Mold life prediction and maintenance decision-making method based on industrial big data

Abstract

The invention relates to the technical field of industrial big data analysis and product life cycle management, and discloses a mold life prediction and maintenance decision-making method based on industrial big data. The method comprises the following steps: collecting mold multi-dimensional operation state data and order task information; performing space-time alignment and time sequence normalization processing; a deep residual convolutional neural network and a gating circulation unit are used for joint modeling, and the remaining service life of the mold is predicted; constructing a double-layer maintenance decision space taking an order delivery window as a hard boundary and a failure risk as a soft constraint; and solving the optimal maintenance intervention opportunity by adopting an improved particle swarm optimization algorithm, generating a maintenance decision instruction and synchronously updating the production schedule. According to the invention, high-precision life prediction and maintenance-production cooperative scheduling are realized, non-planned shutdown can be reduced by more than 45%, the maintenance cost is reduced by 30%, and the on-time delivery rate of key orders is ensured not to be lower than 99%.

Description

Technical Field

[0001] This invention belongs to the field of industrial big data analysis and product lifecycle management technology, specifically involving a method for predicting mold life and making maintenance decisions based on industrial big data. Background Technology

[0002] With the deepening of intelligent manufacturing and Industry 4.0, molds, as core process equipment in high-end manufacturing, directly affect product quality, production efficiency, and delivery cycle through their operational status. In modern flexible manufacturing systems, molds are widely used in key processes such as injection molding, stamping, and die casting, and their performance degradation exhibits high nonlinearity, time-varying characteristics, and condition dependence. Traditional mold maintenance strategies are mostly based on fixed cycles or empirical thresholds, relying on offline detection and static life models, which are difficult to reflect the dynamic evolution characteristics of material fatigue, thermal stress accumulation, and wear rate during actual processing, leading to delayed maintenance decisions or excessive intervention.

[0003] Mold life prediction technology based on industrial big data has become a research hotspot in recent years. It estimates remaining service life by collecting multi-source sensor data such as equipment vibration, temperature, pressure, and process parameters, combined with machine learning models. However, existing methods generally separate life prediction from maintenance decision-making: the prediction module only outputs a single health indicator, while maintenance scheduling still relies on rule-driven or optimization-based solutions, lacking real-time response capabilities to changes in production plans, adjustments to order urgency, and the collaborative status of equipment clusters. Especially in production scenarios with multiple varieties, small batches, and high mixing, static maintenance strategies cannot align with dynamic scheduling goals, often resulting in wasted capacity, delivery delays, or sudden downtime.

[0004] Existing technologies face three major bottlenecks when constructing an integrated prediction-decision framework: First, individual prediction models neglect the mutual influence between molds in shared production lines and shared resources, making it difficult to characterize the group degradation coupling effect; second, maintenance decisions are not embedded in the game environment of production scheduling, failing to balance the dynamic equilibrium between "continued processing benefits" and "failure risk costs"; and third, there is a lack of mechanisms to simulate the adaptive evolution of different maintenance strategies within a group of equipment, leading to strategy rigidity and convergence to local suboptimal solutions. Therefore, in highly dynamic and strongly coupled intelligent manufacturing environments, there is an urgent need for a mold maintenance decision-making method that can integrate multi-agent collaborative learning and strategy evolution mechanisms to achieve deep coupling and global optimization of lifespan prediction, health assessment, and production scheduling. Summary of the Invention

[0005] This invention provides a method for mold life prediction and maintenance decision-making based on industrial big data. By integrating multi-source heterogeneous production operation data, equipment status monitoring data, and order task information, it constructs a dynamically coupled mold health status evolution model. On this basis, it introduces a task priority-driven maintenance window optimization mechanism to achieve high-precision online prediction of the remaining service life of the mold and real-time collaborative scheduling of maintenance decisions. This solves the problem of maintenance and production disconnect caused by static and isolated maintenance strategies being unable to adapt to dynamically changing production plans, order priorities, and real-time equipment status.

