Intelligent early warning analysis platform based on management and control of product life cycle

Abstract

The invention discloses an intelligent early warning analysis platform based on management and control of a product life cycle, and the platform comprises a data collection and preprocessing module which is used for collecting multi-source heterogeneous data and carrying out the preprocessing of the data, and obtaining a standardized input data set; the spectral clustering grouping module is used for carrying out unsupervised grouping by adopting a spectral clustering algorithm and generating a grouping label and a mapping result; the abnormal sample counting and generating module is used for counting the number of abnormal samples and data distribution of each data sub-group and generating abnormal sample data matched with sub-group characteristics through a stable diffusion model; the training and risk detection module is used for constructing a training data set and training a risk identification and anomaly detection method; and the intelligent early warning and response module is used for automatically triggering early warning for the detected abnormal data and linking platform management and control. According to the method, intelligent grouping, abnormal sample enhancement and efficient risk early warning analysis of product full-life-cycle multi-source heterogeneous data are realized.

Description

Technical Field

[0001] This invention relates to the field of industrial internet technology, and in particular to an intelligent early warning and analysis platform based on the management and control of the product lifecycle. Background Technology

[0002] In modern industrial manufacturing and product lifecycle management, with the rapid development of technologies such as intelligent manufacturing, the Industrial Internet, and big data analytics, enterprises have accumulated massive amounts of multi-source heterogeneous data. This data covers multiple stages, including product design, production, quality inspection, supply chain, operation and maintenance, and after-sales service. To achieve risk control and intelligent early warning throughout the product lifecycle, existing technologies widely employ methods such as machine learning, deep learning, and cluster analysis for data mining and anomaly detection. Common practices include using traditional clustering algorithms or single supervised models to perform pattern recognition and anomaly detection on historical data, or manually setting rules to complete risk assessments for certain lifecycle nodes. However, these approaches have several limitations in practical applications: Firstly, the complex structure of multi-source heterogeneous data makes it difficult for traditional clustering algorithms to adapt to the dynamic clustering and high-dimensional feature representation of data from different stages and batches, leading to inaccurate identification of anomaly category sample distribution. Secondly, high-risk anomalies in the industrial sector generally suffer from extremely scarce samples, making it difficult to effectively train a single model. Generative data augmentation methods lack deep coupling with lifecycle characteristics, failing to achieve high-quality anomaly sample expansion. At the same time, existing risk identification and intelligent early warning systems are generally lagging behind in the dynamic detection and response to abnormal data, making it difficult to achieve real-time automated alarms and closed-loop management.

[0003] Existing technologies cannot simultaneously address data clustering, abnormal sample expansion, and efficient risk warning at different stages of the product lifecycle.

[0004] Therefore, how to provide an intelligent early warning and analysis platform based on the management and control of the product lifecycle is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent early warning analysis platform based on the product lifecycle management system. This invention fully leverages the unsupervised clustering capabilities of spectral clustering and the high-quality anomaly sample generation advantages of stable diffusion models. It details a fully automated algorithm scheme for the entire process, from data acquisition and preprocessing, dynamic clustering, anomaly sample augmentation, to risk identification and early warning judgment. Through multi-source heterogeneous data standardization, accurate clustering, adaptive data augmentation in situations where high-risk anomaly samples are scarce, and intelligent identification and closed-loop management of risk states, this invention possesses advantages such as strong data adaptability, high anomaly detection capability, timely early warning response, and high system integration. It can significantly improve the effectiveness of intelligent early warning analysis throughout the product lifecycle and enhance the level of industrial intelligent management.

[0006] According to an embodiment of the present invention, a smart early warning and analysis platform based on the controlled product lifecycle includes:

[0007] The data acquisition and preprocessing module is used to collect multi-source heterogeneous data throughout the product lifecycle and preprocess it to obtain a standardized input dataset.

[0008] The spectral clustering module is used to perform unsupervised clustering based on a standardized input dataset using the spectral clustering algorithm, identifying data subgroups under different stages, batches, or categories, and generating cluster labels and mapping results.

[0009] The abnormal sample statistics and generation module is used to count the number of abnormal samples and data distribution of each data subgroup. For data subgroups with scarce and high-risk abnormal samples, feature vectors are extracted and used as conditional inputs. An improved stable diffusion model is used to generate abnormal sample data that matches the characteristics of the subgroup.

[0010] The training and risk detection module is used to integrate standardized input datasets, cluster labels, clustering mapping results, and abnormal sample data to build a training dataset and train risk identification and anomaly detection methods.

[0011] The intelligent early warning and response module is used to perform spectral clustering discrimination, abnormal sample generation and intelligent early warning analysis when the platform is running. It automatically triggers early warnings for detected abnormal data and links with platform management to complete alarm, risk response and event tracing.

[0012] Optionally, modules can be integrated using the following methods:

[0013] S1. Collect multi-source heterogeneous data at each stage of the product's entire lifecycle, and preprocess the multi-source heterogeneous data to form a standardized input dataset;

[0014] S2. Based on the standardized input dataset, the spectral clustering algorithm is used for unsupervised clustering to automatically identify data subgroups under different life cycle stages, process batches or product categories, and generate clustering labels and clustering mapping results.

