Mold forming process parameter diagnosis method based on genetic algorithm

Abstract

The application discloses a mold forming process parameter diagnosis method based on a genetic algorithm, and particularly relates to the field of data analysis, and comprises multi-source heterogeneous data fusion and enhanced processing, multi-objective dynamic fitness function construction, constraint adaptive genetic algorithm optimization, process parameter-defect correlation mapping, online transfer learning optimization, dynamic environment online correction, and man-machine collaborative decision verification. The application realizes the collaborative optimization of size, energy consumption and cycle time through a multi-objective dynamic fitness function, significantly improves the practicality and flexibility of the parameter scheme, constructs a closed-loop learning system by using transfer learning and a historical case library, enhances the generalization ability and result reliability, compensates for temperature and humidity fluctuations in real time through the environment online correction function, greatly improves the robustness and stability of the production process, and finally forms a comprehensive solution that has both intelligent optimization and manual intervention capabilities.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and more specifically, to a method for diagnosing mold forming process parameters based on genetic algorithms. Background Technology

[0002] Existing technical solutions typically employ relatively isolated optimization strategies. Their operational process generally involves collecting limited process data through sensors, calculating parameters based on a single optimization objective such as product dimensional accuracy or production cycle time using traditional optimization algorithms, and then directly sending the resulting parameter scheme to the injection molding machine for execution. The entire process lacks effective integration of multi-source heterogeneous data and fails to fully consider the complex correlation between multiple optimization objectives and the dynamic needs of actual production orders.

[0003] However, in practical use, it still has some shortcomings, such as the inability to effectively coordinate and optimize multiple conflicting target indicators, making it difficult to ensure high product quality while simultaneously achieving low energy consumption and high efficiency; the lack of an effective online sensing and compensation mechanism for dynamic interference factors such as temperature and humidity fluctuations in the production workshop environment, which makes the optimized static parameters prone to inaccuracy in actual production; at the same time, the optimization process is highly dependent on real-time data, failing to fully explore and utilize the valuable knowledge contained in historical cases, and the optimization decision-making process is like a black box, lacking an explicit explanation of the correlation between parameters and defects, making it difficult for human experience to effectively intervene for verification and fine-tuning, ultimately restricting the reliability, stability and scalability of the process optimization effect. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, the present invention provides a method for diagnosing mold forming process parameters based on genetic algorithms, which solves the problems mentioned in the background art through the following scheme.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for diagnosing mold forming process parameters based on a genetic algorithm, comprising:

[0006] S1: Multi-source heterogeneous data fusion and enhancement processing: Fusion and enhancement of mold sensor data, material property data, and historical defect data, including data alignment and sample enhancement, and then transmission to S2 and S3;

[0007] S2: Construction of multi-objective dynamic fitness function: Extract dimensional accuracy index, energy consumption index and cycle time index, configure dynamic weight coefficients according to the priority of the current production order, and construct a multi-objective fitness function that integrates dimensional error term, energy consumption term and cycle time term for optimization calculation of S3;

[0008] S3: Constrained Adaptive Genetic Algorithm Optimization: The melt temperature parameters, holding pressure parameters, and cooling time parameters are encoded into binary to generate chromosomes. During the iteration process, the repair operator is applied to handle process constraint conflicts. The crossover probability and mutation probability are dynamically adjusted according to the number of generations. The Pareto optimal parameter solution set is obtained and transmitted to S4 for analysis and then transmitted to S5 to build a case library.

[0009] S4: Process parameter-defect association mapping: The decision tree algorithm is used to extract the explicit rules of parameter combination and flash defects and short shot defects. The contribution of melt temperature parameters and holding pressure parameters to the defect rate is calculated through parameter perturbation sensitivity analysis to obtain parameter-defect mapping rules.

[0010] S5: Online transfer learning optimization: Calculate mold similarity based on material type features, cavity complexity features, and gate type features. Match the optimization solutions of similar molds from the historical case library and mix them with the randomly generated population to construct an initial population. The initial transfer population is then fed back to S3 as the initial input.

[0011] S6: Dynamic Environment Online Correction: Real-time acquisition of workshop temperature and humidity sensor data sequences, prediction of melt temperature compensation and holding pressure compensation through LSTM compensation model, and triggering real-time parameter correction when the ambient temperature changes exceed the threshold, outputting environmental compensation parameters to compensate and correct the output of S3.

[0012] S7: Human-Machine Collaborative Decision Verification: Verify and adjust the output of S3, calculate the nearest distance of parameters to the constraint boundary and generate a confidence assessment. When the confidence is lower than the threshold, push sensitive parameters to the human-machine interaction interface, receive fine-tuning feedback from the management terminal and update the historical defect database. The adjusted data is then fed back to the historical database of S1.

