a reference, it performs drift classification detection by combining PSI values ​​with Kolmogorov-Smirnov tests, chi-square tests, etc., and outputs a JSON format detection report. The model fit determination and automatic retraining module determines the model fit based on the drift detection results. If the model fits, new asset data is pushed to the model. If the model does not fit, a retraining dataset is built by mixing 30% new data with 70% of the original dataset, and the model is automatically retrained until it fits. The multi-dimensional asset scoring and decision-making suggestion module receives model inference results, calculates multi-dimensional and comprehensive scores and classifies them into levels A to D, generates corresponding decision-making suggestions and pushes them according to priority. The system management and visualization module enables data archiving, access control, decision suggestion tracking, and system status monitoring and visualization, supporting collaborative operation across the entire process.

Images