[0015] S3. For each data subgroup, count the number of abnormal samples and data distribution characteristics of the data subgroup. For data subgroups with scarce abnormal samples and high-risk characteristics, extract feature vectors as conditional inputs, and use the improved stable diffusion model to generate abnormal samples and data augmentation to obtain abnormal sample data that matches the characteristics of the subgroup.

[0016] S4. Integrate the standardized input dataset, cluster labels and clustering mapping results, and abnormal sample data to construct a training dataset. Based on the training dataset, train the risk identification and anomaly detection methods to realize risk feature modeling, anomaly detection, and early warning judgment at different stages of the product life cycle.

[0017] S5. During the actual operation of the platform, based on the newly collected multi-source heterogeneous data, steps S1 to S4 are executed cyclically to achieve spectral clustering discrimination, abnormal sample generation and intelligent early warning analysis. For detected abnormal data, early warning is automatically triggered, and the platform management process is linked to complete alarm, risk response and event tracing.

[0018] Optionally, the multi-source heterogeneous data specifically includes product design data, manufacturing data, quality inspection data, supply chain logistics data, equipment operation data, sensor monitoring data, and after-sales service data.

[0019] Optionally, the preprocessing of multi-source heterogeneous data specifically includes data cleaning, data normalization, missing value imputation, feature selection, and feature construction.

[0020] Optionally, S2 specifically includes:

[0021] S21. Normalize the various data features in the standardized input dataset, and set weights for each feature according to the data type and business requirements to form a weighted normalized feature dataset.

[0022] S22. Using the weighted normalized feature dataset, calculate the similarity between any two data points, specifically as follows:

[0023] For each pair of data points, first calculate the difference between the two points on each feature, then square each difference and multiply it by the feature's weight coefficient. Finally, sum the weighted squared differences of all features to obtain the total weighted distance.

[0024] Then, the total weighted distance is correlated with the width parameter of the kernel function, and the weighted distance is converted into a similarity value through an exponential function. Finally, a similarity matrix is ​​formed between all data points. The larger the similarity value, the more similar the two points are, and the smaller the value, the less similar the two points are.

[0025] S23. Based on the similarity matrix, calculate the degree value of each data point, that is, add the similarity values ​​of each data point with all other data points to obtain the degree matrix;

[0026] S24. Analyze the distribution density and sample size of each data subgroup in the weighted normalized feature dataset. If the sample density differs significantly between different subgroups, choose the symmetric normalized Laplace matrix method, specifically:

[0027] Based on the degree matrix and similarity matrix, a symmetric normalized Laplacian matrix is ​​calculated to reduce the influence of different densities on the clustering effect.

[0028] If the data subgroups are relatively evenly distributed, then the non-normalized Laplace matrix method is chosen, specifically:

[0029] By directly subtracting the degree matrix from the similarity matrix, the non-normalized Laplacian matrix is ​​calculated, preserving the overall structural information, and obtaining a Laplacian matrix that can accurately reflect the actual data clustering relationship;

[0030] S25. Perform eigenvalue decomposition on the Laplacian matrix, select several eigenvectors with the smallest eigenvalues, and arrange the eigenvectors in columns to form an eigenvector matrix. The number of eigenvectors is equal to the preset number of cluster categories.

[0031] S26. Normalize each row of the eigenvector matrix, ensuring that the vector length of each data point is consistent, to obtain the normalized eigenvector matrix.

[0032] S27. Each row of the normalized feature vector matrix is ​​used as the representation of the data point in the new feature space. An unsupervised clustering algorithm is used to cluster all data points. Based on the cluster labels, all data points are divided into several data subgroups. Each cluster category corresponds to a data subgroup. The data subgroup to which each data point belongs is determined. The cluster labels and cluster mapping results of all data points are output to obtain the member set of each data subgroup.

[0033] Optionally, S3 specifically includes:

[0034] S31. For each data subgroup, count the number of abnormal samples within the data subgroup and analyze the data distribution of each feature dimension. Based on the number of abnormal samples and the feature distribution results, classify all data subgroups into risk levels to form risk level classification information including high-risk level and ordinary risk level.

[0035] S32. Extract weighted normalized feature vectors from data subgroups that are classified as high-risk and have scarce abnormal samples.

[0036] S33. Construct an improved stable diffusion model, wherein the stable diffusion model includes:

[0037] The locally deformable diffusion block consists of a local feature selection unit, a deformation amplitude control unit, and an adaptive noise addition unit. The local feature selection unit determines the feature regions that need to be deformed based on risk weights and anomaly distributions. The deformation amplitude control unit dynamically adjusts the deformation intensity of each region. The adaptive noise addition unit allocates noise disturbance intensity to different regions.

[0038] The reverse consistency reconstruction path consists of a forward feature mapping unit and a reverse reconstruction unit. The forward feature mapping unit realizes the mapping from generated samples to feature space, and the reverse reconstruction unit realizes the restoration from feature space to original data space. The correspondence between the forward and reverse processes is constrained by consistency loss.

[0039] The hierarchical risk-oriented noise modulation module consists of a risk level discrimination unit and a multi-level noise control unit. The risk level discrimination unit identifies the risk level of each data subgroup, and the multi-level noise control unit allocates noise perturbations of corresponding intensity and structure to different subgroups according to the risk level.

