A mold management full life cycle state machine flow transfer method

Abstract

This invention provides a method for state machine transition throughout the entire lifecycle of mold management, belonging to the field of production mold technology. This invention maps a unified semantic set of end-to-end parameters to a three-dimensional topological space through manifold dimensionality reduction and divides state regions. The three-dimensional spatial coordinate set and historical state sequences are input into an artificial intelligence cumulative damage state recognition model constructed based on a sparse connection learning framework using dynamic topology reconstruction. The model outputs the current state label and historical dependency feature vectors. Based on the Euclidean distance between the historical dependency feature vectors and the region boundaries, the state transition trend is determined, and a two-layer game optimization model is initiated to calculate maintenance timing and intensity parameters. After performing maintenance operations, the dataset is updated to achieve closed-loop state machine transition. This solves the technical problem of low accuracy in cumulative damage feature recognition due to the inability to effectively handle cross-stage semantic differences during mold lifecycle state transition prediction.

Description

Technical Field

[0001] This invention belongs to the field of production mold technology, and more specifically, relates to a method for managing the entire life cycle state machine of a mold. Background Technology

[0002] In the field of mold manufacturing and management, traditional condition monitoring methods mainly rely on parameter acquisition and analysis at a single stage, triggering maintenance decisions by setting fixed thresholds. Existing technologies typically employ independent sensor systems to monitor design parameters, machining parameters, production operation parameters, and maintenance parameters separately. Data from each stage uses different semantic standards and physical units, making unified analysis across stages difficult. In current mold lifecycle management, the lack of semantic mapping between geometric parameters in the design stage and temperature field distribution parameters in the production stage, coupled with the inability to establish a causal relationship between surface roughness parameters in the machining stage and wear depth parameters in the maintenance stage, means that condition recognition models can only make judgments based on local information from the current stage, ignoring the impact of historical cumulative damage on the current state. In other words, existing technologies suffer from the technical problem of low accuracy in identifying cumulative damage features due to the inability to effectively handle cross-stage semantic differences when predicting mold lifecycle state transitions. Summary of the Invention

[0003] In view of this, the present invention provides a method for state machine transition throughout the entire life cycle of mold management, which can solve the technical problem in the prior art that the inability to effectively handle cross-stage semantic differences in the prediction of state transitions throughout the entire life cycle of molds leads to low accuracy in the recognition of cumulative damage features.

[0004] This invention is implemented as follows: It provides a method for managing the entire lifecycle state machine of a mold. Parameters from the mold's design, processing, production, and maintenance phases are collected and stored uniformly in a lifecycle dataset. This dataset is then input into a semantic mapping layer for cross-stage semantic unification processing, outputting a unified semantic full-link parameter set. The unified semantic full-link parameter set undergoes manifold dimensionality reduction, mapping it to a three-dimensional topological space and dividing it into normal operation, light wear, moderate fatigue, and severe fault regions. The three-dimensional spatial coordinate set and historical state sequences are input into a cumulative damage state identification model, outputting the current state label and historical dependency feature vectors. The state transition trend is determined based on the Euclidean distance between the historical dependency feature vectors and the boundaries of each region. When the Euclidean distance is less than a preset threshold, a two-layer game optimization model is activated, outputting maintenance timing and intensity parameters. After performing maintenance operations, the lifecycle dataset is updated, and semantic unification processing is re-executed to achieve closed-loop state machine flow.

[0005] The semantic mapping layer is built on an industrial knowledge graph, which defines a standard ontology structure for the entire lifecycle of a mold. The standard ontology structure includes entity nodes and relation edges. Entity nodes include design entity nodes, processing entity nodes, production entity nodes, and maintenance entity nodes. Relation edges define the causal relationships and semantic mapping rules between entity nodes at different stages.

[0006] The semantic mapping layer uses natural language processing technology to parse unstructured data text at each stage to extract parameter names and numerical units, maps parameter names to entity nodes in the industrial knowledge graph, establishes equivalent relationship edges for the same physical quantity at different stages in the industrial knowledge graph, and establishes a correspondence between the expected number of cycles in the design stage and the actual number of cycles in the maintenance stage through a graph query algorithm.

[0007] Among them, the manifold dimensionality reduction process adopts the incremental t-SNE nonlinear dimensionality reduction algorithm. The incremental t-SNE nonlinear dimensionality reduction algorithm calculates the conditional probability distribution between each parameter point in the high-dimensional parameter space, and finds the corresponding probability distribution in the low-dimensional topological space to minimize the KL divergence between the two. The lightweight encoder network is trained to map the unified semantic full-link parameter set into a three-dimensional spatial coordinate set.

[0008] The lightweight encoder network consists of an input layer, three hidden layers, and an output layer. The input layer receives a unified semantic full-link parameter set. The three hidden layers reduce the parameter dimensions to 64, 32, and 16 dimensions respectively. The output layer outputs a three-dimensional spatial coordinate set. The lightweight encoder network is deployed on an edge computing device to perform online dimensionality reduction mapping on the real-time acquired unified semantic full-link parameter set.

[0009] The normal operation area, light wear area, moderate fatigue area and severe fault area in the three-dimensional topological space are determined by geometric boundary. The normal operation area is defined as the area where the Euclidean distance of the coordinate point in the three-dimensional spatial coordinate set from the origin is less than the first threshold. The light wear area is defined as the area where the Euclidean distance is greater than or equal to the first threshold and less than the second threshold.

[0010] Among them, the cumulative damage state recognition model utilizes a sparse connection learning framework based on dynamic topology reconstruction. During network training, it adaptively adds or removes connection edges according to the statistical distribution of neuron activation patterns, evaluates connection importance through information theory criteria and implements probabilistic connection pruning, and introduces grouping regularization constraints on connection weights to maintain the network's expressive power.

[0011] The cumulative damage state recognition model is a cascaded combination of an input layer, a long short-term memory (LSTM) layer, a sparse connection layer, a Bayesian inference layer, and an output layer. The LSTM layer contains 128 LSTM units, each with an input gate, a forget gate, and an output gate. The LSTM layer compresses the cumulative damage features in the historical state sequence into a 64-dimensional latent variable vector.

[0012] The sparse connection layer receives a 64-dimensional latent variable vector and a three-dimensional spatial coordinate set. During training, the importance score of each connection edge is calculated based on the statistical distribution of neuron activation frequency. When the importance score of a connection edge is lower than a dynamic threshold, the connection edge is pruned and removed. When the prediction error of a certain region continues to increase, new connection edges are added between the neurons corresponding to the region.

[0013] In this layer, the connection weights are grouped according to the functional group to which the neurons belong, and an application is made to the connection weights of each group. Regularization constraints ensure that the connection weights within the same functional group maintain a similar numerical range. The Bayesian inference layer calculates the posterior probability of the current state belonging to each region using the Bayesian formula and selects the region type with the highest posterior probability as the label of the current state.

[0014] The dynamic threshold is determined based on the standard deviation of the temperature field distribution parameters, the length of the historical state sequence, and the current network sparsity. The dynamic threshold value is calculated by dividing the standard deviation of the temperature field distribution parameters by the maximum observed value of the standard deviation of the temperature field distribution parameters, dividing the length of the historical state sequence by the maximum set value of the historical state sequence length, and dividing the current network sparsity by the target network sparsity.

[0015] Among them, the cumulative damage state identification model training dataset is generated by collecting real mold life cycle monitoring data, using finite element analysis software to establish a three-dimensional geometric model and material constitutive model of the mold, and simulating physical field data including thermal cracks, fatigue cracks and wear. The simulated physical field data is then mapped to a three-dimensional topological space through manifold dimensionality reduction processing to generate a three-dimensional spatial coordinate set of simulation samples.

[0016] Among them, the objective function of the upper-level model of the two-level game optimization model is used to minimize the state degradation rate. The state degradation rate is defined as the speed at which the historical dependency feature vector moves towards the severe fault region in the three-dimensional topological space per unit time. The constraints of the upper-level model include that the maintenance timing parameter is greater than or equal to the current time and less than or equal to the predicted fault time.