[0006] This invention provides a method for mold life prediction and maintenance decision-making based on industrial big data, which includes: Collect multi-dimensional operational status data of the mold during its service process. The multi-dimensional operational status data includes mold temperature field distribution data, historical stress and strain data, cumulative stamping number data, surface wear image data, and lubrication status monitoring data. The system synchronously acquires current production plan data and order task information. The production plan data includes the production sequence of each batch of products, the set values ​​of process parameters, and the estimated processing time. The order task information includes the customer delivery deadline, the order urgency level indicator, and the product quality level requirements. The multi-dimensional operating status data is subjected to spatiotemporal alignment and time-series normalization to generate a standardized mold state time-series matrix. Based on the mold state time series matrix, the micro-damage accumulation features of the mold are extracted using a deep residual convolutional neural network, and the nonlinear degradation trajectory of the mold health state is modeled by a gated recurrent unit network, outputting the continuous numerical prediction results of the remaining service life of the mold. The remaining useful life prediction results are mapped with the order task information to construct a two-layer maintenance decision space with the order delivery window as the hard boundary and the mold failure risk threshold as the soft constraint. Within the dual-layer maintenance decision space, an improved particle swarm optimization algorithm is used to solve for the optimal maintenance intervention time that satisfies both production continuity and equipment reliability objectives, generating maintenance decision instructions that include maintenance type, execution time period, and resource allocation scheme. The maintenance decision command is pushed to the manufacturing execution system, triggering the generation of maintenance work orders and synchronously updating the production schedule plan.

[0007] Preferably, the multi-dimensional operational status data of the collected mold during its service life specifically includes: The three-dimensional temperature field distribution data of the mold during the stamping cycle is acquired in real time by a distributed fiber optic grating sensor array embedded inside the mold body, with a sampling frequency of no less than 10 times per second. The strain gauge assembly and piezoelectric force sensor installed on the mold support structure synchronously collect the principal stress direction components, maximum principal strain value and load peak data of the mold in each stamping action. The mold uses a counter module to record the cumulative number of stampings since the last major overhaul, and automatically associates the corresponding standard stamping frequency threshold with the product model. The industrial vision camera deployed above the working area of ​​the stamping machine periodically captures high-definition images of the mold working surface. The image resolution is no less than 4 million pixels, and the shooting interval is dynamically adjusted according to the current production cycle, ranging from once every 5 to 30 minutes. The viscosity of the lubricating oil, the concentration of metal abrasive particles, and the moisture content of the mold lubrication system are detected by an online oil monitoring device to form a quantitative index of lubrication status.

[0008] Preferably, the spatiotemporal alignment and time-series normalization processing of the multi-dimensional operational status data specifically includes: Using the completion time of the stamping action as the baseline event, data streams from different sensors are aligned using event-triggered timestamps. For temperature field data and strain data with sampling frequencies higher than the reference event frequency, a sliding window averaging method is used to compress them to be synchronized with the stamping cycle. For visual image data and oil monitoring data with sampling frequencies lower than the baseline event frequency, cubic spline interpolation is used to generate estimated values ​​at the missing time points. All aligned data are organized into a two-dimensional matrix based on the stamping number index, with the row dimension representing the time step and the column dimension representing the characteristic channels of different physical quantities. Min-max normalization is performed independently for each feature channel, mapping the original values ​​to the interval between 0 and 1.

[0009] Preferably, the extraction of microscopic damage accumulation features of the mold using a deep residual convolutional neural network specifically includes: The standardized mold state timing matrix is ​​input into a convolutional encoder consisting of 8 stacked residual blocks. Each residual block contains 2 one-dimensional convolutional layers with a kernel size of 3, and the activation function is a modified linear unit. A global average pooling layer is introduced at the end of the convolutional encoder to compress the temporal feature map into a fixed-length damage feature vector; The gated recurrent unit network receives the damage feature vector as an input sequence, updates the hidden state step by step over time, and finally outputs the mold health index sequence. The mold health index sequence is mapped to a continuous predicted value of the remaining number of stampings via a fully connected regression head, and this predicted value is the remaining service life of the mold.

[0010] Preferably, the construction of the two-layer maintenance decision space, with the order delivery window as the hard boundary and the mold failure risk threshold as the soft constraint, specifically includes: The mold failure risk threshold is defined as a remaining service life prediction that is less than 20% of the number of stampings required for the current order; Arrange all pending orders in ascending order of delivery deadline to form a task queue; For each order in the queue, calculate its earliest start time and latest end time to form the time feasible window for that order; If the predicted remaining service life of the mold is consistently higher than the failure risk threshold within the feasible time window, the order is marked as a low-risk task. If there is a point in time when the predicted remaining useful life is lower than the failure risk threshold, then a mandatory maintenance checkpoint is inserted before that point in time to form a set of maintenance candidate windows.