[0013] Preferably, the data alignment first extracts the timestamp sequence of the PLC signal and the timestamp sequence of the industrial camera image frame, and sorts the two sequences according to the time axis; for each frame of the industrial camera image timestamp, finds two adjacent sampling points in the PLC signal sequence; if the image timestamp is between two sampling points of the PLC signal, then the corresponding process parameter value at that moment is calculated using linear interpolation; if the image timestamp and the PLC signal sampling point completely overlap, then the value of that sampling point is directly used.

[0014] Preferably, the sample augmentation first uses a random forest algorithm to analyze the historical defect dataset, and selects key process parameters that significantly affect rare defects through feature importance assessment, including the instantaneous change in melt temperature, the holding pressure gradient value, and the cooling rate fluctuation range. Then, based on the selected key parameter dimensions, the k nearest neighbor samples of the target rare sample are located. Next, a nearest neighbor sample is randomly selected, and the numerical differences between the nearest neighbor sample and the target sample in each key parameter are calculated. Finally, a synthetic sample is generated based on this difference. Specifically, the key parameter values ​​of the target sample are multiplied by a random weight factor and the result of the parameter difference, and then added to form new sample data. This synthesis process is repeated until the ratio of the number of rare defect samples to the number of common defect samples reaches a preset balance threshold.

[0015] Preferably, firstly, for the dimensional accuracy target, the absolute deviation between the actual product size and the design target size is calculated. This deviation is divided by the design target size and multiplied by 100% to convert it into a percentage dimensional error index. Secondly, for the energy consumption target, the total energy consumption data of the injection molding machine during actual operation is collected and divided by the rated power value marked on the equipment nameplate to obtain a dimensionless energy efficiency ratio. Finally, for the production efficiency target, the measured durations of the mold opening and closing time, melt injection time, and product cooling time are accumulated to obtain complete production cycle time data. After the raw data collection is completed, the system automatically performs normalization transformation: the dimensional error index is directly expressed as a percentage value, the energy consumption index is converted into an efficiency coefficient between zero and one, and the cycle time index is divided by the preset maximum allowable cycle time threshold to convert it into a completion coefficient. The conversion result is used as a standardized input item for the fitness function.

[0016] Preferably, the multi-objective fitness function first automatically selects a preset weight template based on order attributes. When the system identifies an order marked as having high precision requirements, it assigns a weight coefficient of 0.7 to the dimensional accuracy target, a weight coefficient of 0.2 to the energy efficiency target, and a weight coefficient of 0.1 to the cycle time target. If the order is marked as having regular requirements, it assigns a weight coefficient of 0.3 to the dimensional accuracy target, a weight coefficient of 0.4 to the energy efficiency target, and a weight coefficient of 0.3 to the cycle time target. The weight allocation module is integrated into the production order management system and automatically triggers weight configuration when a new order is imported. The fitness function is constructed using a weighted summation method, which multiplies the normalized dimensional accuracy coefficient by the corresponding weight coefficient, adds the normalized energy efficiency coefficient by the corresponding weight coefficient, and adds the normalized cycle time completion coefficient by the corresponding weight coefficient to obtain the final comprehensive fitness evaluation value.

[0017] Preferably, the chromosome uses binary strings to represent the combination of mold forming process parameters, and each parameter is assigned a fixed number of binary code segments according to its physical range and accuracy requirements.

[0018] Preferably, the constraint conflict is handled by real-time correction of invalid solutions generated by the genetic algorithm through a repair operator, and the correction logic is automatically triggered when a new individual violates the process constraint rules.

[0019] Preferably, the rule extraction of the decision tree uses the Pareto optimal solution set output by the genetic algorithm as the input data source. Each solution corresponds to a set of process parameter combinations and their defect type labels. The C4.5 decision tree algorithm is used to train the classification model. The information gain ratio is set as the node splitting standard, and the minimum number of leaf node samples is set to 5 to prevent overfitting. After training, the path from the root node to the leaf node is automatically converted into readable rules in IF-THEN form.

[0020] Preferably, the sensitivity analysis performs a quantitative impact assessment on key process parameters in the decision tree rules. After selecting the parameter to be analyzed, a perturbation of ±5% is applied to its original optimal value. Other parameters keep the Pareto solution set values ​​unchanged. Mold forming simulation is performed and the defect rate change value is recorded. Then, the change is divided by the sum of the absolute values ​​of the defect rate changes of all parameters to obtain the percentage contribution of this parameter to the specific defect.

[0021] Preferably, in step S5, three core features—material type, cavity complexity, and gate type—are selected as matching dimensions. Material type is determined using a binary judgment: it is assigned a value of 1 if they are completely identical and 0 otherwise. Cavity complexity is measured by the absolute difference in the number of surfaces and normalized to the range of 0 to 1. Gate type is assigned a value of 1 if they are of the same type and 0 otherwise. Each of the three features is assigned a preset weight coefficient, and the sum of the weights is 1. The final similarity score is the weighted sum of the features.