[0017] Among them, the objective function of the lower-level model of the two-level game optimization model is used to maximize the equipment availability time. The equipment availability time is defined as the cumulative time from the current moment to the next maintenance moment when the mold is in the normal operation area or the light wear area. The constraints of the lower-level model include that the difference between the maintenance timing parameter and the current moment is greater than the minimum maintenance interval time.

[0018] In the two-layer game optimization model, the coupling term is the product of the maintenance timing parameter and the maintenance strength parameter. The optimal solution is calculated by an iterative solution algorithm. First, the maintenance timing parameter is fixed and the objective function of the lower-level model is solved to obtain the optimal maintenance strength parameter. Then, the optimal maintenance strength parameter is substituted into the objective function of the upper-level model to obtain the optimal maintenance timing parameter. The iteration is repeated until convergence.

[0019] The maintenance operations include surface grinding, heat treatment repair, and coating recoating. The wear depth and crack length parameters re-collected after the maintenance operations are completed reflect the maintenance effect. The maintenance operation is considered successful when the re-collected wear depth parameter drops to the threshold range corresponding to the normal operating area and the re-collected crack length parameter drops below the safety threshold.

[0020] This invention achieves unified semantic processing of parameters across stages by constructing a semantic mapping layer based on an industrial knowledge graph. It maps heterogeneous parameters from the design, processing, production, and maintenance stages to a unified end-to-end parameter set. An incremental t-SNE nonlinear dimensionality reduction algorithm compresses the 100-dimensional parameters into a three-dimensional topological space. A cumulative damage state recognition model is constructed using a sparse connection learning framework based on dynamic topology reconstruction. A long short-term memory (LSTM) layer compresses the cumulative damage features in the historical state sequence into a fixed-dimensional latent variable vector. In the sparse connection layer, the connection topology is adaptively adjusted according to the neuron activation pattern, enabling the model to efficiently infer the current state label in a low-dimensional space while preserving historical dependencies. This reduces computational complexity from being proportional to the length of the historical data to a constant level. This invention solves the technical problem mentioned in the background art, where the inability to effectively handle cross-stage semantic differences during mold lifecycle state transition prediction leads to low accuracy in cumulative damage feature recognition. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2 The convergence curves of the maintenance timing parameter and maintenance strength parameter during the iterative solution of the two-level game optimization model are shown. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0024] like Figure 1 The diagram shown is a flowchart of a mold management full lifecycle state machine transition method provided by the present invention. This method includes the following steps:

[0025] S1: Collect the geometric parameters, material property parameters, and expected number of cycles of the mold during the design phase; collect the tool path parameters and surface roughness parameters during the machining phase; collect the temperature field distribution parameters, pressure field distribution parameters, and vibration spectrum parameters during the production phase; collect the wear depth parameters and crack length parameters during the maintenance phase; and store the geometric parameters, material property parameters, expected number of cycles, tool path parameters, surface roughness parameters, temperature field distribution parameters, pressure field distribution parameters, vibration spectrum parameters, wear depth parameters, and crack length parameters in a unified lifecycle dataset;

[0026] S2: Input the geometric parameters, material property parameters, expected number of cycles, tool path parameters, surface roughness parameters, temperature field distribution parameters, pressure field distribution parameters, vibration spectrum parameters, wear depth parameters, and crack length parameters from the full life cycle dataset into the semantic mapping layer for cross-stage semantic unification processing, and output a unified semantic full-link parameter set;

[0027] S3: Perform manifold dimensionality reduction on the unified semantic full-link parameter set, and map the hundred-dimensional parameters in the unified semantic full-link parameter set to a three-dimensional topological space through a nonlinear mapping algorithm to obtain a three-dimensional spatial coordinate set. Then, divide the three-dimensional topological space into normal operation area, light wear area, moderate fatigue area and severe fault area.

[0028] S4: Input the three-dimensional spatial coordinate set and historical state sequence into the cumulative damage state recognition model. The cumulative damage state recognition model outputs the current state label and the historical dependency feature vector. Determine the region type of the mold based on the current state label. Determine the state transition trend based on the Euclidean distance between the historical dependency feature vector and the boundary of each region.

[0029] S5: When the Euclidean distance value is less than the preset threshold, the two-layer game optimization model is started to perform maintenance decision calculation. The current wear depth parameter, historical average wear rate and historical state duration parameter are input into the two-layer game optimization model. The two-layer game optimization model outputs maintenance timing parameter and maintenance intensity parameter.

[0030] S6: Perform maintenance operations based on maintenance timing parameters and maintenance intensity parameters. After the maintenance operations are completed, re-collect the wear depth parameters and crack length parameters of the maintenance stage, update the full life cycle dataset, and re-execute steps S2 to S5 to achieve closed-loop flow of the state machine.

[0031] The semantic mapping layer is constructed based on an industrial knowledge graph, which defines a standard ontology structure for the entire lifecycle of a mold. The standard ontology structure includes entity nodes and relational edges. The entity nodes include design entity nodes, processing entity nodes, production entity nodes, and maintenance entity nodes. The relational edges define the causal relationships and semantic mapping rules between entity nodes at different stages.

[0032] The semantic mapping layer uses natural language processing technology to parse unstructured data text at each stage, extracting parameter names and numerical units. The extracted parameter names are mapped to entity nodes in an industrial knowledge graph. Equivalence edges are established for the same physical quantity at different stages within the industrial knowledge graph. A graph query algorithm is used to establish a correspondence between the expected number of cycles in the design stage and the actual number of cycles in the maintenance stage. The surface roughness parameters in the processing stage and the temperature field distribution parameters in the production stage are associated using the physical laws of heat conduction. The result is a unified semantic set of parameters across the entire process. The parameter names refer to physical quantity identifiers that characterize the mold state in the data at each stage.

[0033] The manifold dimensionality reduction process employs an incremental t-SNE nonlinear dimensionality reduction algorithm. This algorithm calculates the conditional probability distribution between parameter points in the high-dimensional parameter space and then searches for a corresponding probability distribution in the low-dimensional topological space that minimizes the KL divergence between the two. The incremental t-SNE algorithm trains a lightweight encoder network, which includes an input layer, three hidden layers, and an output layer. The input layer receives a unified semantic full-link parameter set. The three hidden layers sequentially reduce the parameter dimensions to 64, 32, and 16 dimensions, respectively. The output layer outputs a three-dimensional spatial coordinate set. The lightweight encoder network is deployed on an edge computing device to perform online dimensionality reduction mapping on the real-time acquired unified semantic full-link parameter set, outputting the mapping result as a three-dimensional spatial coordinate set.

[0034] The normal operation region, light wear region, moderate fatigue region, and severe failure region in the three-dimensional topological space are determined by geometric boundaries. The normal operation region is defined as the region where the Euclidean distance of the coordinate point in the three-dimensional spatial coordinate set from the origin is less than a first threshold. The light wear region is defined as the region where the Euclidean distance is greater than or equal to the first threshold and less than a second threshold. The moderate fatigue region is defined as the region where the Euclidean distance is greater than or equal to the second threshold and less than a third threshold. The severe failure region is defined as the region where the Euclidean distance is greater than or equal to the third threshold.

[0035] The historical state sequence is a sequence formed by arranging all historical state labels of the mold from its initial use to the current moment in chronological order. The cumulative damage state recognition model utilizes a sparse connection learning framework based on dynamic topology reconstruction. During network training, it adaptively adds and removes connection edges according to the statistical distribution of neuron activation patterns, evaluates connection importance using information theory criteria, and implements probabilistic connection pruning. At the same time, it introduces grouping regularization constraints on connection weights to maintain the network's expressive power.

[0036] The cumulative damage state recognition model is a cascaded combination of an input layer, a long short-term memory (LSM) layer, a sparse connection layer, a Bayesian inference layer, and an output layer. The input layer receives a set of three-dimensional spatial coordinates and a sequence of historical states. The LSM layer contains 128 LSM units, each with an input gate, a forget gate, and an output gate. The input gate controls the amount of current input information received, the forget gate controls the amount of historical information retained, and the output gate controls the activation level of the current output. The LSM layer compresses the cumulative damage features in the historical state sequence into a 64-dimensional latent variable vector.