[0040] S34. In the forward diffusion process of the stable diffusion model, local deformable diffusion blocks are used to dynamically allocate deformation amplitude and noise disturbance intensity to different regions or feature dimensions of the input features, thereby improving the stable diffusion model's ability to capture local anomalous structures and simulate diverse anomalies.

[0041] S35. In the reverse diffusion process of the stable diffusion model, the reverse consistency reconstruction path is used to perform bidirectional mapping between the current data and the weighted normalized feature vector in the diffusion generation process in the feature space and the sample space, and optimization is performed through consistency loss constraints.

[0042] S36. In the noise injection stage, the hierarchical risk-oriented noise modulation module is used to inject differentiated noise into data subgroups of different risk levels according to the subgroup labels, subgroup mapping results and risk level classification information. The high-risk subgroups are subjected to noise perturbation with larger amplitude or more complex structure, while the ordinary risk subgroups are subjected to standard noise perturbation.

[0043] S37. Perform forward diffusion and reverse denoising processes to generate abnormal sample data that matches high-risk, scarce abnormal feature subgroups. Screen the validity of the generated abnormal sample data. The screening criteria include the consistency in distribution and feature diversity with the existing abnormal samples in each data subgroup.

[0044] Optionally, S4 specifically includes:

[0045] S41. Integrate the standardized input dataset cluster labels and clustering mapping results with outlier sample data to form a training dataset;

[0046] S42. Perform feature encoding processing on each sample in the training dataset. Concatenate and normalize the standardized input dataset, cluster labels and abnormal sample data according to the category to obtain the feature vector. Assign a category label or warning status label to each sample to form a training feature matrix and label vector.

[0047] S43. Input the training feature matrix and label vector into the risk identification and anomaly detection module. The risk identification and anomaly detection module includes a feature input unit, a feature representation unit, a risk discrimination unit, and an anomaly identification unit. The feature input unit receives the feature vector, the feature representation unit processes and maps the input features, the risk discrimination unit judges the risk status of the sample, and the anomaly identification unit determines whether the sample is an abnormal sample data.

[0048] S44. Train the risk identification and anomaly detection module by batch iteratively inputting training samples, compare the output risk discrimination results and anomaly identification results with the label vector, adjust the internal parameters according to the comparison results, until the accuracy of risk discrimination and anomaly identification reaches the preset requirements, and complete the training of the risk identification and anomaly detection module.

[0049] S45. For new input samples at different stages of the product lifecycle, perform feature encoding consistent with the training samples, input the feature vector of the new sample into the trained risk identification and anomaly detection module, and obtain the risk judgment result and anomaly identification result of the new sample.

[0050] Optionally, S5 specifically includes:

[0051] S51. During the actual operation of the platform, continuously collect multi-source heterogeneous data from each stage of the product's entire lifecycle, and preprocess it to form a standardized input dataset.

[0052] S52. For the standardized input dataset, perform spectral clustering, cluster label generation, and cluster mapping result generation in sequence, and perform unsupervised clustering and cluster mapping.

[0053] S53. Based on the new clustering results, the number of abnormal samples and data distribution of each data subgroup are counted. For high-risk data subgroups with scarce abnormal samples, an improved stable diffusion model is used to generate abnormal samples and augment data to obtain new abnormal sample data.

[0054] S54. Integrate the newly collected standardized input dataset, the latest cluster labels and clustering mapping results, and the generated abnormal sample data to form data input for risk identification and anomaly detection, perform intelligent early warning analysis, and output risk judgment and anomaly detection results;

[0055] S55. For abnormal data detected during the intelligent early warning analysis process, the platform will automatically trigger an early warning, link the platform management and control process, and complete the alarm, risk response and event tracing.

[0056] The beneficial effects of this invention are:

[0057] This invention establishes an intelligent early warning and analysis platform covering the entire product lifecycle by integrating spectral clustering algorithms and stable diffusion models, effectively overcoming the shortcomings of existing technologies in dynamic data clustering, abnormal sample expansion, and real-time risk warning. Specifically, this invention can automatically identify data subgroups at different stages, processes, or batches for multi-source, heterogeneous lifecycle data. Furthermore, in high-risk scenarios with scarce abnormal samples, it generates high-quality abnormal sample data through an improved stable diffusion model, significantly enhancing the system's ability to detect rare risk events. Simultaneously, through integrated risk identification and abnormal detection method modules, the entire process data is modeled and trained, enabling real-time and accurate abnormal detection and early warning of newly collected data. Through the platform's automated cyclic execution, once abnormal data is detected, it immediately triggers the control system to complete alarms, risk responses, and event tracing, achieving closed-loop monitoring and adaptive optimization of the entire product lifecycle in the field of intelligent manufacturing. Compared with traditional solutions, this invention not only improves the accuracy and response efficiency of risk warnings but also enhances the system's adaptability and scalability to complex industrial scenarios, improving the overall level of product quality control and industrial intelligent management. Attached Figure Description

[0058] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0059] Figure 1 This is a schematic diagram of the structure of the intelligent early warning and analysis platform based on the product lifecycle management proposed in this invention;

[0060] Figure 2 This is a flowchart illustrating the intelligent early warning analysis method based on the product lifecycle management proposed in this invention. Detailed Implementation

[0061] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0062] refer to Figure 1 Based on the intelligent early warning and analysis platform for managing the product lifecycle, including:

[0063] The data acquisition and preprocessing module is used to collect multi-source heterogeneous data throughout the product lifecycle and preprocess it to obtain a standardized input dataset.