[0022] Preferably, the initial population construction adopts a hybrid strategy of similar case inheritance and random generation. From the historical mold case library with the highest similarity selected in the case feature matching stage, the process parameter solution set optimized by the genetic algorithm is extracted as the basic data source. 30% of the high-quality individual solutions are extracted according to a preset ratio. At the same time, 70% of the new parameter individual solutions are randomly generated for the current new mold. The new solutions are randomly sampled and generated evenly within the allowable range of process parameters. Finally, the two sets of solutions are merged to form a complete initial population.

[0023] Preferably, the input data of the LSTM compensation model is continuous time-series data collected by temperature and humidity sensors arranged around the mold. The data sampling frequency is once per minute. The temperature and humidity data of the most recent 10 minutes are arranged in chronological order into a sequence of length 10. The data is first normalized so that all values ​​are in the range of 0 to 1, and then input into a two-layer LSTM network. The first LSTM layer contains 32 neurons for extracting time features, and the second LSTM layer contains 16 neurons for associating feature mapping. Finally, two compensation values ​​are output through a fully connected layer. The first output value is the melt temperature compensation value, which represents the temperature value that needs to be adjusted due to the change in melt heat dissipation efficiency caused by the rise in ambient temperature. The second output value is the holding pressure compensation value, which represents the pressure value that needs to be adjusted due to the difference in material fluidity caused by the change in ambient humidity. The training data comes from paired samples of environmental fluctuations and successfully corrected process parameter offsets recorded in historical production processes.

[0024] Preferably, the system continuously receives monitoring data streams from ambient temperature and humidity sensors. When the absolute value of the difference between the current ambient temperature and the baseline ambient temperature during genetic algorithm optimization exceeds 3°C, a compensation process is immediately triggered. The system calls the environmental compensation model, inputs the temperature and humidity time-series data for the most recent 10 minutes, and the model outputs the corresponding melt temperature compensation amount and holding pressure compensation amount. Subsequently, the system automatically performs correction operations: the melt temperature setpoint obtained from genetic algorithm optimization is added to the melt temperature compensation amount, and a new melt temperature control command is sent to the injection molding machine temperature control unit; at the same time, the optimized holding pressure setpoint is added to the holding pressure compensation amount, and a new holding pressure command is sent to the hydraulic system. The corrected parameters remain in effect until the ambient temperature fluctuation returns to the 3°C threshold range, or the system starts a new round of genetic algorithm optimization.

[0025] Preferably, step S7 automatically calculates a confidence index based on the process parameter combination optimized in step S3. This index reflects the reliability of the parameter combination in actual production. In specific implementation, firstly, the safety margin value is calculated for each process parameter. The safety margin value is calculated by taking the minimum difference between the parameter value and the upper or lower limit of its allowable range, and then dividing it by the total width of the allowable range of the parameter. After completing the safety margin calculation for all parameters, the minimum value among all safety margin values ​​is taken as the overall confidence of the parameter combination. Finally, the system converts the confidence value into a percentage and displays it visually as a progress bar at the top of the solution card on the human-machine interface.

[0026] Preferably, when the confidence level calculated by the system is lower than the 80% threshold, the interactive adjustment process is automatically triggered. The system first identifies the 1-3 key process parameters with the lowest safety margin. At this time, in the process parameter adjustment interface, the input boxes corresponding to these parameters are highlighted with a red border, and a yellow exclamation mark icon is displayed next to the parameter name. The management terminal adjusts the parameters by dragging the slider corresponding to the highlighted parameters. The slider displays the parameter value and the corresponding safety margin change curve in real time. After the management terminal confirms the modification, it clicks the submit button. The system automatically records the adjusted parameter value, the employee number of the person making the modification, and the timestamp. At the same time, it generates a drop-down menu of preset options for the adjustment reason for engineers to select. The new parameter combination after submission is immediately transmitted to the MES system in the workshop for trial production. At the same time, the parameter combination and its adjustment record are written into the parameter adjustment history database as new sample data for subsequent genetic algorithm iterative training.

[0027] The technical effects and advantages of this invention are as follows:

[0028] This invention constructs a multi-objective dynamic fitness function that integrates dimensional accuracy, energy consumption, and cycle time indicators, and innovatively automatically configures the dynamic weight coefficients of each sub-objective based on the priority of real-time production orders. This effectively solves the core problem of multi-objective collaborative optimization in existing technologies, making the process parameter combination obtained by the optimization algorithm no longer a single extreme solution, but a Pareto optimal solution set that can accurately adapt to the needs of different production scenarios. In high-precision orders, it prioritizes product dimensional quality, while in regular orders, it intelligently balances production efficiency and energy consumption, thereby significantly improving the practicality and flexibility of the process parameter scheme under complex and ever-changing actual production conditions.