[0037] The sparse connection layer receives a 64-dimensional latent variable vector and a three-dimensional spatial coordinate set. Initialized as a fully connected structure, the layer calculates the importance score of each connection edge based on the statistical distribution of neuron activation frequencies during training. This importance score is calculated using the information gain criterion. When the importance score of a connection edge falls below a dynamic threshold, the edge is pruned and removed. When the prediction error in a certain region continues to increase, a new connection edge is added between the neurons corresponding to that region. The dynamic threshold is adaptively adjusted based on the current sparsity of the network and the validation set error. The connection weights of the sparse connection layer are grouped according to the functional groups to which the neurons belong. L2 regularization is applied to each group of connection weights to maintain similar numerical ranges within the same functional group, preventing a decrease in network expressive power due to excessive pruning.

[0038] The Bayesian inference layer receives the output of the sparse connection layer and calculates the posterior probability that the current state belongs to the normal operation region, the light wear region, the moderate fatigue region, or the severe fault region using the Bayesian formula. The region type with the highest posterior probability is selected as the current state label. The output layer outputs the current state label and a historical dependency feature vector, which is a 64-dimensional latent variable vector output by the Long Short-Term Memory (LSTM) layer.

[0039] The sparse connection learning framework based on dynamic topology reconstruction continuously monitors the activation mode of each neuron during training, calculates the contribution of each connection edge to the final prediction result, and uses the mutual information index in information theory to quantify the effective information transmitted by the connection edge. When a connection edge is in a low activation state for a long time or the amount of information transmitted is less than a threshold, it is pruned to reduce network complexity. At the same time, new connections are dynamically added in the region where the prediction error is concentrated to enhance local expressive ability. The dynamic adjustment mechanism enables the network to significantly reduce the number of parameters and computation while maintaining high accuracy.

[0040] The sparse connection learning framework based on dynamic topology reconstruction enables the cumulative damage state recognition model to handle historical dependencies in non-Markov processes without storing and backtracking complete historical path data. Instead, it compresses long-term cumulative damage features into a fixed-dimensional 64-dimensional latent variable vector through a long short-term memory (LSTM) layer. The sparse connection layer then efficiently infers the current state label in this low-dimensional space, reducing computational complexity from being proportional to the length of the historical data to a constant level. This allows the cumulative damage state recognition model to achieve millisecond-level real-time inference on edge computing devices. Furthermore, the sparse connection layer inherently possesses anti-overfitting capabilities. When faced with extremely imbalanced small-sample fault data, by preserving key connection edges and dynamically adjusting the network topology, the cumulative damage state recognition model can effectively learn the characteristic patterns of rare faults without being overwhelmed by a large amount of normal state data, improving the accuracy and generalization ability for identifying low-frequency fault events.

[0041] The steps for establishing the training dataset for the cumulative damage state identification model specifically include: collecting full life-cycle monitoring data of real molds, which includes all three-dimensional spatial coordinate sets and corresponding manually labeled status labels from the time the mold is put into use to its scrapping; using finite element analysis software to establish a three-dimensional geometric model and material constitutive model of the mold, setting different temperature boundary conditions, pressure loading paths and cycle numbers, and simulating to generate physical field data including thermal cracks, fatigue cracks and wear; extracting temperature gradient, stress distribution and displacement field parameters from the simulated physical field data, and mapping the temperature gradient, stress distribution and displacement field parameters to a three-dimensional topological space through manifold dimensionality reduction processing to generate a three-dimensional spatial coordinate set of simulation samples; labeling the simulation samples with corresponding status labels according to the fault type set in the simulation; and mixing the full life-cycle monitoring data of real molds and the simulated physical field data in a 7:3 ratio to form a balanced training dataset including normal operation state, light wear state, moderate fatigue state and severe fault state.

[0042] The specific steps for training the cumulative damage state recognition model include: dividing the balanced training dataset into a training set and a validation set according to time order, with the training set accounting for 80% and the validation set accounting for 20%; initializing the weight matrix and bias vector of the long short-term memory unit layer as random values ​​following a normal distribution with a mean of 0 and a standard deviation of 0.1, and initializing the sparse connection layer as a fully connected structure with connection weights following a Xavier initialization distribution; setting the batch size to 32, the learning rate to 0.001, and the training epochs to 200; in each training epoch, randomly selecting batch samples from the training set, inputting the three-dimensional spatial coordinate set and historical state sequence into the cumulative damage state recognition model, and calculating the predicted current state label and historical dependency feature vector through forward propagation; calculating the cross-entropy loss between the predicted current state label and the true state label, and simultaneously calculating the sparse connection layer... The sparse connection layer uses a grouped regularization loss, which is the sum of the L2 norms of the weights in each group of connections. The cross-entropy loss and the grouped regularization loss are weighted and summed to obtain the total loss function. The gradient of the total loss function with respect to the parameters of each layer is calculated using the backpropagation algorithm. The parameters of the Long Short-Term Memory (LSTM) layer and the sparse connection layer are updated using the Adam optimizer. After each training epoch, the activation frequency and information gain of each connection edge in the sparse connection layer are calculated. When the information gain of a connection edge is less than a dynamic threshold, the connection edge is removed. When the prediction error of a region on the validation set is greater than an error threshold, a new connection edge is added between the neurons corresponding to that region. Every 10 training epochs, the performance of the cumulative damage state recognition model is evaluated on the validation set, and the classification accuracy and average loss of the validation set are recorded. Training stops when the validation set loss no longer decreases for 20 consecutive training epochs.

[0043] The dynamic threshold of the sparse connection layer is determined based on the standard deviation of the temperature field distribution parameters, the length of the historical state sequence, and the current network sparsity. The dynamic threshold calculation function is used to adaptively adjust the connection pruning sensitivity of the sparse connection layer. The dynamic threshold calculation function is based on the standard deviation of the temperature field distribution parameters, the length of the historical state sequence, and the current network sparsity to obtain the dynamic threshold value.

[0044] The dynamic threshold calculation function is expressed as follows: divide the standard deviation of the temperature field distribution parameters by the maximum observed value of the standard deviation of the temperature field distribution parameters, divide the length of the historical state sequence by the maximum set value of the historical state sequence length, divide the current network sparsity by the target network sparsity, multiply the normalized standard deviation of the temperature field distribution parameters and the normalized length of the historical state sequence by the square root of the normalized current network sparsity, and then multiply by the benchmark threshold coefficient to obtain the dynamic threshold value.

[0045] When the dynamic threshold value is within [0, 0.3), a high-sensitivity pruning strategy is used to adjust the connection retention rate parameter of the sparse connection layer to 60% to 70%; when the dynamic threshold value is within [0.3, 0.7), a medium-sensitivity pruning strategy is used to adjust the connection retention rate parameter of the sparse connection layer to 70% to 85%; when the dynamic threshold value is within [0.7, 1.0], a low-sensitivity pruning strategy is used to adjust the connection retention rate parameter of the sparse connection layer to 85% to 95%; when the dynamic threshold value is greater than 1.0, the connection pruning operation is paused and the current network topology remains unchanged. The connection retention rate parameter is the ratio of the number of retained connection edges in the sparse connection layer to the total number of connection edges in the initial fully connected structure.

[0046] The Euclidean distance value is the minimum Euclidean distance from the coordinate point of the historical dependent feature vector in the three-dimensional topological space to the boundary of the normal operation area, the boundary of the light wear area, the boundary of the moderate fatigue area, and the boundary of the severe failure area. The preset threshold is determined based on the average Euclidean distance value before the state transition occurred in the historical statistical data.

[0047] The historical average wear rate is the average value of the change in the wear depth parameter in historical statistical data divided by the corresponding time interval. The historical state duration parameter is the length of time the mold remains in the area corresponding to the current state label.

[0048] The objective function of the upper-level model of the two-layer game optimization model is used to minimize the state degradation rate. The state degradation rate is defined as the speed at which the historical dependency feature vector moves towards the critical fault region in the three-dimensional topological space per unit time. The inputs of the upper-level model objective function include maintenance timing parameters, maintenance intensity parameters, and current wear depth parameters. The output of the upper-level model objective function is the minimum value of the state degradation rate.

