A method and system for cross-line collaborative health management of an automobile stamping die

Abstract

The application provides a kind of across line collaborative automobile stamping die health management method and system, the method includes the following steps: collecting the running data of die;Perform anomaly detection to identify abnormal data;From the knowledge graph, retrieve the target historical case similar to the current abnormal feature;Based on the target historical case retrieved, generate a prediction model for the current anomaly;Generate maintenance parameter instructions through a collaborative decision engine and output execution instructions to related equipment;Integrate the feature data of abnormal data and maintenance parameter instructions into the knowledge graph, and update the knowledge base through incremental learning mechanism.By utilizing edge computing and knowledge graph retrieval, cross-line experience sharing is achieved, avoiding knowledge silos;Using meta-learning model to quickly adapt based on historical cases improves response speed and processing accuracy for abnormal data;Through the collaborative decision engine, maintenance instructions are automatically generated, and the knowledge base is dynamically updated to ensure continuous optimization of the system.

Description

Technical Field

[0001] This invention relates to the field of automotive stamping die management technology, and in particular to a cross-production line collaborative method and system for automotive stamping die health management. Background Technology

[0002] Automotive stamping dies are among the most critical process equipment in vehicle manufacturing, and their health directly determines the dimensional accuracy, surface quality, and production cycle time of stamped parts. As the automotive industry moves towards multi-variety, small-batch, and flexible production, the frequency and complexity of die usage are constantly increasing, making die health management increasingly important.

[0003] In existing technologies, mold health management solutions are typically based on single production lines or single factories. For example, they use vibration sensors and MES systems to predict mold lifespan, or utilize AI algorithms for fault diagnosis. While these technologies can achieve basic monitoring, they have significant shortcomings: First, the systems are limited to a single environment and lack cross-production line knowledge sharing mechanisms, resulting in the inability to transfer experience across different production lines, creating "knowledge silos" and increasing the cost of repeated trial and error. Second, existing models have poor generalization capabilities; when abnormal data patterns occur, extensive retraining with numerous samples is required, leading to delayed responses and an inability to quickly generate customized maintenance solutions. Finally, data integration is difficult; due to equipment heterogeneity and inconsistent protocols, knowledge updates rely on manual intervention, hindering continuous system evolution. These drawbacks limit the level of intelligence and economic benefits of mold health management. Summary of the Invention

[0004] In view of the shortcomings of the prior art, the purpose of this invention is to provide a cross-production line collaborative health management method and system for automotive stamping dies, aiming to solve the technical problem of low diagnostic efficiency in the prior art.

[0005] To achieve the above objectives, in a first aspect, the present invention provides: a cross-production line collaborative method for health management of automotive stamping dies, comprising the following steps:

[0006] The operation data of the mold is collected in real time by a multi-modal sensor array deployed on multiple stamping production lines;

[0007] The collected runtime data is preprocessed and features are extracted on the edge computing nodes, and anomaly detection is performed based on the extracted features to identify abnormal data.

[0008] When abnormal data is detected, cross-production line collaborative diagnosis is triggered, and historical cases with similar characteristics to the current abnormality are retrieved from the knowledge graph stored on the central server.

[0009] A meta-learning model is used to adapt the retrieved target historical cases to generate a predictive model for the current anomaly.

[0010] Based on the prediction model and the retrieved target historical cases, maintenance parameter instructions are generated through the collaborative decision engine, and execution instructions are output to the relevant devices.

[0011] The feature data and maintenance parameter instructions of the abnormal data are integrated into the knowledge graph, and the knowledge base is updated through an incremental learning mechanism.

[0012] According to one aspect of the above technical solution, the steps of preprocessing and feature extraction of the collected operational data specifically include:

[0013] The collected operational data is filtered and denoised, wavelet transform is used to remove high-frequency noise components, and time-domain features including the mean, variance, and peak factor of the vibration signal are extracted.

[0014] The frequency domain characteristics, including the dominant frequency and harmonic energy ratio, are calculated using Fast Fourier Transform. The dominant frequency corresponds to the dominant periodic component of the signal, and the harmonic energy ratio represents the energy ratio of the harmonic components to the fundamental frequency.