[0029] This invention provides a high-quality data foundation for optimization by implementing multi-source heterogeneous data fusion and enhancement processing. It fully utilizes historical case knowledge to accelerate search and improve generalization ability through online transfer learning optimization. Furthermore, it introduces human experience into the closed loop through human-machine collaborative decision verification. This not only constitutes an intelligent system that continuously learns and improves, effectively reducing over-reliance on a single current data source, but also significantly enhances the interpretability and reliability of optimization results by extracting parameter-defect explicit rules and calculating the confidence of parameter combination through decision tree. This enables managers to understand the optimization logic and participate in key decisions, ultimately improving the practical value and trust level of the entire optimization system.

[0030] This invention introduces a dynamic environment online correction function, utilizes a long short-term memory network compensation model to analyze the time-series data of workshop temperature and humidity in real time, and accurately predicts the required compensation amount of process parameters. It then performs feedforward real-time correction on the static optimization value output by the genetic algorithm, thereby effectively overcoming the adverse effects of workshop environmental fluctuations on the process stability of the molding process. This significantly improves the robustness and adaptability of the production system against dynamic interference, ensuring that the optimization theoretical parameters are executed accurately and stably in the actual production environment, and realizing a technological leap from static optimization to dynamic real-time optimization. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the overall structure of the present invention.

[0032] Figure 2 This is a schematic diagram of the multi-objective dynamic fitness function construction structure of the present invention.

[0033] Figure 3 This is a schematic diagram of the constraint-adaptive genetic algorithm optimization structure of the present invention.

[0034] Figure 4 This is a schematic diagram of the online transfer learning structure of the present invention. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] refer to Figure 1-4 The method for diagnosing mold forming process parameters based on genetic algorithms, as shown, includes:

[0037] S1: Multi-source heterogeneous data fusion and enhancement processing: Fusion and enhancement of mold sensor data, material property data, and historical defect data, including data alignment and sample enhancement, and then transmission to S2 and S3.

[0038] When the mold sensor data, material property data, and historical defect data are input into the system, they are all structured and stored as process parameter vectors.

[0039] The data alignment operation is performed on the injection molding machine PLC signal and the molding defect frames captured by the high-speed industrial camera. The injection molding machine PLC signal collects process parameter data at a frequency of 10 times per second, including key variables such as melt temperature, injection pressure, and holding time; the high-speed industrial camera captures surface defect images of the product at a frequency of 1000 frames per second, and marks the defect type and occurrence timestamp using image recognition technology. Due to the significant difference in sampling frequencies between the two types of equipment, data fusion needs to be achieved through timestamp alignment. The specific process is as follows: First, extract the timestamp sequence of the PLC signal and the timestamp sequence of the industrial camera image frames, and sort the two sequences by time axis; for each frame of the industrial camera image timestamp, find two adjacent sampling points in the PLC signal sequence; if the image timestamp is between two sampling points of the PLC signal, use linear interpolation to calculate the corresponding process parameter value at that moment; if the image timestamp and the PLC signal sampling point completely overlap, directly use the value of that sampling point; this method ensures that each defect frame is associated with a precisely matched set of process parameters, forming a synchronized dataset.

[0040] Let the target time point be This refers to the specific time when data needs to be supplemented, i.e., the precise timestamp when the high-speed industrial camera captured the defect image, and the process parameters at that time point. The value is Let represent the target parameter value to be solved (such as melt temperature, injection pressure, etc.), then the interpolation formula is... , Target time point The process parameter values ​​are obtained through interpolation calculation. The previous valid sampling time point indicates the PLC sensor's position at the target time point. The most recent actual sampling time, The next valid sampling time point indicates the PLC sensor's position at the target time point. The most recent actual sampling time after that, The measured value at the previous sampling point indicates that the PLC is in... The actual parameter values ​​collected at any given time. The measured value at the next sampling point indicates that the PLC is... The interpolation formula follows the assumption of linear continuity and the principle of time-proportional allocation. The specific process is as follows: First, determine the relative position of the target time within the sampling time interval of adjacent PLCs. Obtain the standardization coefficient by calculating the proportion of the time difference between the target time and the previous sampling point to the sampling interval. Then, according to the actual parameter difference between two adjacent sampling points, the parameter change is weighted and allocated according to this time-proportional coefficient. Finally, the change allocation is superimposed on the measured value of the previous sampling point to synthesize the process parameter estimation for the target time. At the engineering level, it is assumed that the temperature, pressure and other parameters of the injection molding process show an approximately linear gradual change trend within a millisecond time interval, which is consistent with the physical characteristics of melt flow. At the data level, low-frequency sensor sampling and high-frequency visual inspection are synchronized on a unified time axis through mathematical derivation to solve the timing matching problem of multi-source equipment. At the application level, it is ensured that the estimated value has process rationality and provides a reliable time-aligned parameter sequence for defect analysis.