[0049] The objective function of the upper-level model is expressed as follows: divide the current wear depth parameter by the maximum allowable value of the wear depth parameter, divide the maintenance intensity parameter by the maximum applicable value of the maintenance intensity parameter, subtract the square of the normalized maintenance intensity parameter from the product of the normalized current wear depth parameter and the reciprocal of the maintenance timing parameter, and add the product of the current wear depth parameter and the historical average wear rate, divided by the maximum observed value of the historical average wear rate, to obtain the condition deterioration rate.

[0050] The constraints of the upper-level model include a maintenance timing parameter greater than or equal to the current time and less than or equal to the predicted failure time, and a maintenance intensity parameter greater than or equal to the minimum effective maintenance intensity and less than or equal to the maximum applicable value of the maintenance intensity parameter. The predicted failure time is the time when the mold enters the severe failure region, predicted based on the current state degradation rate and historical dependency feature vectors.

[0051] The objective function of the lower-level model in the two-layer game optimization model is used to maximize the equipment availability time. The equipment availability time is defined as the cumulative duration from the current time to the next maintenance time when the mold is in the normal operation area or the light wear area. The input of the lower-level model objective function includes maintenance timing parameters, maintenance intensity parameters, and historical state duration parameters. The output of the lower-level model objective function is the maximum value of the equipment availability time.

[0052] The objective function of the lower-level model is expressed as follows: divide the difference between the maintenance timing parameter and the current time by the difference between the predicted fault time and the current time, divide the maintenance intensity parameter by the maximum applicable value of the maintenance intensity parameter, divide the historical state duration parameter by the maximum observed value of the historical state duration parameter, multiply the difference between the normalized maintenance timing parameter and the current time by the natural logarithm of the normalized maintenance intensity parameter, and add the normalized historical state duration parameter to obtain the equipment availability time.

[0053] The constraints of the lower-level model include that the difference between the maintenance timing parameter and the current time is greater than the minimum maintenance interval, and the ratio of the maintenance intensity parameter to the historical average maintenance intensity is greater than 0.8 and less than 1.5. The historical average maintenance intensity is the average of the maintenance intensity parameters of all maintenance operations in the historical statistical data.

[0054] The coupling term in the two-level game optimization model is the product of the maintenance timing parameter and the maintenance intensity parameter. This coupling term appears simultaneously in the objective functions of both the upper-level and lower-level models, reflecting the mutual influence between timing selection and intensity configuration in maintenance decisions. The two-level game optimization model calculates the optimal solution using an iterative algorithm. This algorithm first fixes the maintenance timing parameter to solve the lower-level model's objective function to obtain the optimal maintenance intensity parameter. Then, it substitutes the optimal maintenance intensity parameter into the upper-level model's objective function to obtain the optimal maintenance timing parameter. This iteration is repeated until the changes in the objective function values ​​of both the upper and lower levels are less than a convergence threshold.

[0055] The maintenance operations include surface grinding, heat treatment repair, and coating recoating. Surface grinding removes the wear layer and surface cracks. Heat treatment repair restores the material microstructure through annealing and quenching. Coating recoating reapplies a wear-resistant coating to the mold surface. After the maintenance operations are completed, the re-collected wear depth and crack length parameters reflect the maintenance effect. When the re-collected wear depth parameter drops to the threshold range corresponding to the normal operating area and the re-collected crack length parameter drops below the safety threshold, the maintenance operation is considered successful. The updated full lifecycle dataset is then re-input into the semantic mapping layer, and steps S2 to S5 are repeated to achieve continuous monitoring and dynamic adjustment of the mold status.

[0056] As one implementation method, the present invention also provides a computer-based approach to form a mold management full lifecycle state machine flow system. The computer is equipped with a readable storage medium that stores program instructions. When the program instructions are run on the computer, they execute the above-described mold management full lifecycle state machine flow method.

[0057] The specific implementation methods of the above steps are described in detail below.

[0058] The specific implementation of step S1 is as follows: During the mold design stage, the digital model file of the mold is read using 3D computer-aided design software. Geometric parameters such as length, wall thickness, and cavity depth are extracted from the model. Simultaneously, material property parameters such as the elastic modulus, Poisson's ratio, and coefficient of thermal expansion of the selected material are retrieved from the material database. The expected number of cycles for the mold is then set according to the design requirements. This expected number of cycles is typically set to... to In the machining stage, the CNC machine tool's control system records the tool's motion trajectory coordinate sequence in three-dimensional space in real time, forming tool trajectory parameters. A surface roughness measuring instrument is used to perform multi-point sampling measurements on the machined mold surface to obtain surface roughness parameters, typically ranging from 0.8 to 6.3 micrometers. In the production stage, temperature data is collected using a thermocouple sensor array positioned at key locations on the mold, forming temperature field distribution parameters. A pressure sensor array collects force data from various areas, forming pressure field distribution parameters. Vibration signals from the mold during operation are collected using a vibration acceleration sensor and processed through a fast Fourier transform to obtain vibration spectrum parameters. In the maintenance stage, the wear depth of the mold surface is measured using a laser displacement sensor or ultrasonic thickness gauge, obtaining wear depth parameters. Cracks inside or on the surface of the mold are detected using magnetic particle testing or penetrant testing, and their lengths are measured, obtaining crack length parameters. All parameters collected in each stage are indexed and associated according to a unified timestamp and mold number, and stored in a distributed database supporting structured queries, forming a full lifecycle dataset. This dataset provides the data foundation for subsequent cross-stage semantic unified processing and state recognition.

[0059] The specific implementation of step S2 is as follows: First, an industrial knowledge graph is constructed as the core knowledge base of the semantic mapping layer. This knowledge graph uses ontology modeling to define entity nodes and relational edges for each stage of the mold's entire lifecycle. Entity nodes include geometric model entities and material specification entities in the design stage, tool path entities and surface quality entities in the machining stage, temperature field entities and stress field entities in the production stage, and defect feature entities in the maintenance stage. Relational edges define the causal relationships between entities at different stages. For example, material property parameters in the design stage are mapped to temperature field distribution parameters in the production stage through thermophysical property relational edges, and surface roughness parameters in the machining stage are associated with temperature gradients in the production stage through contact thermal resistance relational edges. After receiving the entire lifecycle dataset, the semantic mapping layer uses named entity recognition algorithms in natural language processing to parse the unstructured text descriptions in the data of each stage, extract parameter names and numerical units, semantically match the extracted parameter names with entity nodes in the knowledge graph, and retrieve the same physical quantity's expression form at different stages in the knowledge graph using graph query algorithms to establish equivalence relational edges to achieve semantic unification. For example, the expected number of cycles in the design phase is mapped to the actual number of cycles in the maintenance phase using a cycle life relationship. Similarly, the surface roughness parameters in the machining phase are quantitatively mapped to the temperature field distribution parameters in the production phase using a contact thermal resistance model based on the laws of thermal conduction. After semantic mapping, the same physical quantity across different phases is unified into standardized parameter names and units, outputting a unified semantic end-to-end parameter set containing parameters from all phases. This parameter set typically has 100 to 200 dimensions.