[0011] Preferably, the step of using an improved particle swarm optimization algorithm to solve for the optimal maintenance intervention timing specifically includes: Encode the position vector of each particle into a set of maintenance intervention time points, with the dimension equal to the number of maintenance candidate windows; The fitness function is defined as a weighted sum, where the first term is the production delay cost caused by maintenance, the second term is the integral expectation of the probability of sudden mold failure, and the weighting coefficients are dynamically adjusted according to the urgency level of the order. An inertial weight adjustment factor based on the mold degradation rate is introduced into the particle velocity update formula. When the degradation rate exceeds a preset threshold, the inertial weight is increased to enhance the global search capability. The maximum number of iterations is set to 200 generations, the population size is 50 particles, and the process is terminated early when the fitness value changes by less than 1‰ for 30 consecutive generations. The particle with the best fitness is selected from the final population and decoded into the actual maintenance intervention time series.

[0012] Furthermore, the maintenance types include three levels: preventative replacement, partial repair, and complete overhaul, and the selection criteria are as follows: When the predicted remaining service life is greater than 60% of the standard service life and no cracks are detected in the surface wear image, local repair is performed; When the remaining service life prediction is between 30% and 60% of the standard service life, or when a micron-sized crack is detected but does not penetrate the working surface, preventive replacement shall be performed. A full overhaul shall be performed when the predicted remaining service life is less than 30% of the standard service life, or when the stress and strain data show abnormal changes, or when the lubrication condition index exceeds the safety limit three times in a row.

[0013] This invention provides a mold life prediction and maintenance decision-making system based on industrial big data, which includes: The multi-source data acquisition module is used to collect multi-dimensional operating status data, current production plan data, and order task information of the mold during its service process; The data preprocessing module is used to perform spatiotemporal alignment and time-series normalization on the multi-dimensional operating status data to generate a standardized mold state time-series matrix. The lifespan prediction modeling module is used to jointly model the mold state time series matrix using a deep residual convolutional neural network and a gated recurrent unit network, and output continuous numerical prediction results of the remaining lifespan of the mold. The task constraint mapping module is used to map the remaining useful life prediction results with the order task information to construct a two-layer maintenance decision space; The maintenance timing optimization module is used to solve for the optimal maintenance intervention timing using an improved particle swarm optimization algorithm within the two-layer maintenance decision space, and generate maintenance decision instructions. The instruction execution and feedback module is used to push the maintenance decision instruction to the manufacturing execution system, trigger the generation of maintenance work orders and synchronously update the production schedule plan, and at the same time receive the actual effect data after maintenance execution for online model correction.

[0014] Preferably, the lifetime prediction modeling module further includes: The convolutional feature extraction submodule, consisting of multiple stacked residual blocks, is used to extract joint spatial-temporal features from the temporal matrix of the mold state. The degradation trajectory modeling submodule adopts a gated recurrent unit structure to perform time-series modeling on the extracted feature sequences and output the health index evolution curve; The regression output submodule, consisting of fully connected layers, is used to map the health index to a numerical prediction of the remaining number of stampings.

[0015] Preferably, the maintenance timing optimization module further includes: The particle coding submodule is used to encode maintenance intervention time points as search variables for the optimization algorithm; The fitness evaluation submodule is used to calculate the comprehensive cost of each candidate solution based on the production delay cost and the probability of failure risk. The dynamic parameter adjustment submodule is used to adjust the control parameters of the optimization algorithm in real time according to the current degradation rate of the mold; The decoding output submodule is used to decode the optimal particle into a specific maintenance execution schedule and resource requirement list.

[0016] Preferably, the instruction execution and feedback module further includes: The work order generation submodule is used to automatically generate electronic work orders based on maintenance decision instructions, including maintenance type, required spare parts, technical personnel qualification requirements, and estimated working hours. The scheduling synchronization submodule is used to write the maintenance-occupied time period into the master production plan database of the manufacturing execution system, triggering the automatic rescheduling of adjacent orders; The model calibration submodule is used to collect the actual replacement mold component status data and the prediction results after maintenance is completed, perform deviation analysis, and use the deviation data to fine-tune the output layer weights of the life prediction model online.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a data-driven closed-loop system for mold life prediction and maintenance decision-making by deeply integrating real-time mold operating status, dynamic production plans, and order task priorities. This system abandons the traditional static maintenance model based on fixed cycles or simple thresholds, achieving real-time collaboration between maintenance decisions and production scheduling.