[0041] To address the insufficient data volume of rare defect samples during the mold forming process, the sample augmentation employs a synthetic minority class oversampling technique based on random forests. First, the random forest algorithm is used to analyze historical defect datasets, and key process parameters significantly affecting rare defects are selected through feature importance assessment. These parameters include instantaneous changes in melt temperature, holding pressure gradient, and cooling rate fluctuation range. Then, the k nearest neighbors of the target rare sample are located along the selected key parameter dimensions. Next, a nearest neighbor is randomly selected, and the numerical differences between this nearest neighbor and the target sample in each key parameter are calculated. Finally, a synthetic sample is generated based on these differences. Specifically, the key parameter values ​​of the target sample are multiplied by a random weighting factor and the result of the parameter differences, forming new sample data. This synthesis process is repeated until the ratio of rare defect samples to common defect samples reaches a preset equilibrium threshold.

[0042] Let a rare defect sample point in the feature space be... This represents the set of defective process parameters that have occurred in production, and is related to the feature space. One of the k most similar neighbor sample points is , The random weighting factor is determined through random selection. For real numbers whose values ​​range from 0 to 1, the composition rule is as follows: , The virtual defect sample points newly generated through sample augmentation technology are essentially synthetic data points calculated by linear interpolation along the feature space connecting the original rare defect sample and its nearest neighbor sample. This synthesis rule is based on the local linear distribution characteristics of rare defect samples in the feature space. It selects real rare defect sample points and their randomly selected nearest neighbor sample points for linear interpolation along the key process parameter dimensions selected by random forest filtering. The generation of new sample points is strictly limited to the interval between the two points, ensuring that the parameter combination conforms to the constraints of the actual process feasible region. The random weight coefficients are designed with a uniform distribution to simulate the natural fluctuation characteristics of process parameters in injection molding production, ensuring that the new samples are distributed with equal probability density in the neighborhood of the original sample. The physical significance of this method is twofold: First, by filling a reasonable fluctuation range in the process parameter space, virtual defect samples with the same statistical distribution characteristics as real production data are constructed. For example, the data distribution boundary is extended along the actual parameter fluctuation direction on a two-dimensional plane of melt temperature and holding pressure. Second, the core feature patterns of the original defect samples are inherited, so that the synthesized samples retain both the basic defect causes, such as the material degradation characteristics caused by high temperature, and the auxiliary process features of neighboring samples, such as the flow behavior under a specific cooling gradient. This allows the data to restore the performance variations of the same defect type under different equipment states or environmental conditions in actual production, ultimately providing physically reasonable augmented data for subsequent genetic algorithms to avoid the risk of overfitting.

[0043] In this embodiment, it should be specifically noted that, in the mold forming process, rare defects refer to defect types that occur less than 5% of the time in actual production but seriously affect product quality. These mainly include two types: scorching and weld lines. Scorching defects are specifically manifested as brown or black scorch marks on the surface of the plastic part. The cause is that the melt stays in the high-temperature area for too long or local shear overheating leads to material decomposition. This defect is highly sudden in high-speed thin-wall injection molding and is difficult to eliminate through conventional process adjustments. Weld lines are specifically manifested as visible joint lines on the surface of the plastic part. The cause is that the fusion of the front edges of two melt flows is insufficient. This often occurs in mold structures with multiple gates or complex inserts. Its occurrence is random and sensitive to material flowability.

[0044] S2: Construction of multi-objective dynamic fitness function: Extract dimensional accuracy index, energy consumption index and cycle time index, configure dynamic weight coefficients according to the priority of the current production order, and construct a multi-objective fitness function that integrates dimensional error term, energy consumption term and cycle time term for optimization calculation of S3.

[0045] First, for the dimensional accuracy target, the absolute deviation between the actual product size and the design target size is calculated. This deviation is divided by the design target size and multiplied by 100% to convert it into a percentage dimensional error index. Second, for the energy consumption target, the total energy consumption data of the injection molding machine during actual operation is collected and divided by the rated power value marked on the equipment nameplate to obtain a dimensionless energy efficiency ratio. Finally, for the production efficiency target, the measured durations of the mold opening and closing time, melt injection time, and product cooling time are accumulated to obtain complete production cycle time data. After the raw data collection is completed, the system automatically performs normalization transformation: the dimensional error index is directly expressed as a percentage value, the energy consumption index is converted into an efficiency coefficient between zero and one, and the cycle time index is divided by the preset maximum allowable cycle time threshold to convert it into a completion coefficient. The conversion result is used as a standardized input item for the fitness function.