[0064] The spectral clustering module is used to perform unsupervised clustering based on a standardized input dataset using the spectral clustering algorithm, identifying data subgroups under different stages, batches, or categories, and generating cluster labels and mapping results.

[0065] The abnormal sample statistics and generation module is used to count the number of abnormal samples and data distribution of each data subgroup. For data subgroups with scarce and high-risk abnormal samples, feature vectors are extracted and used as conditional inputs. An improved stable diffusion model is used to generate abnormal sample data that matches the characteristics of the subgroup.

[0066] The training and risk detection module is used to integrate standardized input datasets, cluster labels, clustering mapping results, and abnormal sample data to build a training dataset and train risk identification and anomaly detection methods.

[0067] The intelligent early warning and response module is used to perform spectral clustering discrimination, abnormal sample generation and intelligent early warning analysis when the platform is running. It automatically triggers early warnings for detected abnormal data and links with platform management to complete alarm, risk response and event tracing.

[0068] refer to Figure 2 The intelligent early warning analysis method based on the product lifecycle management includes the following steps:

[0069] S1. Collect multi-source heterogeneous data at each stage of the product's entire lifecycle, and preprocess the multi-source heterogeneous data to form a standardized input dataset;

[0070] S2. Based on the standardized input dataset, the spectral clustering algorithm is used for unsupervised clustering to automatically identify data subgroups under different life cycle stages, process batches or product categories, and generate clustering labels and clustering mapping results.

[0071] S3. For each data subgroup, count the number of abnormal samples and data distribution characteristics of the data subgroup. For data subgroups with scarce abnormal samples and high-risk characteristics, extract feature vectors as conditional inputs, and use the improved stable diffusion model to generate abnormal samples and data augmentation to obtain abnormal sample data that matches the characteristics of the subgroup.

[0072] S4. Integrate the standardized input dataset, cluster labels and clustering mapping results, and abnormal sample data to construct a training dataset. Based on the training dataset, train the risk identification and anomaly detection methods to realize risk feature modeling, anomaly detection, and early warning judgment at different stages of the product life cycle.

[0073] S5. During the actual operation of the platform, based on the newly collected multi-source heterogeneous data, steps S1 to S4 are executed cyclically to achieve spectral clustering discrimination, abnormal sample generation and intelligent early warning analysis. For detected abnormal data, early warning is automatically triggered, and the platform management process is linked to complete alarm, risk response and event tracing.

[0074] In this embodiment, the multi-source heterogeneous data specifically includes product design data, manufacturing data, quality inspection data, supply chain logistics data, equipment operation data, sensor monitoring data, and after-sales service data.

[0075] In this embodiment, the preprocessing of multi-source heterogeneous data specifically includes data cleaning, data normalization, missing value imputation, feature selection, and feature construction.

[0076] In this embodiment, S2 specifically includes:

[0077] S21. Normalize the various data features in the standardized input dataset, and set weights for each feature according to the data type and business requirements to form a weighted normalized feature dataset.

[0078] S22. Using the weighted normalized feature dataset, calculate the similarity between any two data points, specifically as follows:

[0079] For each pair of data points, first calculate the difference between the two points on each feature, then square each difference and multiply it by the feature's weight coefficient. Finally, sum the weighted squared differences of all features to obtain the total weighted distance.

[0080] Then, the total weighted distance is correlated with the width parameter of the kernel function, and the weighted distance is converted into a similarity value through an exponential function. Finally, a similarity matrix is ​​formed between all data points. The larger the similarity value, the more similar the two points are, and the smaller the value, the less similar the two points are.

[0081] S23. Based on the similarity matrix, calculate the degree value of each data point, that is, add the similarity values ​​of each data point with all other data points to obtain the degree matrix;

[0082] S24. Analyze the distribution density and sample size of each data subgroup in the weighted normalized feature dataset. If the sample density differs significantly between different subgroups, choose the symmetric normalized Laplace matrix method, specifically:

[0083] Based on the degree matrix and similarity matrix, a symmetric normalized Laplacian matrix is ​​calculated to reduce the influence of different densities on the clustering effect.

[0084] If the data subgroups are relatively evenly distributed, then the non-normalized Laplace matrix method is chosen, specifically:

[0085] By directly subtracting the degree matrix from the similarity matrix, the non-normalized Laplacian matrix is ​​calculated, preserving the overall structural information, and obtaining a Laplacian matrix that can accurately reflect the actual data clustering relationship;

[0086] S25. Perform eigenvalue decomposition on the Laplacian matrix, select several eigenvectors with the smallest eigenvalues, and arrange the eigenvectors in columns to form an eigenvector matrix. The number of eigenvectors is equal to the preset number of cluster categories.