[0060] The specific implementation of step S3 is as follows: An incremental t-SNE nonlinear dimensionality reduction algorithm is used to perform manifold dimensionality reduction on the unified semantic full-link parameter set. This algorithm is based on the principle of probability distribution similarity. It calculates the conditional probability distribution between parameter points in the high-dimensional parameter space, uses a Gaussian kernel function to measure the similarity between point pairs in the high-dimensional space, and then constructs the corresponding probability distribution in the low-dimensional topological space. By minimizing the KL divergence between the probability distributions of the high-dimensional and low-dimensional spaces, a nonlinear mapping from high-dimensional parameters to low-dimensional space is achieved. To achieve real-time dimensionality reduction, a lightweight encoder network is trained. This network includes an input layer that receives a unified semantic full-link parameter set of 100 to 200 dimensions; a first hidden layer containing 64 neurons that uses the ReLU activation function to reduce the parameter dimension to 64 dimensions; a second hidden layer containing 32 neurons that further reduces the dimension to 32 dimensions; a third hidden layer containing 16 neurons that reduces the dimension to 16 dimensions; and an output layer containing 3 neurons that outputs a three-dimensional spatial coordinate set. A trained lightweight encoder network is deployed on an edge computing device to perform online dimensionality reduction mapping on real-time acquired parameters, resulting in three-dimensional spatial coordinates. In this three-dimensional topological space, a spherical region partitioning model is established centered on the origin. Based on the distribution patterns of coordinate points corresponding to different states in historical statistical data, a first threshold of 2.5 is set, defining regions with an Euclidean distance less than 2.5 as normal operating regions. A second threshold of 5.0 is set, defining regions with an Euclidean distance greater than or equal to 2.5 and less than 5.0 as lightly worn regions. A third threshold of 8.0 is set, defining regions with an Euclidean distance greater than or equal to 5.0 and less than 8.0 as moderately fatigued regions, and regions with an Euclidean distance greater than or equal to 8.0 as severely faulty regions. This geometric boundary partitioning method provides a spatial reference benchmark for subsequent state identification.

[0061] The specific implementation of step S4 is as follows: The cumulative damage state recognition model receives a set of three-dimensional spatial coordinates and a historical state sequence as input. The historical state sequence is a sequence of all state labels recorded in chronological order from the time the mold was put into use to the current time. The input layer of the model encodes the three-dimensional spatial coordinates and the historical state sequence into numerical vectors respectively. The long short-term memory unit layer contains 128 long short-term memory units. Each unit selectively receives the current input information through an input gate, decides to retain or discard historical information through a forget gate, and controls the output activation at the current time through an output gate. Utilizing the memory mechanism of the long short-term memory units, the cumulative damage features in the historical state sequence are compressed into a 64-dimensional latent variable vector. This vector encodes the cumulative damage history information of the mold from the past to the present. The sparse connection layer is initialized as a fully connected structure. During training, the activation frequency of each connection edge is counted according to the activation frequency of neurons, and the information gain transmitted by the connection edge is calculated. The information gain is quantified by comparing the difference in model prediction error with and without the connection. When the information gain of a connection edge is less than a dynamic threshold, a pruning operation is performed to remove the connection. When the prediction error in a certain region on the validation set continues to increase, a new connection is added between the corresponding neurons. To prevent excessive pruning from reducing expressive power, the connection weights of the sparse connection layer are grouped according to neuron function, and L2 regularization is applied to each group of connection weights to maintain similar numerical ranges within the same group. The Bayesian inference layer receives the output of the sparse connection layer, calculates the posterior probability of the current state belonging to each region using the Bayesian formula, and selects the region with the highest posterior probability as the current state label. The model outputs the current state label and a 64-dimensional history dependency feature vector. The history dependency feature vector is mapped to a three-dimensional topological space, and the minimum Euclidean distance from its coordinates to the boundaries of each region is calculated. This distance reflects the trend of the mold state transitioning to each region. When the Euclidean distance is less than a preset threshold of 1.2, it indicates that the mold state is about to transition, and maintenance decision calculation needs to be initiated.

[0062] The specific implementation of step S5 is as follows: after determining that a state transition trend exists, a two-layer game optimization model is initiated to calculate maintenance decisions. This model receives the current wear depth parameter, the historical average wear rate, and the historical state duration parameter as inputs. The upper-layer model aims to minimize the state degradation rate, which is defined as the speed at which the historically dependent feature vector moves towards the severe fault region in the three-dimensional topological space per unit time. The objective function of the upper-layer model is obtained by multiplying the normalized current wear depth parameter by the reciprocal of the maintenance timing parameter, subtracting the square of the normalized maintenance intensity parameter, and adding the normalized product of the current wear depth and the historical average wear rate. This objective function reflects the physical law that the later the maintenance timing and the weaker the maintenance intensity, the greater the degradation rate. The upper-layer model's constraints require that the maintenance timing parameter be no earlier than the current time and no later than the predicted fault time, and that the maintenance intensity parameter be between the minimum effective maintenance intensity of 0.3 and the maximum applicable value of 1.0. The lower-level model aims to maximize equipment availability time, defined as the cumulative duration from the current moment to the next maintenance time when the mold is in the normal operating or lightly worn area. The objective function of the lower-level model normalizes the difference between the maintenance timing and the current moment, multiplies it by the natural logarithm of the maintenance intensity parameter, and adds the normalized historical state duration parameter to obtain the equipment availability time. This function reflects the relationship that delaying maintenance timing and increasing maintenance intensity can extend equipment availability time. The constraints of the lower-level model require that the difference between the maintenance timing and the current moment be greater than the minimum maintenance interval of 24 hours, and that the ratio of maintenance intensity to the historical average maintenance intensity be between 0.8 and 1.5. The two-layer game optimization model calculates the optimal solution through an iterative algorithm. First, the maintenance timing parameter is fixed. The partial derivative of the objective function of the lower-layer model with respect to the maintenance intensity parameter is taken and set to zero to obtain the optimal maintenance intensity parameter. Then, this parameter is substituted into the objective function of the upper-layer model, and the partial derivative with respect to the maintenance timing parameter is taken and set to zero to obtain the optimal maintenance timing parameter. The iteration is repeated until the change in the values ​​of the upper and lower-layer objective functions is less than the convergence threshold of 0.01. Finally, the optimal maintenance timing parameter and maintenance intensity parameter are output.

[0063] The specific implementation of step S6 is as follows: according to the maintenance timing parameters and maintenance intensity parameters output by the two-layer game optimization model, corresponding maintenance operations are performed. When the maintenance intensity parameter is between 0.3 and 0.5, surface grinding is performed, using a precision grinding machine to remove the wear layer and microcracks on the mold surface, with the grinding depth controlled between 0.1 and 0.3 mm. When the maintenance intensity parameter is between 0.5 and 0.8, heat treatment repair is performed, heating the mold to the recrystallization temperature in an annealing furnace and holding it for a certain time before slowly cooling it to eliminate internal stress and restore the material structure. When the maintenance intensity parameter is between 0.8 and 1.0, coating recoating is performed, first removing the original coating, and then reapplying a wear-resistant coating such as titanium nitride or tungsten carbide to the mold surface through physical vapor deposition or chemical vapor deposition technology, with the coating thickness controlled between 5 and 15 micrometers. After maintenance is completed, the wear depth parameter is measured again using a laser displacement sensor, and the crack length parameter is detected using a magnetic particle inspection device. When the wear depth parameter drops below 0.05 mm and the crack length parameter drops below 0.5 mm, the maintenance operation is considered successful. The updated wear depth and crack length parameters are written into the full lifecycle dataset, triggering a dataset update event. The system automatically re-executes the semantic mapping processing in step S2, the manifold dimensionality reduction processing in step S3, the state recognition in step S4, and the maintenance decision calculation in step S5, forming a closed-loop flow mechanism to achieve continuous monitoring and dynamic adjustment of the mold state, ensuring that the mold always operates within a controllable state area.