[0015] According to one aspect of the above technical solution, the steps for performing anomaly detection based on the extracted features specifically include:

[0016] Principal component analysis is used to perform dimensionality reduction and reconstruction on the feature vector including the time-domain features and the frequency-domain features to obtain the reconstructed feature vector. The reconstruction error is then calculated based on the following formula:

[0017] ;

[0018] ;

[0019] In the formula, The reconstruction error is... For feature vectors, For the reconstructed feature vector, , , These are the mean, variance, and peak factor of the vibration signal, respectively. , These are the main frequency and harmonic energy ratio, respectively.

[0020] If the reconstruction error is greater than a preset threshold, abnormal data identification is triggered.

[0021] According to one aspect of the above technical solution, the steps of adapting based on retrieved target historical cases using a meta-learning model specifically include:

[0022] Calculate the similarity distance between the current abnormal feature vector and the feature vector of historical cases in the knowledge graph, and select the historical case with the highest similarity as the target historical case based on the similarity distance;

[0023] Define the current anomaly feature vector and the feature vector of the target historical cases, and calculate the attention weight at each time point using the time point as the index;

[0024] The meta-learning model is trained using weighted target historical case data, and the model parameters are updated through gradient optimization.

[0025] According to one aspect of the above technical solution, the expression for calculating the attention weight at a given time point is:

[0026] ;

[0027] ;

[0028] In the formula, For attention weights, This is the current abnormal feature vector. The feature vector of the target historical case. The similarity score is given, where t is the index of the time point. Let be the weight coefficient of the k-th feature, k be the feature index, K be the total dimension of the feature vector, j be the summation index, and T be the number of time points in the target historical cases. The k-th component of the current abnormal feature vector. The k-th component of the feature vector of the target historical case at time point t;

[0029] The loss function for updating the meta-learning model is:

[0030] ;

[0031] In the formula, For loss function, For model parameters, For predicted values, Let i be the true value, i be the sample index, and n be the number of samples. Let be the weight of the i-th sample.

[0032] According to one aspect of the above technical solution, the steps of generating maintenance parameter instructions through the collaborative decision engine specifically include:

[0033] Based on the prediction results output by the prediction model, a preliminary maintenance parameter instruction set is obtained;

[0034] Retrieve the maintenance parameter instruction set and maintenance effect data of the target historical cases from the knowledge graph, and calculate the parameter adjustment amount based on the linear adjustment function of the historical maintenance effect and the maintenance effect data;

[0035] The initial maintenance parameter instruction set and parameter adjustment values ​​are modified to generate maintenance parameter instructions.

[0036] According to one aspect of the above technical solution, the steps of integrating the feature data and maintenance parameter instructions of abnormal data into the knowledge graph and updating the knowledge base through an incremental learning mechanism specifically include:

[0037] Extract the feature data and maintenance parameter instructions of the current abnormal data, and map them to the knowledge graph;

[0038] Create a new node corresponding to the current abnormal data, calculate the correlation degree with the current node based on the similarity distance, and calculate the knowledge update amount based on the correlation degree and the update weight factor;

[0039] Incremental learning algorithms are used to apply knowledge updates to the knowledge graph to adjust node weights and edge relationship strength, thereby updating the knowledge base.

[0040] According to one aspect of the above technical solution, the expression for calculating the correlation degree is:

[0041] ;

[0042] In the formula, A is the relevance degree, and D is the similarity distance;

[0043] The formula for calculating the updated knowledge amount is:

[0044] ;

[0045] ;

[0046] In the formula, To update knowledge, To update the weighting factors, n is the number of new cases, and N is the total number of cases in the knowledge base. For abnormal feature vectors, To maintain parameter commands, , The characteristic contribution coefficient.

[0047] Secondly, this solution also provides a cross-production line collaborative automotive stamping die health management system, including:

[0048] The data acquisition module is used to collect real-time operating data of the mold through a multi-modal sensor array deployed on multiple stamping production lines;

[0049] The identification module is used to preprocess and extract features from the collected runtime data on the edge computing node, and perform anomaly detection based on the extracted features to identify abnormal data.