[0018] 2. The joint network structure of deep residual convolution and gated recurrent units can effectively capture the spatiotemporal evolution of micro-damage in molds, significantly improving the accuracy and robustness of remaining service life prediction. The task constraint mapping mechanism transforms the urgency of orders at the business level into maintenance constraints at the technical level, ensuring the production continuity of high-priority orders. The improved particle swarm optimization algorithm takes into account both production efficiency loss and equipment failure risk when solving for maintenance timing, achieving Pareto optimality in both economy and reliability. Practical applications show that this invention can reduce unplanned mold downtime by more than 45%, reduce maintenance costs by 30%, and ensure that the on-time delivery rate of critical orders is no less than 99%. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the mold health state degradation trajectory jointly modeled by depth residual convolution and gated recurrent unit in this invention; Figure 3 This is a logical flowchart of the multi-source heterogeneous data acquisition and spatiotemporal alignment preprocessing in this invention; Figure 4 This is a logical framework diagram of the two-layer maintenance decision space construction driven by task priority in this invention; Figure 5 This is a flowchart illustrating the logical process of the improved particle swarm optimization algorithm in this invention for determining the optimal maintenance intervention timing. Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between the mold life prediction and maintenance decision system and the manufacturing execution system in this invention. Detailed Implementation

[0020] This invention provides a method for mold life prediction and maintenance decision-making based on industrial big data. Its core lies in constructing a dynamically coupled mold health state evolution model by integrating multi-source heterogeneous production operation data, equipment status monitoring data, and order task information. Based on this, a task priority-driven maintenance window optimization mechanism is introduced to achieve high-precision online prediction of the remaining mold life and real-time collaborative scheduling of maintenance decisions. The following will be combined with the appendix... Figure 1 To be continued Figure 6 This section provides a detailed implementation description of each functional module of the system, expanding upon it layer by layer.

[0021] The method includes the following steps: S1, collecting multi-dimensional operational status data of the mold during its service life; S2, simultaneously acquiring current production plan data and order task information; S3, performing spatiotemporal alignment and time-series normalization processing on the multi-dimensional operational status data to generate a standardized mold status time-series matrix; S4, based on the mold status time-series matrix, using a deep residual convolutional neural network to extract the cumulative features of microscopic damage in the mold, and combining a gated recurrent unit network to model the nonlinear degradation trajectory of the mold's health status, outputting continuous numerical prediction results of the mold's remaining service life; S5, mapping the prediction results of the remaining service life with the order task information to construct a two-layer maintenance decision space with the order delivery window as the hard boundary and the mold failure risk threshold as the soft constraint; S6, within the two-layer maintenance decision space, using an improved particle swarm optimization algorithm to solve for the optimal maintenance intervention time that satisfies both production continuity and equipment reliability objectives, generating maintenance decision instructions that include maintenance type, execution period, and resource allocation scheme; S7, pushing the maintenance decision instructions to the manufacturing execution system to trigger the generation of maintenance work orders and synchronously update the production schedule plan.

[0022] In step S1, multi-dimensional operational status data of the mold during its service life is collected. This multi-dimensional operational status data includes mold temperature field distribution data, historical stress and strain data, cumulative stamping count data, surface wear image data, and lubrication status monitoring data. Specifically, a distributed fiber optic grating sensor array embedded inside the mold body acquires the three-dimensional temperature field distribution data of the mold in real time during the stamping cycle, with a sampling frequency of no less than 10 times per second. This sensor array is deployed along key thermally sensitive areas of the mold, such as the cavity wall and the periphery of cooling channels. Each sensing point independently records local temperature changes, forming a dynamic thermal image sequence with a spatial resolution of millimeters.

[0023] Simultaneously, strain gauges and piezoelectric force sensors mounted on the mold support structure synchronously collect data on the principal stress direction components, maximum principal strain values, and peak loads of the mold during each stamping action. The strain gauges are arranged orthogonally, covering the mold base, side walls, and ejection mechanism connection to ensure the capture of anisotropic deformation responses. The piezoelectric force sensor is integrated between the press slide and the upper die, directly measuring the instantaneous amplitude and duration of the impact load. Furthermore, a mold usage counter module records the cumulative number of stamping operations since the last major overhaul. This counter is linked to the press's main control system, automatically incrementing by 1 after each complete stroke and automatically associating with the corresponding standard stamping frequency threshold based on the product model, achieving lifespan benchmark calibration for different product specifications.

[0024] Furthermore, an industrial vision camera deployed above the working area of ​​the stamping machine periodically captures high-definition images of the mold's working surface. The image resolution is no less than 4 megapixels, and the shooting interval is dynamically adjusted according to the current production cycle, ranging from once every 5 to 30 minutes. The camera is equipped with a ring LED light source and a polarizing filter to suppress interference from metal reflections, ensuring clear imaging of surface microcracks, peeling, or sticking defects. Finally, an online oil monitoring device detects the viscosity of the lubricating oil, the concentration of metal abrasive particles, and the moisture content of the mold's lubrication system, forming a quantitative index of lubrication status. This device incorporates a micro-fluid channel and a multi-parameter sensor chip to analyze the degree of oil deterioration in the return oil line in real time and outputs a numerical sequence of three core indicators in digital signal form.