[0046] The multi-objective fitness function first automatically selects a preset weight template based on order attributes. When the system identifies an order marked as having high precision requirements, it assigns a weight coefficient of 0.7 to the dimensional accuracy objective, 0.2 to the energy efficiency objective, and 0.1 to the cycle time objective. If the order is marked as having regular requirements, it assigns a weight coefficient of 0.3 to the dimensional accuracy objective, 0.4 to the energy efficiency objective, and 0.3 to the cycle time objective. The weight allocation module is integrated into the production order management system and automatically triggers weight configuration when a new order is imported. The fitness function is constructed using a weighted summation method, multiplying the normalized dimensional accuracy coefficient by its corresponding weight coefficient, adding the normalized energy efficiency coefficient multiplied by its corresponding weight coefficient, and adding the normalized cycle time completion coefficient multiplied by its corresponding weight coefficient to obtain the final comprehensive fitness evaluation value. This calculation process is executed in real time within the genetic algorithm's iterative loop. Each time an individual solution is evaluated, the weight configuration module is called to obtain the weight value corresponding to the current order type, ensuring that the optimization direction strictly matches the actual production needs.

[0047] The high-precision requirement specifically refers to order types with stringent dimensional tolerances for molded products, typically used in precision parts manufacturing. The core criterion is that the dimensional tolerance range is explicitly stated in the order's technical specifications as ≤±0.5%. The conventional requirement, on the other hand, refers to order types with relatively relaxed dimensional tolerances, typically used in general industrial products manufacturing. The criterion is that the dimensional tolerance range is >±0.5% but ≤±2%. The distinction between the two types of requirements is automatically identified by the order preprocessing system. When the dimensional tolerance field value is ≤0.5, it is marked as a high-precision requirement; otherwise, it is marked as a conventional requirement.

[0048] Let the dimensional error percentage be The energy efficiency ratio is The actual cycle time is Then the fitness function , Overall fitness evaluation value Maximum allowable cycle time , , The dynamic weighting coefficients are used to first transform the original production indicators into a unified, positive evaluation scale based on efficiency. This transformation ensures that the larger the values ​​of the three indicators, the better the performance, and that the dimensions are consistent. Then, a dynamic weighting mechanism is introduced, and the system automatically assigns weighting coefficients according to the order type: high-precision orders are assigned a weight of 0.7 for the dimensional accuracy target to strengthen quality control, and 0.2 and 0.1 for the energy consumption target and the cycle time target, respectively; regular orders are assigned a weight of 0.3 for the dimensional accuracy target, while the weights of the energy consumption and cycle time targets are increased to 0.4 and 0.3, respectively, to achieve balanced optimization. Finally, the three transformed indicators are combined into a single score value through weighted summation. Its physical meaning is to quantitatively characterize the overall effectiveness of a specific combination of process parameters within the order demand framework. A score close to 100% indicates that the current parameters simultaneously meet the production requirements of high precision, low energy consumption, and short cycle time. A lower score reflects defects such as dimensional deviations, energy waste, or insufficient efficiency. This score directly serves as the decision-making basis for the genetic algorithm to select superior process solutions.

[0049] S3: Constrained Adaptive Genetic Algorithm Optimization: The melt temperature parameter, holding pressure parameter, and cooling time parameter are encoded into binary to generate chromosomes. During the iteration process, the repair operator is applied to handle process constraint conflicts. The crossover probability and mutation probability are dynamically adjusted according to the number of generations. The Pareto optimal parameter solution set is obtained and transmitted to S4 for analysis and then transmitted to S5 to build the case library.

[0050] The chromosome uses binary strings to represent combinations of mold forming process parameters. Each parameter is assigned a fixed number of binary code segments based on its physical range and accuracy requirements. The melt temperature parameter has an allowable range of 180℃ to 250℃, with an 8-bit binary code and a minimum resolution of 0.27℃ corresponding to one least significant bit. The injection time parameter has an allowable range of 1.0 second to 5.0 seconds, with a 6-bit binary code and a resolution of 0.06 seconds. The holding pressure parameter has an allowable range of 30MPa to 90MPa, with a 7-bit binary code and a resolution of 0.47MPa. All parameter code segments are sequentially connected to form a complete chromosome, and the length of a single chromosome is determined by the number of parameters.

[0051] The handling of constraint conflicts involves real-time correction of invalid solutions generated by the genetic algorithm using a repair operator. Correction logic is automatically triggered when a new individual violates process constraint rules. If the holding pressure exceeds 80% of the clamping force safety threshold, the parameter is forcibly reset to 80% of the clamping force. If the melt temperature exceeds the upper limit of the material's allowable range, it is adjusted to the maximum tolerance temperature recommended by the material supplier. If the cooling time is lower than the minimum cooling requirement of the mold, the theoretical minimum value is recalculated based on the runner diameter and wall thickness data and replaced. All correction operations are completed before the fitness assessment to ensure that all individuals entering the selection phase meet the basic feasibility requirements.

[0052] The theoretical minimum value is specifically calculated using a one-dimensional unsteady-state heat conduction equation. Based on Fourier's heat conduction law, it calculates the shortest time required for the melt to drop from the initial temperature to the ejection temperature within the mold cavity. This embodiment will not elaborate on this in detail.