[0050] The retrieval module is used to trigger cross-production line collaborative diagnosis when abnormal data is detected, and to retrieve target historical cases with similar characteristics to the current abnormality from the knowledge graph stored on the central server.

[0051] The adaptation module is used to adapt the retrieved target historical cases using a meta-learning model to generate a predictive model for the current anomaly.

[0052] The maintenance module is used to generate maintenance parameter instructions based on the prediction model and the retrieved target historical cases through a collaborative decision engine, and output execution instructions to relevant devices.

[0053] The update module is used to integrate the feature data and maintenance parameter instructions of abnormal data into the knowledge graph, and update the knowledge base through an incremental learning mechanism.

[0054] According to one aspect of the above technical solution, the identification module is specifically used for:

[0055] The collected operational data is filtered and denoised, wavelet transform is used to remove high-frequency noise components, and time-domain features including the mean, variance, and peak factor of the vibration signal are extracted.

[0056] The frequency domain characteristics, including the dominant frequency and harmonic energy ratio, are calculated using Fast Fourier Transform. The dominant frequency corresponds to the dominant periodic component of the signal, and the harmonic energy ratio represents the energy ratio of the harmonic components to the fundamental frequency.

[0057] According to one aspect of the above technical solution, the identification module is further used for:

[0058] Principal component analysis is used to perform dimensionality reduction and reconstruction on the feature vector including the time-domain features and the frequency-domain features to obtain the reconstructed feature vector. The reconstruction error is then calculated based on the following formula:

[0059] ;

[0060] ;

[0061] In the formula, The reconstruction error is... For feature vectors, For the reconstructed feature vector, , , These are the mean, variance, and peak factor of the vibration signal, respectively. , These are the main frequency and harmonic energy ratio, respectively.

[0062] If the reconstruction error is greater than a preset threshold, abnormal data identification is triggered.

[0063] According to one aspect of the above technical solution, the adapter module is specifically used for:

[0064] Calculate the similarity distance between the current abnormal feature vector and the feature vector of historical cases in the knowledge graph, and select the historical case with the highest similarity as the target historical case based on the similarity distance;

[0065] Define the current anomaly feature vector and the feature vector of the target historical cases, and calculate the attention weight at each time point using the time point as the index;

[0066] The meta-learning model is trained using weighted target historical case data, and the model parameters are updated through gradient optimization.

[0067] According to one aspect of the above technical solution, the maintenance module is specifically used for:

[0068] Based on the prediction results output by the prediction model, a preliminary maintenance parameter instruction set is obtained;

[0069] Retrieve the maintenance parameter instruction set and maintenance effect data of the target historical cases from the knowledge graph, and calculate the parameter adjustment amount based on the linear adjustment function of the historical maintenance effect and the maintenance effect data;

[0070] The initial maintenance parameter instruction set and parameter adjustment values ​​are modified to generate maintenance parameter instructions.

[0071] According to one aspect of the above technical solution, the updating module is specifically used for:

[0072] Extract the feature data and maintenance parameter instructions of the current abnormal data, and map them to the knowledge graph;

[0073] Create a new node corresponding to the current abnormal data, calculate the correlation degree with the current node based on the similarity distance, and calculate the knowledge update amount based on the correlation degree and the update weight factor;

[0074] Incremental learning algorithms are used to apply knowledge updates to the knowledge graph to adjust node weights and edge relationship strength, thereby updating the knowledge base.

[0075] Compared with existing technologies, the beneficial effects of this invention are as follows: by deploying a multimodal sensor array to achieve real-time acquisition of multi-source data, combined with edge computing and knowledge graph retrieval, it breaks through the limitations of a single production line, realizes cross-production line experience sharing, and avoids knowledge silos; by using a meta-learning model to quickly adapt based on historical cases, it can generate predictive models with a small number of samples, improving the response speed and processing accuracy of abnormal data; by automatically generating maintenance instructions through a collaborative decision engine and dynamically updating the knowledge base with the help of an incremental learning mechanism, it ensures continuous system optimization, thereby reducing downtime losses, lowering maintenance costs, and enhancing the adaptability and intelligence level of overall management. Attached Figure Description

[0076] Figure 1 This is a flowchart illustrating the cross-production line collaborative health management method for automotive stamping dies in the first embodiment of the present invention.