[0064] It should be noted that the key technical ideas of this invention are as follows: First, an industrial knowledge graph is used to construct a semantic mapping layer to achieve a unified representation of parameters across stages. Through ontology modeling, heterogeneous parameters from design, processing, production, and maintenance stages are mapped to a unified semantic space. A graph query algorithm is used to establish equivalence relationships for the same physical quantity at different stages, solving the information gap problem caused by isolated data and semantic inconsistencies in traditional methods. This allows for correlation analysis of full lifecycle data within a unified framework, providing complete feature input for subsequent state identification. Second, a sparse connection learning framework based on dynamic topology reconstruction is used to construct a cumulative damage state identification model. Network connections are dynamically adjusted using information theory criteria. During training, pruning and connection addition operations are adaptively performed based on the information gain of the connection edges. This reduces the number of network parameters and computational complexity, enabling the model to be deployed on edge devices for millisecond-level real-time inference. Furthermore, grouping regularization constraints maintain the network's expressive power, solving the overfitting problem of traditional deep learning models when dealing with small-sample imbalanced fault data and improving the accuracy of identifying rare fault patterns. Third, a two-layer game optimization model is adopted for maintenance decision-making. The upper layer aims to minimize the rate of state degradation, while the lower layer aims to maximize equipment availability. By using coupling terms, the mutual influence between maintenance timing and intensity is modeled as a game process. An iterative solution algorithm is used to find the Nash equilibrium solution, which solves the contradiction that traditional single-objective optimization methods cannot simultaneously consider equipment performance maintenance and production efficiency, and achieves optimal allocation of maintenance resources. The synergistic effect of these three technical ideas is that the semantic mapping layer provides high-quality end-to-end feature input for the state recognition model, the sparse connection learning framework achieves efficient and accurate state recognition and trend prediction, the two-layer game model formulates the optimal maintenance strategy based on the recognition results, and the data update after maintenance execution is fed back to the semantic mapping layer to form a closed loop. This constitutes a complete technical chain from data collection, semantic unification, state recognition to decision optimization. Compared with traditional methods that only focus on a single stage or use empirical threshold judgment for maintenance strategies, this invention realizes intelligent management and predictive maintenance of the entire mold life cycle.

[0065] It should be noted that this invention also solves the following technical problem: In existing mold maintenance decisions, the selection of maintenance timing and maintenance intensity are often independent of each other, lacking consideration of the coupling relationship between the two. This leads to the technical problem that maintenance decisions cannot achieve a dynamic balance between minimizing the rate of state degradation and maximizing equipment availability. This invention constructs a two-layer game optimization model. The upper-layer model optimizes the maintenance timing parameter with the objective function of minimizing the rate of state degradation, while the lower-layer model optimizes the maintenance intensity parameter with the objective function of maximizing equipment availability. The two models are coupled through the product term of the maintenance timing parameter and the maintenance intensity parameter. An iterative solution algorithm is used to solve for the optimal maintenance intensity parameter when the maintenance timing parameter is fixed. Then, the optimal maintenance intensity parameter is substituted into the upper-layer model to solve for the optimal maintenance timing parameter. This process is repeated iteratively until convergence, enabling maintenance decisions to simultaneously consider state degradation suppression and production efficiency assurance, thus achieving synergistic optimization of maintenance timing and maintenance intensity.

[0066] Specifically, the principle of this invention is as follows: The solution to this technical problem lies in the fact that the industrial knowledge graph establishes semantic mapping rules between entity nodes at different stages by defining a standard ontology structure. It establishes a correspondence between the expected number of cycles in the design stage and the actual number of cycles in the maintenance stage through a graph query algorithm, and establishes an association mapping between the surface roughness parameters in the processing stage and the temperature field distribution parameters in the production stage through the physical laws of heat conduction, thus eliminating cross-stage semantic differences. The incremental t-SNE nonlinear dimensionality reduction algorithm establishes a probability distribution mapping between the high-dimensional parameter space and the low-dimensional topological space, causing parameter points in similar states to cluster in the three-dimensional space and parameter points in different states to separate in the three-dimensional space, providing a geometric basis for state boundary partitioning. The sparse connection learning framework based on dynamic topology reconstruction dynamically adjusts connection edges by monitoring neuron activation patterns. It adds connections in regions with concentrated prediction errors to enhance local expressive power and removes redundant connections in low-activation regions to reduce the risk of overfitting. This allows the model to effectively learn the feature patterns of rare faults without being overwhelmed by a large amount of normal state data, improving the accuracy of cumulative damage feature recognition.

[0067] The following provides a specific embodiment 1 of the present invention. The specific implementation methods of steps S1, S2 and S6 in this embodiment 1 are the same as those described above, and will not be repeated in detail here. The specific implementation methods of other steps are described in detail below.

[0068] In step S3, the manifold dimensionality reduction process employs an incremental t-SNE nonlinear dimensionality reduction algorithm. This algorithm maps the unified semantic end-link parameter set to a three-dimensional topological space through a lightweight encoder network. The forward propagation process of the lightweight encoder network includes the mapping from the input layer to the first hidden layer, the cascaded mapping between the three hidden layers, and the coordinate generation of the output layer. Layer to the first The mapping relationship of the layers is expressed as follows:

[0069] .

[0070] In the formula, For the first The activation output vector of the layer; For the first Layer to the first Layer weight matrix; For the first The activation output vector of the layer; For the first Layer bias vector; The activation function is a modified linear unit function. Wherein, The set of end-to-end parameters for unified semantics received by the input layer, with a dimension of 100. For the first hidden layer Output vector; This is the 32-dimensional output vector of the second hidden layer; This is the 16-dimensional output vector of the third hidden layer; the three-dimensional spatial coordinate set generated by the output layer is represented as... , It is a three-dimensional coordinate set with dimension 3. This is the weight matrix from the third hidden layer to the output layer. This represents the bias vector of the output layer. The optimization objective of the incremental t-SNE nonlinear dimensionality reduction algorithm is to minimize the KL divergence between the high-dimensional space and the low-dimensional space. The KL divergence calculation formula is as follows:

[0071] .

[0072] In the formula, This represents the KL divergence value. For the first in a high-dimensional parameter space The parameter point and the first Conditional probability distribution among parameter points; For the first in a low-dimensional topological space The coordinates of the nth point and the th Probability distribution between coordinate points; This is a natural logarithm operation. The division of regions in the three-dimensional topological space is achieved through Euclidean distance determination, with coordinate points... The Euclidean distance to the origin is calculated as follows:

[0073] .

[0074] In the formula, For the first The Euclidean distance from each coordinate point to the origin; , , The first The three-dimensional components of each coordinate point. When At that time, the coordinate point belongs to the normal operating range. This area is considered to be slightly worn. This falls within the moderate fatigue range. This area is considered a critical fault zone. The first threshold is empirically set at 0.5. The second threshold has an empirical value of 1.2. The third threshold is 2.0, based on experience.

[0075] In step S4, the cumulative damage state identification model outputs the current state label and the historical dependency feature vector. The Euclidean distance value of the historical dependency feature vector in the three-dimensional topological space is calculated as follows:

[0076] .

[0077] In the formula, This represents the minimum Euclidean distance from the historically dependent feature vectors to the boundaries of each region. The coordinates of the history-dependent feature vector in the three-dimensional topological space; For the first Coordinate points on the boundary of each region; For Euclidean norm operations. When When the state transition trend is determined to be significant, the two-level game optimization model is initiated, in which... The preset threshold is determined based on historical statistical data and is typically set to 0.15. The information gain calculation for edges in a sparse connection layer is described below:

[0078] .

[0079] In the formula, For the first Information gain of each connecting edge; The entropy of the output state label; The activation value of the connecting edge is equal to The number of samples; This represents the total number of samples; Given the activation value of the connection edge The conditional entropy of the output status label is used.

[0080] The dynamic threshold calculation function for the sparse connection layer in step S5 is described as follows:

[0081] .

[0082] In the formula, The value is a dynamic threshold, dimensionless. This is the baseline threshold coefficient, dimensionless, and typically takes a value of 0.5. represents the standard deviation of the temperature field distribution parameters, in °C; This represents the maximum observed standard deviation of the temperature field distribution parameters, in °C. The length of the historical state sequence is dimensionless. This is the maximum set value for the length of the historical state sequence, and it is dimensionless. The current network sparsity is dimensionless; The target network sparsity is dimensionless and defaults to 0.3. Based on... The numerical range of the sparse connection layer's connection retention rate parameter is adjusted when... A highly sensitive pruning strategy was employed, with the connection retention rate parameter adjusted to 60% to 70%; when A moderately sensitive pruning strategy was adopted, with the connection retention rate parameter adjusted to 70% to 85%; when A low-sensitivity pruning strategy was adopted, and the connection retention rate parameter was adjusted to 85% to 95%; when The connection pruning operation is paused at any time. The historical average wear rate is calculated as follows:

[0083] .

[0084] In the formula, The historical average wear rate is expressed in μm / h. This represents the total number of wear and tear events recorded in historical statistics. For the first The change in wear depth parameter recorded each time, in μm; For the first The time interval corresponding to each record is in hours (h). The calculation method for predicting the fault time is as follows:

[0085] .