[0053] The crossover and mutation probabilities are dynamically adjusted according to the iteration process of the genetic algorithm. The initial crossover probability is set to 0.8 and the mutation probability is set to 0.1. As the number of generations increases, the crossover probability decreases linearly, and the crossover probability decreases by a fixed step size for each additional generation. This step size is determined by the maximum number of iterations. The mutation probability increases linearly, and the mutation probability increases by a fixed step size for each additional generation. When the evolution reaches the maximum number of generations, the crossover probability drops to 0.2 and the mutation probability rises to 0.4. The specific adjustment amount is calculated by the ratio of the total number of iterations to the current generation, without the need for manual intervention in parameter configuration.

[0054] Let the current iteration algebra be Maximum iteration algebra Then the crossover probability Probability of mutation , Initial crossover probability: This indicates that at the start of the algorithm, 80% of individuals will participate in gene exchange. Crossover decay coefficient: This represents the rate at which the crossover probability decreases from its initial value to its final value. Initial mutation probability: This indicates that only 10% of gene loci undergo random mutations in the early stages. The mutation growth coefficient represents the rate at which the mutation probability increases from its initial value to its final value; the crossover probability... Based on the principle of dynamic equilibrium in evolutionary stages: In the initial stage, a high crossover probability (0.8) is set to fully utilize population diversity. The solution space is rapidly explored through large-scale exchange of chromosome gene segments (such as optimization of the combination of melt temperature and holding pressure), accelerating convergence to the potential region; as the iteration progresses ( (Increase), linear attenuation coefficient 0.6 makes The crossover frequency is gradually reduced to 0.2 because high-quality gene combinations (such as stable pairings of cooling time and clamping force) gradually form in the later stages. Reducing the crossover frequency can minimize the damage to existing optimized structures and shift towards local fine-grained search; mutation probability Following a gradual enhancement mechanism of diversity: initially maintaining a low mutation probability (0.1) to avoid random perturbations interfering with the directed search process, while with increasing evolutionary generations, a linear growth coefficient of 0.3 drives the process. When the value is increased to 0.4, the enhanced mutation ability can break through the local optimum trap when the population tends to be homogeneous (for example, introducing mutations in the temperature-sensitive area of ​​the material to explore new process windows), while also dealing with environmental uncertainties (such as parameter mismatch caused by temperature and humidity fluctuations in the workshop).

[0055] S4: Process parameter-defect association mapping: The decision tree algorithm is used to extract the explicit rules of parameter combination and flash defects and short shot defects. The contribution of melt temperature parameters and holding pressure parameters to the defect rate is calculated through parameter perturbation sensitivity analysis to obtain parameter-defect mapping rules.

[0056] The rule extraction of the decision tree uses the Pareto optimal solution set output by the genetic algorithm as the input data source. Each solution corresponds to a set of process parameter combinations and its defect type label. The C4.5 decision tree algorithm is used to train the classification model, and the information gain ratio is set as the node splitting criterion. The minimum number of leaf node samples is set to 5 to prevent overfitting. After training, the path from the root node to the leaf node is automatically converted into a readable rule in IF-THEN form. For example, when the splitting point with a melt temperature feature value greater than 230 degrees Celsius is combined with the splitting point with a holding time feature value less than 1.5 seconds, the leaf node points to the flash defect type. The generated rule is expressed as: IF Melt temperature > 230℃ AND Holding time < 1.5s THEN Defect type = "Flash". Finally, the rule set corresponding to all paths is output for process diagnosis.

[0057] The sensitivity analysis quantifies the impact of key process parameters in the decision tree rules. After selecting the parameter to be analyzed, a perturbation of ±5% is applied to its original optimal value, while other parameters remain unchanged in their Pareto solution set values. Mold forming simulation is performed, and the defect rate changes are recorded. For example, a 5% increase in melt temperature raises the short-shot defect rate by 3%, or a 5% decrease in holding pressure lowers the flash defect rate by 2%. The absolute change in defect rate caused by the parameter perturbation is calculated, and then this change is divided by the sum of the absolute values ​​of the defect rate changes of all parameter perturbations to obtain the percentage contribution of this parameter to a specific defect. A higher contribution indicates a more significant impact of the parameter on the defect.

[0058] Set parameters The absolute change in defect rate caused by applying a ±5% disturbance is All analyzed parameters The change in defect rate is ,but , :parameter Contribution to a specific defect, ranging from [0,1]. A larger value indicates a more significant impact of the parameter on the defect. The sum of the absolute values ​​of the defect rate changes caused by all parameter disturbances, used to normalize the contribution. :common 1 analytical parameter; among which molecules This directly characterizes the intensity of the direct impact of small fluctuations in parameter i on the defect rate. For example, a 5% increase in melt temperature leads to a 4% increase in the short-shot defect rate. The denominator is a positive value of 4; It reflects the total scale of defect rate change when all parameters are disturbed together, providing a benchmark scale for relative influence; final contribution. This reveals the sensitivity priority of parameters to defects. For example, when the contribution of melt temperature is calculated to be 57%, it indicates that the contribution of the fluctuation of this parameter to the formation of defects exceeds the sum of other parameters. In engineering, its stability should be controlled first to reduce the risk of defects. This result can explain the optimization logic of the parameter being strictly constrained to a narrow range in the genetic algorithm.