[0077] Figure 2 This is a structural block diagram of the cross-production line collaborative automotive stamping die health management system in the fourth embodiment of the present invention;

[0078] The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation

[0079] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Several embodiments of the invention are illustrated in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0080] It should be noted that when a component is said to be "fixed to" another component, it can be directly on the other component or there may be an intervening component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there may be an intervening component. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0081] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0082] Example 1

[0083] Please see Figure 1 The figure shows a flowchart of a cross-production line collaborative automotive stamping die health management method according to the first embodiment of the present invention. As shown in the figure, the method includes the following steps:

[0084] Step S100 involves collecting real-time mold operation data using a multimodal sensor array deployed on multiple stamping production lines. Specifically, in this embodiment, a multimodal sensor array is deployed at key mold locations (such as guide pillars, guide sleeves, and mold cavities) on three stamping production lines in factory A and two stamping production lines in factory B. The sensors include vibration sensors, acoustic sensors, and temperature sensors.

[0085] Step S200 involves preprocessing and extracting features from the collected runtime data at the edge computing node, and performing anomaly detection based on the extracted features to identify abnormal data. Specifically, the steps of preprocessing and extracting features from the collected runtime data include:

[0086] The collected operational data is filtered and denoised, wavelet transform is used to remove high-frequency noise components, and time-domain features including the mean, variance, and peak factor of the vibration signal are extracted.

[0087] The frequency domain characteristics, including the dominant frequency and harmonic energy ratio, are calculated using Fast Fourier Transform. The dominant frequency corresponds to the dominant periodic component of the signal, and the harmonic energy ratio represents the energy ratio of the harmonic components to the fundamental frequency.

[0088] The specific steps for performing anomaly detection based on extracted features include:

[0089] Principal component analysis is used to perform dimensionality reduction and reconstruction on the feature vector including the time-domain features and the frequency-domain features to obtain the reconstructed feature vector. The reconstruction error is then calculated based on the following formula:

[0090] ;

[0091] ;

[0092] In the formula, The reconstruction error is... For feature vectors, For the reconstructed feature vector, , , These are the mean, variance, and peak factor of the vibration signal, respectively. , These are the main frequency and harmonic energy ratio, respectively.

[0093] If the reconstruction error exceeds a preset threshold, abnormal data identification is triggered. The threshold is set to 0.1 (based on historical data calibration). If E > 0.1, abnormal data pattern identification is triggered. For example, a vibration signal is detected on production line 2 of factory B. If the frequency changes to 800Hz and E=0.15, it is identified as abnormal data.

[0094] Step S300: When abnormal data is detected, cross-production line collaborative diagnosis is triggered, and a target historical case with similar characteristics to the current abnormality is retrieved from the knowledge graph stored on the central server. Edge nodes send the abnormal feature vector (e.g., data [0.5, 0.02, 4.2, 800, 0.4]) to the central server (cloud server). The knowledge graph is built on the Neo4j database and contains historical case nodes (e.g., case ID, feature data, maintenance records) for factories A and B. The retrieval algorithm uses the Euclidean norm to calculate the similarity distance between the abnormal data features and the feature vectors of all historical cases in the graph. A similarity threshold of 0.5 is set, and cases with D < 0.5 are selected as target historical cases. For example, the retrieval finds a similar case in factory B (data [0.48, 0.019, 4.0, 790, 0.38]), whose abnormality was caused by guide post wear, and the repair solution was "replace the guide post and adjust the lubrication".

[0095] Step S400: Use a meta-learning model to adapt the retrieved target historical cases to generate a prediction model for the current anomaly.

[0096] Preferably, the step of adapting based on retrieved target historical cases using a meta-learning model specifically includes:

[0097] Calculate the similarity distance between the current abnormal feature vector and the feature vector of historical cases in the knowledge graph, and select the historical case with the highest similarity as the target historical case based on the similarity distance;

[0098] Define the current anomaly feature vector and the feature vector of the target historical cases, and calculate the attention weight at each time point using the time point as the index;

[0099] The meta-learning model is trained using weighted target historical case data, and the model parameters are updated through gradient optimization. The meta-learning model employs an optimization-based MAML (Model-Agnostic Meta-Learning) algorithm and is deployed on a central server. Input data includes the current anomaly feature vector and time-series data of target historical cases (e.g., 100 time points of vibration signals in historical cases).