[0025] In step S2, current production plan data and order task information are acquired synchronously. The production plan data includes the production schedule, process parameter settings, and estimated processing time for each batch of products. The production schedule is exported from the manufacturing execution system's main database, with time slots allocated to the minute level. The process parameter settings cover key control variables such as punching force, closing height, holding time, and cooling rate. The estimated processing time is calculated based on the historical average cycle time of similar products and the current equipment efficiency coefficient. The order task information includes the customer delivery deadline, order urgency level identifier, and product quality level requirements. The delivery deadline is stored in a standard timestamp format. The urgency level identifier uses a three-level coding system: Level 1 is "Urgent," Level 2 is "Regular," and Level 3 is "Delayable." Product quality level requirements are divided into three categories: "High Precision," "Standard," and "Trial Production," corresponding to different mold surface roughness tolerances and dimensional tolerance zones.

[0026] In step S3, the multi-dimensional operating status data undergoes spatiotemporal alignment and time-series normalization to generate a standardized mold state time-series matrix. First, using the completion time of the stamping action as the baseline event, event-triggered timestamp alignment is performed on the data streams from different sensors. All sensor data carries a hardware-level timestamp, and the system's main control unit broadcasts a synchronization pulse at the end of each stroke. Each acquisition module then adds a unified logical time tag to its locally cached data accordingly. For temperature field data and strain data with sampling frequencies higher than the baseline event frequency, a sliding window averaging method is used to compress them to be synchronized with the stamping cycle. The sliding window length is equal to the time interval between two adjacent strokes, and the arithmetic mean of all sampling points within the window is taken as the representative value for that stroke. For visual image data and oil monitoring data with sampling frequencies lower than the baseline event frequency, cubic spline interpolation is used to generate estimated values ​​at missing time points. The interpolation nodes are determined by the already acquired valid data points, ensuring that the generated sequence is continuous and smooth in the time dimension.

[0027] Subsequently, all aligned data were organized into a two-dimensional matrix based on the number of stamping operations. The row dimension represents the time step, with each row being a snapshot of the overall state of a complete stamping stroke. The column dimension represents the characteristic channels for different physical quantities, including nine channels: mean temperature, temperature gradient, principal stress amplitude, maximum principal strain, cumulative number of stamping operations, image grayscale entropy, lubricating oil viscosity, metal abrasive concentration, and moisture content. Finally, min-max normalization was performed independently on each characteristic channel, mapping the original values ​​to the 0-1 range. The normalization formula is:

[0028] in, These are the original eigenvalues. and These are the minimum and maximum values ​​of the channel in the entire historical dataset, which are determined in advance through offline statistics and fixed in the preprocessing configuration file.

[0029] In step S4, based on the mold state time-series matrix, a deep residual convolutional neural network is used to extract the micro-damage accumulation features of the mold, and a gated recurrent unit network is combined to model the nonlinear degradation trajectory of the mold's health state, outputting a continuous numerical prediction result of the mold's remaining service life. Specifically, the standardized mold state time-series matrix is ​​input into a convolutional encoder composed of 8 stacked residual blocks. Each residual block contains two 1D convolutional layers with a kernel size of 3, and the activation function is a modified linear unit. The input sequence length is a snapshot of the state of the most recent 200 strokes. After the first convolutional layer, the number of channels is expanded to 64, and the number of channels doubles every two residual blocks thereafter, resulting in a final output feature map with a dimension of 25×512. A global average pooling layer is introduced at the end of the convolutional encoder to compress the time-series feature map into a fixed-length damage feature vector with a dimension of 512. This vector represents the microstructural damage patterns accumulated by the mold during recent service, such as the tendency of thermal fatigue crack initiation, concentrated areas of plastic deformation, and precursors of lubrication failure. Subsequently, the gated recurrent unit network receives the damage feature vector as an input sequence and updates the hidden state step by step. The gated recurrent unit includes a reset gate and an update gate, mathematically expressed as follows:

[0030] in, For the first Damage feature vectors at each time step, For the corresponding hidden state, For the sigmoid function, This represents element-wise multiplication. After 20 time steps of recursion, the final output is a mold health index sequence, which monotonically decreases, reflecting the continuous degradation of mold performance. This mold health index sequence is mapped to a continuous predicted value of the remaining number of stamping operations via a fully connected regression head; this predicted value represents the remaining service life of the mold. The regression head consists of two fully connected layers: the middle layer has a dimension of 128, and the output layer has a dimension of 1, with a linear activation function.