[0087] S26. Normalize each row of the eigenvector matrix, ensuring that the vector length of each data point is consistent, to obtain the normalized eigenvector matrix.

[0088] S27. Each row of the normalized feature vector matrix is ​​used as the representation of the data point in the new feature space. An unsupervised clustering algorithm is used to cluster all data points. Based on the cluster labels, all data points are divided into several data subgroups. Each cluster category corresponds to a data subgroup. The data subgroup to which each data point belongs is determined. The cluster labels and cluster mapping results of all data points are output to obtain the member set of each data subgroup.

[0089] In this embodiment, S3 specifically includes:

[0090] S31. For each data subgroup, count the number of abnormal samples within the data subgroup and analyze the data distribution of each feature dimension. Based on the number of abnormal samples and the feature distribution results, classify all data subgroups into risk levels to form risk level classification information including high-risk level and ordinary risk level.

[0091] S32. Extract weighted normalized feature vectors from data subgroups that are classified as high-risk and have scarce abnormal samples.

[0092] S33. Construct an improved stable diffusion model, wherein the stable diffusion model includes:

[0093] The locally deformable diffusion block consists of a local feature selection unit, a deformation amplitude control unit, and an adaptive noise addition unit. The local feature selection unit determines the feature regions that need to be deformed based on risk weights and anomaly distributions. The deformation amplitude control unit dynamically adjusts the deformation intensity of each region. The adaptive noise addition unit allocates noise disturbance intensity to different regions.

[0094] The reverse consistency reconstruction path consists of a forward feature mapping unit and a reverse reconstruction unit. The forward feature mapping unit realizes the mapping from generated samples to feature space, and the reverse reconstruction unit realizes the restoration from feature space to original data space. The correspondence between the forward and reverse processes is constrained by consistency loss.

[0095] The hierarchical risk-oriented noise modulation module consists of a risk level discrimination unit and a multi-level noise control unit. The risk level discrimination unit identifies the risk level of each data subgroup, and the multi-level noise control unit allocates noise perturbations of corresponding intensity and structure to different subgroups according to the risk level.

[0096] S34. In the forward diffusion process of the stable diffusion model, local deformable diffusion blocks are used to dynamically allocate deformation amplitude and noise disturbance intensity to different regions or feature dimensions of the input features, thereby improving the stable diffusion model's ability to capture local anomalous structures and simulate diverse anomalies.

[0097] S35. In the reverse diffusion process of the stable diffusion model, the reverse consistency reconstruction path is used to perform bidirectional mapping between the current data and the weighted normalized feature vector in the diffusion generation process in the feature space and the sample space, and optimization is performed through consistency loss constraints.

[0098] S36. In the noise injection stage, the hierarchical risk-oriented noise modulation module is used to inject differentiated noise into data subgroups of different risk levels according to the subgroup labels, subgroup mapping results and risk level classification information. The high-risk subgroups are subjected to noise perturbation with larger amplitude or more complex structure, while the ordinary risk subgroups are subjected to standard noise perturbation.

[0099] S37. Perform forward diffusion and reverse denoising processes to generate abnormal sample data that matches high-risk, scarce abnormal feature subgroups. Screen the validity of the generated abnormal sample data. The screening criteria include the consistency in distribution and feature diversity with the existing abnormal samples in each data subgroup.

[0100] In this embodiment, S4 specifically includes:

[0101] S41. Integrate the standardized input dataset cluster labels and clustering mapping results with outlier sample data to form a training dataset;

[0102] S42. Perform feature encoding processing on each sample in the training dataset. Concatenate and normalize the standardized input dataset, cluster labels and abnormal sample data according to the category to obtain the feature vector. Assign a category label or warning status label to each sample to form a training feature matrix and label vector.

[0103] S43. Input the training feature matrix and label vector into the risk identification and anomaly detection module. The risk identification and anomaly detection module includes a feature input unit, a feature representation unit, a risk discrimination unit, and an anomaly identification unit. The feature input unit receives the feature vector, the feature representation unit processes and maps the input features, the risk discrimination unit judges the risk status of the sample, and the anomaly identification unit determines whether the sample is an abnormal sample data.

[0104] S44. Train the risk identification and anomaly detection module by batch iteratively inputting training samples, compare the output risk discrimination results and anomaly identification results with the label vector, adjust the internal parameters according to the comparison results, until the accuracy of risk discrimination and anomaly identification reaches the preset requirements, and complete the training of the risk identification and anomaly detection module.

[0105] S45. For new input samples at different stages of the product lifecycle, perform feature encoding consistent with the training samples, input the feature vector of the new sample into the trained risk identification and anomaly detection module, and obtain the risk judgment result and anomaly identification result of the new sample.

[0106] In this embodiment, S5 specifically includes:

[0107] S51. During the actual operation of the platform, continuously collect multi-source heterogeneous data from each stage of the product's entire lifecycle, and preprocess it to form a standardized input dataset.

[0108] S52. For the standardized input dataset, perform spectral clustering, cluster label generation, and cluster mapping result generation in sequence, and perform unsupervised clustering and cluster mapping.