[0100] Furthermore, the expression for calculating the attention weights at each time point is:

[0101] ;

[0102] ;

[0103] In the formula, For attention weights, This is the current abnormal feature vector. The feature vector of the target historical case. The similarity score is given, where t is the index of the time point. Let be the weight coefficient of the k-th feature, k be the feature index, K be the total dimension of the feature vector, j be the summation index, and T be the number of time points in the target historical cases. The k-th component of the current abnormal feature vector. The k-th component of the feature vector of the target historical case at time point t;

[0104] The loss function for updating the meta-learning model is:

[0105] ;

[0106] In the formula, For loss function, For model parameters, For predicted values, Let i be the true value, i be the sample index, and n be the number of samples. Let be the weight of the i-th sample.

[0107] In some application scenarios of this embodiment, firstly, the attention weight is calculated: for each time point t, the similarity score is: （ Feature dimension K=5, weight coefficient (Obtained through historical training). Attention weights (T = 100 time points). Then, a meta-learning model is trained using weighted historical data, with the loss function... (Sample weights) (Based on the severity setting of the anomaly). The model parameter θ is optimized through gradient descent and converges quickly within 10 samples to generate a predictive model (e.g., outputting the remaining lifespan of the guide post wear as 50 stamping cycles).

[0108] Step S500: Based on the prediction model and the retrieved target historical cases, maintenance parameter instructions are generated through the collaborative decision engine, and execution instructions are output to the relevant devices.

[0109] Preferably, the step of generating maintenance parameter instructions through the collaborative decision engine specifically includes:

[0110] Based on the prediction results output by the prediction model, a preliminary maintenance parameter instruction set is obtained;

[0111] Retrieve the maintenance parameter instruction set and maintenance effect data of the target historical cases from the knowledge graph, and calculate the parameter adjustment amount based on the linear adjustment function of the historical maintenance effect and the maintenance effect data;

[0112] The initial maintenance parameter instruction set and parameter adjustment values ​​are modified to generate maintenance parameter instructions.

[0113] In some application scenarios of this embodiment, the collaborative decision-making engine runs on a central server. First, a preliminary maintenance parameter set is obtained from the prediction model. (e.g., lubrication interval = 100 stamping cycles). Then, retrieve the maintenance parameter set of the target historical cases from the knowledge graph. (e.g., lubrication interval = 120 times) and maintenance effect data (For example, actual wear rate = 0.01 mm / thousand cycles). Calculate parameter adjustment amount. Where k is the adjustment factor (k=0.8 based on...) (Calculation). Optimize and maintain the parameter set. (e.g., adjust the lubrication interval to 92 times). Finally, generate maintenance parameter instructions (e.g., "reduce the lubrication interval to 92 times, check the guide post").

[0114] Step S600: Integrate the feature data and maintenance parameter instructions of the abnormal data into the knowledge graph, and update the knowledge base through an incremental learning mechanism.

[0115] Preferably, the step of integrating the feature data and maintenance parameter instructions of abnormal data into the knowledge graph and updating the knowledge base through an incremental learning mechanism specifically includes:

[0116] Extract the feature data and maintenance parameter instructions of the current abnormal data, and map them to the knowledge graph;

[0117] Create a new node corresponding to the current abnormal data, calculate the correlation degree with the current node based on the similarity distance, and calculate the knowledge update amount based on the correlation degree and the update weight factor;

[0118] Incremental learning algorithms are used to apply knowledge updates to the knowledge graph to adjust node weights and edge relationship strength, thereby updating the knowledge base.

[0119] The expression for calculating the degree of association is as follows:

[0120] ;

[0121] In the formula, A is the relevance degree, and D is the similarity distance;

[0122] The formula for calculating the updated knowledge amount is:

[0123] ;

[0124] ;

[0125] In the formula, To update knowledge, To update the weighting factors, n is the number of new cases, and N is the total number of cases in the knowledge base. For abnormal feature vectors, To maintain parameter commands, , The characteristic contribution coefficient.