[0086] In the formula, The time of failure is predicted in hours (h). The current time is expressed in hours (h). The minimum Euclidean distance from the historically dependent feature vector to the boundaries of each region is dimensionless. The rate of degradation is expressed in units of 1000 m / s. The objective function of the upper-level model in the two-level game optimization model is expressed as follows:

[0087] .

[0088] In the formula, The rate of degradation is expressed in units of 1000 m / s. ; This is the current wear depth parameter, in μm; This represents the maximum allowable value for the wear depth parameter, in μm. To maintain timing parameters, the unit is hours; To maintain the strength parameters, they are dimensionless; To maintain the maximum applicable value of the strength parameter, it is dimensionless; The historical average wear rate is expressed in μm / h. This represents the maximum observed historical average wear rate, expressed in μm / h. The constraints of the upper-level model are as follows: and ,in The minimum effective maintenance intensity is dimensionless. The historical average maintenance intensity is calculated as follows:

[0089] .

[0090] In the formula, The historical average maintenance intensity is dimensionless. This represents the total number of maintenance operations in the historical statistics. For the first The maintenance intensity parameter for this maintenance operation is dimensionless. The objective function of the lower-level model in the two-level game optimization model is expressed as follows:

[0091] .

[0092] In the formula, The available time of the equipment is dimensionless. This is a parameter representing the duration of the historical state, in hours (h). The maximum observed value of the historical state duration parameter, in hours; It is the natural logarithm function. The constraints of the lower-level model are: and ,in The minimum maintenance interval is expressed in hours (h). The two-level game optimization model calculates the optimal solution using an iterative algorithm, first fixing... Solving the objective function of the lower-level model yields the optimal solution. Then the optimal Substituting the values ​​into the objective function of the upper-level model, we obtain the optimal solution. The iteration is repeated until the changes in the objective function values ​​of both the upper and lower models are less than the convergence threshold, which is usually set to 0.001.

[0093] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0094] To verify the effectiveness of this invention, technicians set up a test environment and selected a precision injection mold as the monitoring object. This mold is used to produce automotive dashboard parts, and the designed cycle number is [number missing]. During the mold design phase, technicians collected geometric parameters, including cavity length of 185 mm, cavity width of 92 mm, and wall thickness of 12 mm. Material properties included elastic modulus of 210 GPa, Poisson's ratio of 0.28, and coefficient of thermal expansion. / ℃, expected number of cycles set to During the machining phase, the CNC machine tool recorded a total of 37,856 coordinate points for tool path parameters. A surface roughness measuring instrument selected 15 measuring points on the mold surface, obtaining an average surface roughness parameter of 1.8 micrometers. During the production phase, technicians placed 18 thermocouple sensors on the mold cavity surface, cooling channels, and mold edges to collect temperature field distribution parameters. Simultaneously, pressure sensors were installed at 8 key stress locations to collect pressure field distribution parameters, and a triaxial vibration accelerometer was installed on the mold base to collect vibration spectrum parameters. During the maintenance phase, a laser displacement sensor measured the wear depth parameter to be 0.15 mm, and a magnetic particle inspection device detected the crack length parameter to be 0.8 mm. Technicians stored these parameters in a distributed database indexed by timestamp and mold number, forming a full lifecycle dataset with a total capacity of 2.3 GB.

[0095] Technicians input the entire lifecycle dataset into a semantic mapping layer, which performs cross-stage semantic unification processing based on an industrial knowledge graph. The industrial knowledge graph defines design entity nodes (including geometric models and material specifications), machining entity nodes (including toolpaths and surface quality), production entity nodes (including temperature and stress fields), and maintenance entity nodes (including defect features). Through graph query algorithms, technicians determine the expected number of iterations in each design phase. The second and maintenance phases have actually been running. The correspondence is established by mapping the surface roughness parameter of 1.8 micrometers in the processing stage to the temperature gradient in the production stage through the contact thermal resistance model, and outputting a unified semantic full-link parameter set with a parameter dimension of 156.

[0096] Technicians employed an incremental t-SNE nonlinear dimensionality reduction algorithm to perform manifold dimensionality reduction on the 156-dimensional full-link parameter set. After receiving the 156-dimensional parameters, the lightweight encoder network reduced the dimension to 64 dimensions in the first hidden layer (64 neurons), 32 in the second hidden layer (32 neurons), and 16 in the third hidden layer (16 neurons). The output layer, consisting of 3 neurons, outputs three-dimensional spatial coordinates of 2.85, 4.23, and 1.67. Technicians set a first threshold of 2.5, a second threshold of 5.0, and a third threshold of 8.0 in the three-dimensional topology space. Based on the Euclidean distance of 5.38 from the origin to the current coordinate point, the mold was determined to be in a moderate fatigue region. The boundaries of each region in the three-dimensional topology space were clearly visible: the normal operating region was the sphere's center, the slightly worn region was the first spherical shell, the moderately fatigued region was the second spherical shell, and the severely faulty region was the outermost region. The current coordinate point was located within the moderate fatigue region.

[0097] Technicians input a three-dimensional spatial coordinate set and a historical state sequence into the cumulative damage state recognition model. The historical state sequence contains 127 state labels from the time the mold was put into use to the current time, in chronological order: 92 normal operation states, 28 mild wear states, and 7 moderate fatigue states. As shown in Table 1, the long short-term memory (LSTM) layer of the cumulative damage state recognition model contains 128 units, compressing the historical state sequence into a 64-dimensional latent variable vector. The sparse connection layer is initialized as a fully connected structure. During training, the information gain of the connection edges is calculated based on the neuron activation frequency. When the information gain is less than the dynamic threshold of 0.42, the connection edge is removed. After training, the connection retention rate of the sparse connection layer is 78%. The Bayesian inference layer calculates the posterior probability of the current state belonging to the normal operation region as 0.08, the posterior probability of belonging to the mild wear region as 0.15, the posterior probability of belonging to the moderate fatigue region as 0.71, and the posterior probability of belonging to the severe fault region as 0.06. The moderate fatigue region with the highest posterior probability is selected as the current state label. The coordinates of the historical dependency feature vector output by the model in the three-dimensional topological space are 2.92, 4.35, and 1.73. The minimum Euclidean distance from this coordinate point to the boundary of the moderate fatigue region is calculated to be 0.98, which is less than the preset threshold of 1.2, indicating that the mold state is about to shift to the severe failure region.

[0098] Table 1. Parameter Configuration Table for Each Layer of the Cumulative Damage State Identification Model

[0099]

[0100] Technicians initiated a two-level game optimization model to calculate maintenance decisions. The current wear depth parameter is 0.15 mm, and the historical average wear rate is... The wear rate is calculated per millimeter per cycle, with a historical state duration parameter of 136 hours. The upper-level model's objective function calculates the condition degradation rate, normalizing the current wear depth parameter of 0.15 mm to 0.3 and the initial maintenance intensity parameter of 0.6 to 0.6, resulting in a degradation rate of 0.58. The lower-level model's objective function calculates the equipment availability time, normalizing the difference between the maintenance timing parameter and the current time (72 hours) to 0.45 and the maintenance intensity parameter of 0.6 to 0.6, resulting in an equipment availability time of 0.67. Technicians used an iterative algorithm to calculate the optimal maintenance timing parameter. In the first iteration, the maintenance timing parameter was fixed at 72 hours, and the optimal maintenance intensity parameter was found to be 0.68 in the lower-level model. In the second iteration, 0.68 was substituted into the upper-level model, yielding an optimal maintenance timing parameter of 84 hours. In the third iteration, the lower-level model was fixed at 84 hours, yielding a maintenance intensity parameter of 0.72. In the fourth iteration, the upper-level model was solved, yielding a maintenance timing parameter of 86 hours. After eight iterations, the changes in the objective function values ​​of both the upper and lower levels were less than the convergence threshold of 0.01, resulting in the output optimal maintenance timing parameter of 86 hours and the optimal maintenance intensity parameter of 0.73. Figure 2 As shown, the iterative convergence process of the two-layer game optimization model clearly demonstrates the dynamic adjustment trajectory of the maintenance timing parameter and the maintenance strength parameter. After 8 iterations, both parameters tend to stabilize.