[0109] S53. Based on the new clustering results, the number of abnormal samples and data distribution of each data subgroup are counted. For high-risk data subgroups with scarce abnormal samples, an improved stable diffusion model is used to generate abnormal samples and augment data to obtain new abnormal sample data.

[0110] S54. Integrate the newly collected standardized input dataset, the latest cluster labels and clustering mapping results, and the generated abnormal sample data to form data input for risk identification and anomaly detection, perform intelligent early warning analysis, and output risk judgment and anomaly detection results;

[0111] S55. For abnormal data detected during the intelligent early warning analysis process, the platform will automatically trigger an early warning, link the platform management and control process, and complete the alarm, risk response and event tracing.

[0112] Example 1:

[0113] To verify the feasibility of this invention in practice, it was applied to the production site of a smart equipment manufacturing company. The company's production line covers the entire process of designing, producing, assembling, testing, delivering, and after-sales service of high-end CNC machine tools. Due to the long manufacturing cycle, numerous stages, and diverse equipment types of this type of equipment, the company has accumulated a vast amount of multi-source heterogeneous data, including production process parameters, equipment sensor signals, quality inspection reports, and operation and maintenance logs. Traditional data analysis and risk monitoring methods typically rely on static rules or single supervised learning models, which are difficult to adapt to complex and dynamically changing data structures. Especially when abnormal samples are extremely scarce in certain stages and data clusters are mixed, the system often fails to provide timely warnings, posing a risk of missed quality incidents and slow response.

[0114] To address the aforementioned issues, the company piloted the product lifecycle intelligent early warning analysis platform proposed in this invention, based on the deep integration of spectral clustering algorithm and stable diffusion model, starting in March 2025. The platform first automatically collects data from each stage through a data acquisition and preprocessing module, capable of processing 120,000 new data entries daily. For existing quality inspection data and maintenance records, the platform standardizes and completes missing data to obtain a high-quality standardized input dataset. In the spectral clustering module, the platform performs unsupervised clustering of each batch of product data, automatically identifying data subgroups at different stages such as assembly, factory inspection, and customer maintenance. For example, in April 2025, the factory inspection data of a single machine tool was divided into five data subgroups, each aggregating specific process or component characteristics, effectively improving the data's structure and usability.

[0115] In practical risk management, the platform's anomaly sample statistics and generation module counts the number of anomalies for each data subgroup. Taking the CNC machine tool spindle anomaly subgroup as an example, only four high-risk failures have occurred historically, resulting in only 10 related samples, insufficient to support deep model training. Based on this, the platform utilizes an improved stable diffusion model, employing locally deformable diffusion blocks and a risk-oriented noise modulation module to quickly generate 60 high-quality, structurally diverse spindle anomaly samples. After review by quality control engineers, over 80% of the generated anomaly samples accurately reflect real failure scenarios, significantly alleviating the problem of anomaly sample scarcity.

[0116] Subsequently, the platform integrated standardized input data, cluster labels, and enhanced anomaly samples to construct a brand-new training dataset. The risk identification and anomaly detection module, through multiple rounds of training, achieved accurate identification of high-risk states and anomalous events.

[0117] Table 1 Comparison of Key Indicators Before and After the Implementation of the Intelligent Early Warning Platform

[0118]

[0119] As can be seen from the comparison of the data in Table 1, the solution of this invention has achieved significant results in practical applications. Firstly, the platform's data processing capacity has been greatly improved, with the average daily processing volume increasing from 70,000 records in the traditional solution to 120,000 records, an increase of 71.4%, reflecting the system's efficient access and processing capabilities for massive amounts of multi-source data. Secondly, the failure rate of major equipment decreased from 2.1% to 1.35%, a reduction of 35.7%, indicating that the platform has significantly reduced the probability of sudden major failures in terms of full-process risk monitoring and anomaly early warning.

[0120] The average warning response time was significantly reduced from 14.7 seconds to 3.9 seconds, a decrease of 73.5%, achieving near real-time response to risk events. Anomaly detection accuracy improved from 84.5% to 97.1%, an increase of 12.6 percentage points. This means the system is more accurate in identifying anomalies in complex scenarios, with significantly reduced false negatives and missed positives. In scenarios where anomalous samples are extremely scarce in the main axis subgroup, the number of high-quality anomaly samples generated by this invention increased from 10 to 70, a six-fold increase, mitigating the impact of insufficient data on model training and enhancing the ability to detect rare anomalies.

[0121] The first-pass yield rate increased from 93.4% to 98.6%, demonstrating that the platform effectively promoted product quality improvement through precise risk management. The average downtime decreased from 35.6 minutes per unit per month to 20.3 minutes, a reduction of 43.0%, reflecting the platform's early warning and response mechanism's role in ensuring production continuity. The early warning rate for high-risk anomalies increased from 25% to 86%, an increase of 61 percentage points, fully demonstrating the outstanding advantages of this invention in preventing major risks and ensuring production safety.