[0126] Specifically, in some application scenarios of this implementation, =0.6, =0.4, through incremental learning algorithm (online gradient descent) When applied to a knowledge graph, node weights are adjusted (e.g., a new node's weight is set to 1.0, and the weight of related nodes is increased by 0.1) and relationship strength is updated (i.e., edge strength is updated). Finally, a consistency check is performed to remove redundant nodes and ensure the knowledge base is updated.

[0127] In summary, the cross-production line collaborative automotive stamping die health management method in the above embodiments of the present invention achieves real-time multi-source data acquisition by deploying a multi-modal sensor array, and combines edge computing and knowledge graph retrieval to overcome the limitations of a single production line, realize cross-production line experience sharing, and avoid knowledge silos; it utilizes a meta-learning model to quickly adapt based on historical cases, and can generate predictive models with a small number of samples, improving the response speed and processing accuracy of abnormal data; it automatically generates maintenance instructions through a collaborative decision engine, and dynamically updates the knowledge base with the help of an incremental learning mechanism to ensure continuous system optimization, thereby reducing downtime losses, lowering maintenance costs, and enhancing the adaptability and intelligence level of overall management.

[0128] Example 2

[0129] A second embodiment of this application also provides a cross-production line collaborative automotive stamping die health management system. This system is used to implement the embodiments and preferred embodiments described herein, and will not be repeated hereafter. As used below, the terms "module," "unit," "subunit," etc., can refer to a combination of software and / or hardware that performs a predetermined function. Although the system described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0130] like Figure 2 As shown, the system includes: a data acquisition module 100, an identification module 200, a retrieval module 300, an adaptation module 400, a maintenance module 500, and an update module 600.

[0131] The acquisition module 100 is used to acquire the mold operation data in real time through a multi-modal sensor array deployed on multiple stamping production lines;

[0132] The identification module 200 is used to preprocess and extract features from the collected running data on the edge computing node, and to perform anomaly detection based on the extracted features to identify abnormal data.

[0133] The retrieval module 300 is used to trigger cross-production line collaborative diagnosis when abnormal data is detected, and to retrieve target historical cases with similar characteristics to the current abnormality from the knowledge graph stored in the central server.

[0134] The adaptation module 400 is used to adapt based on retrieved target historical cases using a meta-learning model to generate a predictive model for the current anomaly.

[0135] The maintenance module 500 is used to generate maintenance parameter instructions based on the prediction model and the retrieved target historical cases through a collaborative decision engine, and output execution instructions to relevant devices;

[0136] The update module 600 is used to integrate the feature data and maintenance parameter instructions of abnormal data into the knowledge graph, and update the knowledge base through an incremental learning mechanism.

[0137] Preferably, in this embodiment, the identification module 200 is specifically used for:

[0138] The collected operational data is filtered and denoised, wavelet transform is used to remove high-frequency noise components, and time-domain features including the mean, variance, and peak factor of the vibration signal are extracted.

[0139] The frequency domain characteristics, including the dominant frequency and harmonic energy ratio, are calculated using Fast Fourier Transform. The dominant frequency corresponds to the dominant periodic component of the signal, and the harmonic energy ratio represents the energy ratio of the harmonic components to the fundamental frequency.

[0140] Preferably, in this embodiment, the identification module 200 is further configured to:

[0141] Principal component analysis is used to perform dimensionality reduction and reconstruction on the feature vector including the time-domain features and the frequency-domain features to obtain the reconstructed feature vector. The reconstruction error is then calculated based on the following formula:

[0142] ;

[0143] ;

[0144] In the formula, The reconstruction error is... For feature vectors, For the reconstructed feature vector, , , These are the mean, variance, and peak factor of the vibration signal, respectively. , These are the main frequency and harmonic energy ratio, respectively.

[0145] If the reconstruction error is greater than a preset threshold, abnormal data identification is triggered.