[0101] Technicians performed heat treatment repair based on the optimal maintenance strength parameter of 0.73. The mold was heated to 650℃ in an annealing furnace and held for 4 hours, then slowly cooled to room temperature at a rate of 20℃ per hour to eliminate internal stress and restore the material microstructure. After the maintenance operation, technicians re-measured the wear depth parameter using a laser displacement sensor, which was 0.04 mm, and the crack length parameter using magnetic particle testing equipment, which was 0.3 mm, indicating successful maintenance. Technicians wrote the updated wear depth and crack length parameters into the full lifecycle dataset. The system automatically triggered a dataset update event, re-executing semantic mapping processing, manifold dimensionality reduction processing, state recognition, and maintenance decision calculation. The updated 3D spatial coordinates were 1.32, 1.85, and 0.67, with an Euclidean distance of 2.38, indicating the mold had re-entered the normal operating region. The current state label output by the cumulative damage state recognition model was normal operating, with a posterior probability of 0.89. Technicians continued to monitor the mold's operating status and, in subsequent... During the cycle, the mold always operated between the normal operating range and the light wear range, and no more moderate fatigue or serious failure conditions occurred.

[0102] The advancement of this invention over traditional methods lies in the fact that traditional mold maintenance methods typically trigger maintenance operations based on fixed cycles or experience thresholds, failing to dynamically adjust maintenance strategies according to the actual cumulative damage state of the mold, leading to untimely or excessive maintenance. This invention achieves semantic unification of parameters across stages through an industrial knowledge graph, linking heterogeneous data from design, processing, production, and maintenance stages into complete end-to-end features. This allows state recognition to comprehensively consider information throughout the entire lifecycle of the mold from design to operation, avoiding the one-sided judgments caused by traditional methods relying solely on single-stage data. Furthermore, this invention employs a sparse connection learning framework based on dynamic topology reconstruction to construct a cumulative damage state recognition model. By adaptively adjusting network connections using information theory criteria, it reduces computational complexity, enabling the model to be deployed on edge devices for real-time inference. Additionally, by using long short-term memory units to compress the historical dependencies of non-Markov processes into a fixed-dimensional latent variable vector, it solves the problem that traditional methods cannot effectively handle long-term cumulative damage features. This invention employs a two-layer game optimization model for maintenance decisions. The upper layer aims to minimize the rate of state degradation, while the lower layer aims to maximize equipment availability. By using an iterative solution algorithm to find the Nash equilibrium solution, it achieves coordinated optimization of maintenance timing and intensity. This avoids the contradiction between maintaining equipment performance and production efficiency that traditional single-objective optimization cannot balance, making maintenance resource allocation more rational.

[0103] It should be noted that the variables involved in this invention are explained in detail in Tables 2 and 3.

[0104] Table 2. Variable Explanation Table (Part 1)

[0105]

[0106] Table 3. Variable Explanation Table (Part Two)

[0107]

[0108] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for managing the entire lifecycle state machine of a mold, characterized in that, Parameters of the mold are collected during the design, processing, production, and maintenance stages and stored uniformly in a full lifecycle dataset. The full lifecycle dataset is input into a semantic mapping layer for cross-stage semantic unification processing, outputting a unified semantic full-link parameter set. The unified semantic full-link parameter set is then subjected to manifold dimensionality reduction processing and mapped to a three-dimensional topological space, dividing it into normal operation area, light wear area, moderate fatigue area, and severe fault area. The three-dimensional spatial coordinate set and historical state sequence are input into a cumulative damage state recognition model, outputting the current state label and historical dependency feature vector. The state transition trend is determined based on the Euclidean distance between the historical dependency feature vector and the boundary of each region. When the Euclidean distance is less than a preset threshold, a two-layer game optimization model is activated to output maintenance timing parameters and maintenance intensity parameters. After the maintenance operation is performed, the full lifecycle dataset is updated and semantic unification processing is re-executed to achieve closed-loop flow of the state machine.

2. The mold management full lifecycle state machine flow method according to claim 1, characterized in that, The semantic mapping layer is built on an industrial knowledge graph, which defines a standard ontology structure for the entire lifecycle of a mold. The standard ontology structure includes entity nodes and relation edges. Entity nodes include design entity nodes, processing entity nodes, production entity nodes, and maintenance entity nodes. Relation edges define the causal relationships and semantic mapping rules between entity nodes at different stages.

3. The mold management full lifecycle state machine flow method according to claim 2, characterized in that, The semantic mapping layer uses natural language processing technology to parse unstructured data text at each stage to extract parameter names and numerical units, maps parameter names to entity nodes in the industrial knowledge graph, establishes equivalence edges for the same physical quantity at different stages in the industrial knowledge graph, and establishes a correspondence between the expected number of cycles in the design stage and the actual number of cycles in the maintenance stage through a graph query algorithm.

4. The mold management full lifecycle state machine flow method according to claim 3, characterized in that, The manifold dimensionality reduction process employs an incremental t-SNE nonlinear dimensionality reduction algorithm. This algorithm calculates the conditional probability distribution between parameter points in the high-dimensional parameter space and then searches for the corresponding probability distribution in the low-dimensional topological space to minimize the KL divergence between the two. The lightweight encoder network is then trained to map the unified semantic full-link parameter set into a three-dimensional spatial coordinate set.

5. The mold management full lifecycle state machine flow method according to claim 4, characterized in that, The lightweight encoder network consists of an input layer, three hidden layers, and an output layer. The input layer receives a unified semantic full-link parameter set. The three hidden layers reduce the parameter dimensions to 64, 32, and 16 dimensions respectively. The output layer outputs a three-dimensional spatial coordinate set. The lightweight encoder network is deployed on an edge computing device to perform online dimensionality reduction mapping on the real-time acquired unified semantic full-link parameter set.

6. The mold management full lifecycle state machine flow method according to claim 5, characterized in that, The normal operation area, light wear area, moderate fatigue area and severe failure area in the three-dimensional topological space are determined by geometric boundaries. The normal operation area is defined as the area where the Euclidean distance of the coordinate point in the three-dimensional spatial coordinate set from the origin is less than the first threshold. The light wear area is defined as the area where the Euclidean distance is greater than or equal to the first threshold and less than the second threshold.

7. The mold management full lifecycle state machine flow method according to claim 6, characterized in that, The cumulative damage state recognition model utilizes a sparse connection learning framework based on dynamic topology reconstruction. During network training, it adaptively adds or removes connection edges according to the statistical distribution of neuron activation patterns. It evaluates connection importance using information theory criteria and implements probabilistic connection pruning. It also introduces grouping regularization constraints on connection weights to maintain the network's expressive power.

8. The mold management full lifecycle state machine flow method according to claim 7, characterized in that, The cumulative damage state recognition model is a cascaded combination of an input layer, a long short-term memory (LSTM) layer, a sparse connection layer, a Bayesian inference layer, and an output layer. The LSTM layer contains 128 LSTM units, each with an input gate, a forget gate, and an output gate. The LSTM layer compresses the cumulative damage features in the historical state sequence into a 64-dimensional latent variable vector.

9. The mold management full lifecycle state machine flow method according to claim 8, characterized in that, The sparse connection layer receives a 64-dimensional latent variable vector and a three-dimensional spatial coordinate set. During training, the importance score of each connection edge is calculated based on the statistical distribution of neuron activation frequency. When the importance score of a connection edge is lower than a dynamic threshold, the connection edge is pruned and removed. When the prediction error of a certain region continues to increase, new connection edges are added between the neurons corresponding to the region.

10. The mold management full lifecycle state machine flow method according to claim 9, characterized in that, The connection weights of the sparse connection layer are grouped according to the functional group to which the neurons belong, and an application is made to the connection weights of each group. Regularization constraints ensure that the connection weights within the same functional group maintain a similar numerical range. The Bayesian inference layer calculates the posterior probability of the current state belonging to each region using the Bayesian formula and selects the region type with the highest posterior probability as the label of the current state.

Images