[0122] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A product lifecycle intelligent early warning and analysis platform, characterized in that: include: The data acquisition and preprocessing module is used to collect multi-source heterogeneous data throughout the product lifecycle and preprocess it to obtain a standardized input dataset. The spectral clustering module is used to perform unsupervised clustering based on a standardized input dataset using the spectral clustering algorithm, identifying data subgroups under different stages, batches, or categories, and generating cluster labels and mapping results. The abnormal sample statistics and generation module is used to count the number of abnormal samples and data distribution of each data subgroup. For data subgroups with scarce and high-risk abnormal samples, feature vectors are extracted and used as conditional inputs. An improved stable diffusion model is used to generate abnormal sample data that matches the characteristics of the subgroup. The training and risk detection module is used to integrate standardized input datasets, cluster labels, clustering mapping results, and abnormal sample data to build a training dataset and train risk identification and anomaly detection methods. The intelligent early warning and response module is used to perform spectral clustering discrimination, abnormal sample generation and intelligent early warning analysis when the platform is running. It automatically triggers early warnings for detected abnormal data and links with platform management to complete alarm, risk response and event tracing.

2. A method for intelligent early warning analysis based on the product lifecycle of controlled products, applied to the intelligent early warning analysis platform based on the product lifecycle of controlled products as described in claim 1, characterized in that, Includes the following steps: S1. Collect multi-source heterogeneous data at each stage of the product's entire lifecycle, and preprocess the multi-source heterogeneous data to form a standardized input dataset; S2. Based on the standardized input dataset, the spectral clustering algorithm is used for unsupervised clustering to automatically identify data subgroups under different life cycle stages, process batches or product categories, and generate clustering labels and clustering mapping results. S3. For each data subgroup, count the number of abnormal samples and data distribution characteristics of the data subgroup. For data subgroups with scarce abnormal samples and high-risk characteristics, extract feature vectors as conditional inputs, and use the improved stable diffusion model to generate abnormal samples and data augmentation to obtain abnormal sample data that matches the characteristics of the subgroup. S4. Integrate the standardized input dataset, cluster labels and clustering mapping results, and abnormal sample data to construct a training dataset. Based on the training dataset, train the risk identification and anomaly detection methods to realize risk feature modeling, anomaly detection, and early warning judgment at different stages of the product life cycle. S5. During the actual operation of the platform, based on the newly collected multi-source heterogeneous data, steps S1 to S4 are executed cyclically to achieve spectral clustering discrimination, abnormal sample generation and intelligent early warning analysis. For detected abnormal data, early warning is automatically triggered, and the platform management process is linked to complete alarm, risk response and event tracing.

3. The intelligent early warning analysis method based on the product lifecycle management system according to claim 2, characterized in that, The multi-source heterogeneous data specifically includes product design data, manufacturing data, quality inspection data, supply chain logistics data, equipment operation data, sensor monitoring data, and after-sales service data.

4. The intelligent early warning analysis method based on the product lifecycle management system according to claim 2, characterized in that, The preprocessing of multi-source heterogeneous data specifically includes data cleaning, data normalization, missing value imputation, feature selection, and feature construction.

5. The intelligent early warning analysis method based on the product lifecycle management system according to claim 2, characterized in that, S2 specifically includes: S21. Normalize the various data features in the standardized input dataset, and set weights for each feature according to the data type and business requirements to form a weighted normalized feature dataset. S22. Using the weighted normalized feature dataset, calculate the similarity between any two data points, specifically as follows: For each pair of data points, first calculate the difference between the two points on each feature, then square each difference and multiply it by the feature's weight coefficient. Finally, sum the weighted squared differences of all features to obtain the total weighted distance. Then, the total weighted distance is correlated with the width parameter of the kernel function, and the weighted distance is converted into a similarity value through an exponential function. Finally, a similarity matrix is ​​formed between all data points. The larger the similarity value, the more similar the two points are, and the smaller the value, the less similar the two points are. S23. Based on the similarity matrix, calculate the degree value of each data point, that is, add the similarity values ​​of each data point with all other data points to obtain the degree matrix; S24. Analyze the distribution density and sample size of each data subgroup in the weighted normalized feature dataset. If the sample density differs significantly between different subgroups, choose the symmetric normalized Laplace matrix method, specifically: Based on the degree matrix and similarity matrix, a symmetric normalized Laplacian matrix is ​​calculated to reduce the influence of different densities on the clustering effect. If the data subgroups are relatively evenly distributed, then the non-normalized Laplace matrix method is chosen, specifically: By directly subtracting the degree matrix from the similarity matrix, the non-normalized Laplacian matrix is ​​calculated, preserving the overall structural information, and obtaining a Laplacian matrix that can accurately reflect the actual data clustering relationship; S25. Perform eigenvalue decomposition on the Laplacian matrix, select several eigenvectors with the smallest eigenvalues, and arrange the eigenvectors in columns to form an eigenvector matrix. The number of eigenvectors is equal to the preset number of cluster categories. S26. Normalize each row of the eigenvector matrix, ensuring that the vector length of each data point is consistent, to obtain the normalized eigenvector matrix. S27. Each row of the normalized feature vector matrix is ​​used as the representation of the data point in the new feature space. An unsupervised clustering algorithm is used to cluster all data points. Based on the cluster labels, all data points are divided into several data subgroups. Each cluster category corresponds to a data subgroup. The data subgroup to which each data point belongs is determined. The cluster labels and cluster mapping results of all data points are output to obtain the member set of each data subgroup.