[0146] Preferably, in this embodiment, the adapter module 400 is specifically used for:

[0147] Calculate the similarity distance between the current abnormal feature vector and the feature vector of historical cases in the knowledge graph, and select the historical case with the highest similarity as the target historical case based on the similarity distance;

[0148] Define the current anomaly feature vector and the feature vector of the target historical cases, and calculate the attention weight at each time point using the time point as the index;

[0149] The meta-learning model is trained using weighted target historical case data, and the model parameters are updated through gradient optimization.

[0150] Preferably, in this embodiment, the maintenance module 500 is specifically used for:

[0151] Based on the prediction results output by the prediction model, a preliminary maintenance parameter instruction set is obtained;

[0152] Retrieve the maintenance parameter instruction set and maintenance effect data of the target historical cases from the knowledge graph, and calculate the parameter adjustment amount based on the linear adjustment function of the historical maintenance effect and the maintenance effect data;

[0153] The initial maintenance parameter instruction set and parameter adjustment values ​​are modified to generate maintenance parameter instructions.

[0154] Preferably, in this embodiment, the update module 600 is specifically used for:

[0155] Extract the feature data and maintenance parameter instructions of the current abnormal data, and map them to the knowledge graph;

[0156] Create a new node corresponding to the current abnormal data, calculate the correlation degree with the current node based on the similarity distance, and calculate the knowledge update amount based on the correlation degree and the update weight factor;

[0157] Incremental learning algorithms are used to apply knowledge updates to the knowledge graph to adjust node weights and edge relationship strength, thereby updating the knowledge base.

[0158] It should be noted that the modules can be functional modules or program modules, and can be implemented in software or hardware. For modules implemented in hardware, the modules can reside in the same processor; or the modules can be located in different processors in any combination.

[0159] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A cross-production line collaborative method for health management of automotive stamping dies, characterized in that, Includes the following steps: The operation data of the mold is collected in real time by a multi-modal sensor array deployed on multiple stamping production lines; The collected runtime data is preprocessed and features are extracted on the edge computing nodes, and anomaly detection is performed based on the extracted features to identify abnormal data. When abnormal data is detected, cross-production line collaborative diagnosis is triggered, and historical cases with similar characteristics to the current abnormality are retrieved from the knowledge graph stored on the central server. A meta-learning model is used to adapt the model to retrieved historical target cases, generating a predictive model for the current anomaly. The specific steps of adapting the model to retrieved historical target cases include: Calculate the similarity distance between the current abnormal feature vector and the feature vector of historical cases in the knowledge graph, and select the historical case with the highest similarity as the target historical case based on the similarity distance; Define the current anomaly feature vector and the feature vector of the target historical cases, and calculate the attention weight at each time point using the time point as the index; The meta-learning model is trained using weighted target historical case data, and the model parameters are updated through gradient optimization. The expression for calculating the attention weights at each time point is: ； ； In the formula, For attention weights, This is the current abnormal feature vector. The feature vector of the target historical case. The similarity score is given, where t is the index of the time point. Let be the weight coefficient of the k-th feature, k be the feature index, K be the total dimension of the feature vector, j be the summation index, and T be the number of time points in the target historical cases. The k-th component of the current abnormal feature vector. The k-th component of the feature vector of the target historical case at time point t; Based on the prediction model and the retrieved target historical cases, maintenance parameter instructions are generated through a collaborative decision engine, and execution instructions are output to the relevant devices. Specifically, the step of generating maintenance parameter instructions through the collaborative decision engine includes: Based on the prediction results output by the prediction model, a preliminary maintenance parameter instruction set is obtained; Retrieve the maintenance parameter instruction set and maintenance effect data of the target historical cases from the knowledge graph, and calculate the parameter adjustment amount based on the linear adjustment function of the historical maintenance effect and the maintenance effect data; The initial maintenance parameter instruction set and parameter adjustment values ​​are corrected to generate maintenance parameter instructions; The feature data and maintenance parameter instructions of the abnormal data are integrated into the knowledge graph, and the knowledge base is updated through an incremental learning mechanism.