[0031] In step S5, the remaining service life prediction result is mapped to the order task information to construct a two-layer maintenance decision space with the order delivery window as the hard boundary and the mold failure risk threshold as the soft constraint. First, the mold failure risk threshold is defined as the remaining service life prediction value being less than 20% of the number of stampings required for the current order. This threshold is derived from historical failure case statistics to ensure sufficient maintenance window is reserved before the risk occurs. Subsequently, all orders to be executed are arranged in ascending order of delivery deadline to form a task queue. For each order in the queue, its earliest start time and latest end time are calculated to constitute the time feasible window for that order. The earliest start time is the larger of the current system time or the end time of the previous order; the latest end time is the delivery deadline minus the safety buffer period. The buffer period is set according to the product quality level: 4 hours for high-precision products, 2 hours for standard products, and no buffer for trial production products. If the remaining service life prediction value of the mold is always higher than the failure risk threshold within the time feasible window, the order is marked as a low-risk task and normal production scheduling is allowed. If there exists a point in time when the predicted remaining useful life falls below the failure risk threshold, a mandatory maintenance checkpoint is inserted before that point, forming a set of maintenance candidate windows. The start time of each candidate window is the moment the risk first occurs minus the maintenance preparation time (usually 30 minutes), and the end time is the latest time that the order can start.

[0032] In step S6, within the two-layer maintenance decision space, an improved particle swarm optimization algorithm is used to solve for the optimal maintenance intervention timing that satisfies both production continuity and equipment reliability objectives, generating maintenance decision instructions that include maintenance type, execution time period, and resource allocation scheme. The position vector of each particle is encoded as a set of maintenance intervention time points, with the dimension equal to the number of maintenance candidate windows. The value range of each dimension is limited to the start and end times of the corresponding candidate window to ensure the feasibility of the solution. The fitness function is defined as a weighted sum form:

[0033] in, The cost of production delays caused by maintenance is calculated as the sum of the products of the delay time of all affected orders and their urgency level weights. The integral expectation of the probability of sudden mold failure is obtained by integrating the failure probability density function of the remaining service life prediction curve under maintenance-free conditions; weighting coefficients. and The queue is dynamically adjusted based on the urgency level of the orders. When there are "urgent" orders in the queue, Increased to 0.7, The inertial weight is reduced to 0.3; otherwise, both are 0.5. An inertial weight adjustment factor based on the die degradation rate is introduced into the particle velocity update formula. The degradation rate is defined as the average decline slope of the health index over the last 10 strokes. When this slope exceeds a preset threshold (e.g., 0.005 per stroke), the inertial weight increases from the base value of 0.7 to 0.9 to enhance global search capability and avoid getting trapped in local optima.

[0034] The maximum number of iterations is set to 200 generations, and the population size is 50 particles. The process terminates early when the fitness value changes by less than 1‰ over 30 consecutive generations. The particle with the best fitness is selected from the final population and decoded into the actual maintenance intervention time series. The maintenance type is determined based on the predicted remaining service life and surface wear image analysis results: local repair is performed when the predicted remaining service life is greater than 60% of the standard service life and no cracks are detected in the surface wear image; preventative replacement is performed when the predicted remaining service life is between 30% and 60% of the standard service life, or when micron-level cracks are detected but do not penetrate the working surface; and a full overhaul is performed when the predicted remaining service life is less than 30% of the standard service life, or when stress-strain data shows abnormal abrupt changes, or when lubrication status indicators exceed safety limits three times consecutively.

[0035] In step S7, the maintenance decision instruction is pushed to the Manufacturing Execution System (MES), triggering the generation of a maintenance work order and synchronously updating the production schedule. The work order includes the maintenance type, a list of required spare parts, the qualification requirements for technical personnel, and the estimated working hours. Upon receiving the instruction, the MES automatically locks the relevant mold resources, generates an electronic work order, and assigns it to the maintenance team terminal. Simultaneously, the scheduling synchronization submodule writes the maintenance-occupied time period into the master production plan database, triggering automatic rescheduling of adjacent orders to ensure overall load balance on the production line. After maintenance is completed, the model calibration submodule collects the status data of the actually replaced mold components, including the initial health index of the old and new molds, the actual number of strokes in service, and the failure mode description. It performs deviation analysis with the prediction results and uses this deviation data to fine-tune the output layer weights of the life prediction model online. A mini-batch gradient descent method is used, with a learning rate set to 0.001, updating only the regression head parameters to maintain the stability of the feature extraction part.