6. The intelligent early warning analysis method based on the product lifecycle management system according to claim 2, characterized in that, S3 specifically includes: S31. For each data subgroup, count the number of abnormal samples within the data subgroup and analyze the data distribution of each feature dimension. Based on the number of abnormal samples and the feature distribution results, classify all data subgroups into risk levels to form risk level classification information including high-risk level and ordinary risk level. S32. Extract weighted normalized feature vectors from data subgroups that are classified as high-risk and have scarce abnormal samples. S33. Construct an improved stable diffusion model, wherein the stable diffusion model includes: The locally deformable diffusion block consists of a local feature selection unit, a deformation amplitude control unit, and an adaptive noise addition unit. The local feature selection unit determines the feature regions that need to be deformed based on risk weights and anomaly distributions. The deformation amplitude control unit dynamically adjusts the deformation intensity of each region. The adaptive noise addition unit allocates noise disturbance intensity to different regions. The reverse consistency reconstruction path consists of a forward feature mapping unit and a reverse reconstruction unit. The forward feature mapping unit realizes the mapping from generated samples to feature space, and the reverse reconstruction unit realizes the restoration from feature space to original data space. The correspondence between the forward and reverse processes is constrained by consistency loss. The hierarchical risk-oriented noise modulation module consists of a risk level discrimination unit and a multi-level noise control unit. The risk level discrimination unit identifies the risk level of each data subgroup, and the multi-level noise control unit allocates noise perturbations of corresponding intensity and structure to different subgroups according to the risk level. S34. In the forward diffusion process of the stable diffusion model, local deformable diffusion blocks are used to dynamically allocate deformation amplitude and noise disturbance intensity to different regions or feature dimensions of the input features, thereby improving the stable diffusion model's ability to capture local anomalous structures and simulate diverse anomalies. S35. In the reverse diffusion process of the stable diffusion model, the reverse consistency reconstruction path is used to perform bidirectional mapping between the current data and the weighted normalized feature vector in the diffusion generation process in the feature space and the sample space, and optimization is performed through consistency loss constraints. S36. In the noise injection stage, the hierarchical risk-oriented noise modulation module is used to inject differentiated noise into data subgroups of different risk levels according to the subgroup labels, subgroup mapping results and risk level classification information. The high-risk subgroups are subjected to noise perturbation with larger amplitude or more complex structure, while the ordinary risk subgroups are subjected to standard noise perturbation. S37. Perform forward diffusion and reverse denoising processes to generate abnormal sample data that matches high-risk, scarce abnormal feature subgroups. Screen the validity of the generated abnormal sample data. The screening criteria include the consistency in distribution and feature diversity with the existing abnormal samples in each data subgroup.

7. The intelligent early warning analysis method based on the product lifecycle management system according to claim 2, characterized in that, S4 specifically includes: S41. Integrate the standardized input dataset cluster labels and clustering mapping results with outlier sample data to form a training dataset; S42. Perform feature encoding processing on each sample in the training dataset. Concatenate and normalize the standardized input dataset, cluster labels and abnormal sample data according to the category to obtain the feature vector. Assign a category label or warning status label to each sample to form a training feature matrix and label vector. S43. Input the training feature matrix and label vector into the risk identification and anomaly detection module. The risk identification and anomaly detection module includes a feature input unit, a feature representation unit, a risk discrimination unit, and an anomaly identification unit. The feature input unit receives the feature vector, the feature representation unit processes and maps the input features, the risk discrimination unit judges the risk status of the sample, and the anomaly identification unit determines whether the sample is an abnormal sample data. S44. Train the risk identification and anomaly detection module by batch iteratively inputting training samples, compare the output risk discrimination results and anomaly identification results with the label vector, adjust the internal parameters according to the comparison results, until the accuracy of risk discrimination and anomaly identification reaches the preset requirements, and complete the training of the risk identification and anomaly detection module. S45. For new input samples at different stages of the product lifecycle, perform feature encoding consistent with the training samples, input the feature vector of the new sample into the trained risk identification and anomaly detection module, and obtain the risk judgment result and anomaly identification result of the new sample.

8. The intelligent early warning analysis method based on the product lifecycle management system according to claim 2, characterized in that, S5 specifically includes: S51. During the actual operation of the platform, continuously collect multi-source heterogeneous data from each stage of the product's entire lifecycle, and preprocess it to form a standardized input dataset. S52. For the standardized input dataset, perform spectral clustering, cluster label generation, and cluster mapping result generation in sequence, and perform unsupervised clustering and cluster mapping. S53. Based on the new clustering results, the number of abnormal samples and data distribution of each data subgroup are counted. For high-risk data subgroups with scarce abnormal samples, an improved stable diffusion model is used to generate abnormal samples and augment data to obtain new abnormal sample data. S54. Integrate the newly collected standardized input dataset, the latest cluster labels and clustering mapping results, and the generated abnormal sample data to form data input for risk identification and anomaly detection, perform intelligent early warning analysis, and output risk judgment and anomaly detection results; S55. For abnormal data detected during the intelligent early warning analysis process, the platform will automatically trigger an early warning, link the platform management and control process, and complete the alarm, risk response and event tracing.

Images