2. The cross-production line collaborative health management method for automotive stamping dies according to claim 1, characterized in that, The specific steps for preprocessing and feature extraction of the collected operational data include: The collected operational data is filtered and denoised, wavelet transform is used to remove high-frequency noise components, and time-domain features including the mean, variance, and peak factor of the vibration signal are extracted. The frequency domain characteristics, including the dominant frequency and harmonic energy ratio, are calculated using Fast Fourier Transform. The dominant frequency corresponds to the dominant periodic component of the signal, and the harmonic energy ratio represents the energy ratio of the harmonic components to the fundamental frequency.

3. The cross-production line collaborative health management method for automotive stamping dies according to claim 2, characterized in that, The specific steps for performing anomaly detection based on extracted features include: Principal component analysis is used to perform dimensionality reduction and reconstruction on the feature vector including the time-domain features and the frequency-domain features to obtain the reconstructed feature vector. The reconstruction error is then calculated based on the following formula: ； ； In the formula, The reconstruction error is... For feature vectors, For the reconstructed feature vector, , , These are the mean, variance, and peak factor of the vibration signal, respectively. , These are the main frequency and harmonic energy ratio, respectively. If the reconstruction error is greater than a preset threshold, abnormal data identification is triggered.

4. The cross-production line collaborative health management method for automotive stamping dies according to claim 1, characterized in that, The loss function for updating the meta-learning model is: ； In the formula, For loss function, For model parameters, For predicted values, Let i be the true value, i be the sample index, and n be the number of samples. Let be the weight of the i-th sample.

5. The cross-production line collaborative health management method for automotive stamping dies according to claim 1, characterized in that, The steps of integrating the feature data and maintenance parameter instructions of abnormal data into the knowledge graph and updating the knowledge base through an incremental learning mechanism specifically include: Extract the feature data and maintenance parameter instructions of the current abnormal data, and map them to the knowledge graph; Create a new node corresponding to the current abnormal data, calculate the correlation degree with the current node based on the similarity distance, and calculate the knowledge update amount based on the correlation degree and the update weight factor; Incremental learning algorithms are used to apply knowledge updates to the knowledge graph to adjust node weights and edge relationship strength, thereby updating the knowledge base.

6. The cross-production line collaborative health management method for automotive stamping dies according to claim 5, characterized in that, The expression for calculating the degree of association is: ； In the formula, A is the relevance degree, and D is the similarity distance; The formula for calculating the updated knowledge amount is: ； ； In the formula, To update knowledge, To update the weighting factors, n is the number of new cases, and N is the total number of cases in the knowledge base. For abnormal feature vectors, To maintain parameter commands, , The characteristic contribution coefficient.

7. A cross-production-line collaborative automotive stamping die health management system for implementing the cross-production-line collaborative automotive stamping die health management method according to any one of claims 1-6, characterized in that, include: The data acquisition module is used to collect real-time operating data of the mold through a multi-modal sensor array deployed on multiple stamping production lines; The identification module is used to preprocess and extract features from the collected runtime data on the edge computing node, and perform anomaly detection based on the extracted features to identify abnormal data. The retrieval module is used to trigger cross-production line collaborative diagnosis when abnormal data is detected, and to retrieve target historical cases with similar characteristics to the current abnormality from the knowledge graph stored on the central server. The adaptation module is used to adapt the retrieved target historical cases using a meta-learning model to generate a predictive model for the current anomaly. The maintenance module is used to generate maintenance parameter instructions based on the prediction model and the retrieved target historical cases through a collaborative decision engine, and output execution instructions to relevant devices. The update module is used to integrate the feature data and maintenance parameter instructions of abnormal data into the knowledge graph, and update the knowledge base through an incremental learning mechanism.

8. The cross-production line collaborative automotive stamping die health management system according to claim 7, characterized in that, The identification module is specifically used for: The collected operational data is filtered and denoised, wavelet transform is used to remove high-frequency noise components, and time-domain features including the mean, variance, and peak factor of the vibration signal are extracted. The frequency domain characteristics, including the dominant frequency and harmonic energy ratio, are calculated using Fast Fourier Transform. The dominant frequency corresponds to the dominant periodic component of the signal, and the harmonic energy ratio represents the energy ratio of the harmonic components to the fundamental frequency.

Images