[0036] The aforementioned method, through a data-driven approach, transforms mold maintenance from passive response to proactive prediction, effectively addressing the disconnect between static maintenance strategies and dynamic production demands. At the system level, this invention also provides a mold life prediction and maintenance decision-making system based on industrial big data, comprising a multi-source data acquisition module, a data preprocessing module, a life prediction modeling module, a task constraint mapping module, a maintenance timing optimization module, and an instruction execution and feedback module. The multi-source data acquisition module integrates various sensor interfaces and the manufacturing execution system data bus, achieving millisecond-level data aggregation; the data preprocessing module is deployed on edge computing nodes to complete real-time alignment and normalization; the life prediction modeling module runs on a cloud training server, supporting regular model retraining and online inference; the task constraint mapping module interfaces with the enterprise resource planning system to obtain the latest order status; the maintenance timing optimization module adopts a parallel computing architecture to accelerate particle swarm optimization; and the instruction execution and feedback module communicates bidirectionally with the manufacturing execution system through a message queue to ensure closed-loop instruction execution. Compared with existing technologies, this system significantly improves the intelligence level of mold management and achieves synergistic optimization of production efficiency and equipment reliability.

Claims

1. A method for predicting mold life and making maintenance decisions based on industrial big data, characterized in that, include: Collect multi-dimensional operational status data of the mold during its service process. The multi-dimensional operational status data includes mold temperature field distribution data, historical stress and strain data, cumulative stamping number data, surface wear image data, and lubrication status monitoring data. The system synchronously acquires current production plan data and order task information. The production plan data includes the production sequence of each batch of products, the set values ​​of process parameters, and the estimated processing time. The order task information includes the customer delivery deadline, the order urgency level indicator, and the product quality level requirements. The multi-dimensional operating status data is subjected to spatiotemporal alignment and time-series normalization to generate a standardized mold state time-series matrix. Based on the mold state time series matrix, the micro-damage accumulation features of the mold are extracted using a deep residual convolutional neural network, and the nonlinear degradation trajectory of the mold health state is modeled by a gated recurrent unit network, outputting the continuous numerical prediction results of the remaining service life of the mold. The remaining useful life prediction results are mapped with the order task information to construct a two-layer maintenance decision space with the order delivery window as the hard boundary and the mold failure risk threshold as the soft constraint. Within the dual-layer maintenance decision space, an improved particle swarm optimization algorithm is used to solve for the optimal maintenance intervention time that satisfies both production continuity and equipment reliability objectives, generating maintenance decision instructions that include maintenance type, execution time period, and resource allocation scheme. The maintenance decision command is pushed to the manufacturing execution system, triggering the generation of maintenance work orders and synchronously updating the production schedule plan.

2. The mold life prediction and maintenance decision-making method based on industrial big data according to claim 1, characterized in that, Collect multi-dimensional operational status data of the mold during its service life, including: The three-dimensional temperature field distribution data of the mold during the stamping cycle is acquired in real time by a distributed fiber optic grating sensor array embedded inside the mold body, with a sampling frequency of no less than 10 times per second. The strain gauge assembly and piezoelectric force sensor installed on the mold support structure synchronously collect the principal stress direction components, maximum principal strain value and load peak data of the mold in each stamping action. The mold uses a counter module to record the cumulative number of stampings since the last major overhaul, and automatically associates the corresponding standard stamping frequency threshold with the product model. The industrial vision camera deployed above the working area of ​​the stamping machine periodically captures high-definition images of the mold working surface. The image resolution is no less than 4 million pixels, and the shooting interval is dynamically adjusted according to the current production cycle, ranging from once every 5 to 30 minutes. The viscosity of the lubricating oil, the concentration of metal abrasive particles, and the moisture content of the mold lubrication system are detected by an online oil monitoring device to form a quantitative index of lubrication status.

3. The mold life prediction and maintenance decision-making method based on industrial big data according to claim 2, characterized in that, The multi-dimensional operational status data undergoes spatiotemporal alignment and time-series normalization processing, including: Using the completion time of the stamping action as the baseline event, data streams from different sensors are aligned using event-triggered timestamps. For temperature field data and strain data with sampling frequencies higher than the reference event frequency, a sliding window averaging method is used to compress them to be synchronized with the stamping cycle. For visual image data and oil monitoring data with sampling frequencies lower than the baseline event frequency, cubic spline interpolation is used to generate estimated values ​​at the missing time points. All aligned data are organized into a two-dimensional matrix based on the stamping number index, with the row dimension representing the time step and the column dimension representing the characteristic channels of different physical quantities. Min-max normalization is performed independently for each feature channel, mapping the original values ​​to the interval between 0 and 1.

4. The mold life prediction and maintenance decision-making method based on industrial big data according to claim 3, characterized in that, Deep residual convolutional neural networks are used to extract microscopic damage accumulation features of molds, including: The standardized mold state timing matrix is ​​input into a convolutional encoder consisting of 8 stacked residual blocks. Each residual block contains 2 one-dimensional convolutional layers with a kernel size of 3, and the activation function is a modified linear unit. A global average pooling layer is introduced at the end of the convolutional encoder to compress the temporal feature map into a fixed-length damage feature vector; The gated recurrent unit network receives the damage feature vector as an input sequence, updates the hidden state step by step over time, and finally outputs the mold health index sequence. The mold health index sequence is mapped to a continuous predicted value of the remaining number of stampings via a fully connected regression head, and this predicted value is the remaining service life of the mold.

5. The mold life prediction and maintenance decision-making method based on industrial big data according to claim 4, characterized in that, A two-layer maintenance decision space is constructed, with the order delivery window as the hard boundary and the mold failure risk threshold as the soft constraint, including: The mold failure risk threshold is defined as a remaining service life prediction that is less than 20% of the number of stampings required for the current order; Arrange all pending orders in ascending order of delivery deadline to form a task queue; For each order in the queue, calculate its earliest start time and latest end time to form the time feasible window for that order; If the predicted remaining service life of the mold is consistently higher than the failure risk threshold within the feasible time window, the order is marked as a low-risk task. If there is a point in time when the predicted remaining useful life is lower than the failure risk threshold, then a mandatory maintenance checkpoint is inserted before that point in time to form a set of maintenance candidate windows.

6. The mold life prediction and maintenance decision-making method based on industrial big data according to claim 5, characterized in that, An improved particle swarm optimization algorithm is used to solve for the optimal maintenance intervention timing, including: Encode the position vector of each particle into a set of maintenance intervention time points, with the dimension equal to the number of maintenance candidate windows; The fitness function is defined as a weighted sum, where the first term is the production delay cost caused by maintenance, the second term is the integral expectation of the probability of sudden mold failure, and the weighting coefficients are dynamically adjusted according to the urgency level of the order. An inertial weight adjustment factor based on the mold degradation rate is introduced into the particle velocity update formula. When the degradation rate exceeds a preset threshold, the inertial weight is increased to enhance the global search capability. The maximum number of iterations is set to 200 generations, the population size is 50 particles, and the process is terminated early when the fitness value changes by less than 1‰ for 30 consecutive generations. The particle with the best fitness is selected from the final population and decoded into the actual maintenance intervention time series.

7. The mold life prediction and maintenance decision-making method based on industrial big data according to claim 6, characterized in that, The maintenance types include three levels: preventative replacement, partial repair, and complete overhaul, and the selection criteria are as follows: When the predicted remaining service life is greater than 60% of the standard service life and no cracks are detected in the surface wear image, local repair is performed; When the remaining service life prediction is between 30% and 60% of the standard service life, or when a micron-sized crack is detected but does not penetrate the working surface, preventive replacement shall be performed. A full overhaul shall be performed when the predicted remaining service life is less than 30% of the standard service life, or when the stress and strain data show abnormal changes, or when the lubrication condition index exceeds the safety limit three times in a row.

8. The mold life prediction and maintenance decision-making method based on industrial big data according to claim 7, characterized in that, Pushing the maintenance decision command to the manufacturing execution system includes: Based on maintenance decision instructions, an electronic work order is automatically generated, which includes the maintenance type, required spare parts, technical personnel qualification requirements, and estimated working hours. Write the maintenance-occupied period into the master production schedule database of the manufacturing execution system, triggering automatic rearrangement of adjacent orders; After maintenance is completed, the actual replacement mold component status data is collected and compared with the prediction results for deviation analysis. The deviation data is then used to fine-tune the output layer weights of the life prediction model online.

9. The mold life prediction and maintenance decision-making method based on industrial big data according to claim 1, characterized in that, The input sequence of the deep residual convolutional neural network is a snapshot of the state of the most recent 200 strokes. After the first convolution, the number of channels is expanded to 64. The number of channels doubles after every two residual blocks. The final output feature map has a dimension of 25×512.

10. The mold life prediction and maintenance decision-making method based on industrial big data according to claim 4, characterized in that, The gated loop unit network includes a reset gate and an update gate, and its hidden state update formula is:

11. Among them, For the first Damage feature vectors at each time step, For the corresponding hidden state, For the sigmoid function, This represents element-wise product.

Images