[0059] S5: Online transfer learning optimization: Calculate mold similarity based on material type features, cavity complexity features, and gate type features. Match the optimization solutions of similar molds from the historical case library and mix them with the randomly generated population to construct an initial population. The initial transfer population is then fed back to S3 as the initial input.

[0060] S5 selects three core features as matching dimensions: material type, cavity complexity, and gate type. Material type is judged by binary judgment, that is, it is assigned a value of 1 when they are completely consistent and 0 otherwise. Cavity complexity is measured by the absolute difference in the number of surfaces and normalized to the range of 0 to 1. Gate type is assigned a value of 1 if they are the same type and 0 otherwise. The three features are assigned preset weight coefficients and the sum of the weights is 1. The final similarity score is the weighted sum of the features.

[0061] The formula for calculating the similarity score is as follows: , Material type weighting coefficient, fixed at 0.5. Material type indicator function Cavity complexity weighting coefficient, fixed at 0.3. The number of surface features in the new mold cavity. The number of surface features in the cavity of historical molds. The maximum number of cavity surfaces preset by the system. : Gate type weighting coefficient, fixed value of 0.2. : Gate type indicator function.

[0062] The initial population construction adopts a hybrid strategy of similar case inheritance and random generation. From the historical mold case library with the highest similarity selected in the case feature matching stage, the process parameter solution set optimized by the genetic algorithm is extracted as the basic data source. 30% of the high-quality individual solutions are extracted according to a preset ratio. At the same time, 70% of the new parameter individual solutions are randomly generated for the current new mold. The new solutions are randomly sampled and generated evenly within the allowable range of process parameters. Finally, the two sets of solutions are merged to form a complete initial population.

[0063] S6: Dynamic Environment Online Correction: Real-time acquisition of workshop temperature and humidity sensor data sequences, prediction of melt temperature compensation and holding pressure compensation through LSTM compensation model, and triggering real-time parameter correction when the ambient temperature changes beyond the threshold, outputting environmental compensation parameters to compensate and correct the output of S3.

[0064] The input data for the LSTM compensation model consists of continuous time-series data collected by temperature and humidity sensors arranged around the mold. The data sampling frequency is once per minute. The temperature and humidity data of the most recent 10 minutes are arranged in chronological order into a sequence of length 10. The data is first normalized so that all values ​​are in the range of 0 to 1, and then input into a two-layer LSTM network. The first LSTM layer contains 32 neurons for extracting time features, and the second LSTM layer contains 16 neurons for associating feature mapping. Finally, two compensation values ​​are output through a fully connected layer. The first output value is the melt temperature compensation value, which represents the temperature value that needs to be adjusted due to the change in melt heat dissipation efficiency caused by the ambient temperature rise. The second output value is the holding pressure compensation value, which represents the pressure value that needs to be adjusted due to the difference in material fluidity caused by the change in ambient humidity. The training data comes from paired samples of environmental fluctuations and successfully corrected process parameter offsets recorded in historical production processes.

[0065] The system continuously receives monitoring data streams from ambient temperature and humidity sensors. When the absolute value of the difference between the current ambient temperature and the baseline ambient temperature during genetic algorithm optimization exceeds 3°C, a compensation process is immediately triggered. The system calls the environmental compensation model, inputs the temperature and humidity time-series data for the most recent 10 minutes, and the model outputs the corresponding melt temperature compensation and holding pressure compensation. Subsequently, the system automatically performs correction operations: the melt temperature setpoint obtained from genetic algorithm optimization is added to the melt temperature compensation, and this is sent as a new melt temperature control command to the injection molding machine temperature control unit; at the same time, the optimized holding pressure setpoint is added to the holding pressure compensation, and this is sent as a new holding pressure command to the hydraulic system. The corrected parameters remain in effect until the ambient temperature fluctuation returns to the 3°C threshold range, or the system starts a new round of genetic algorithm optimization.

[0066] Let the melt temperature setpoint obtained by the genetic algorithm be... The melt temperature compensation amount is determined by the Pareto solution set output by the S3-constrained adaptive genetic algorithm. The corrected actual setpoint for the melt temperature is determined using the LSTM compensation model. This formula is based on the principle of superposition compensation and the assumption of linear additivity. By retaining the theoretical optimal value optimized by the genetic algorithm as the benchmark, it uses the compensation amount predicted by the environmental compensation model as an independent supplementary term, thereby achieving decoupling between the static optimization results and the dynamic environmental response. When the ambient temperature rises and the heat dissipation efficiency of the mold decreases, the reduced melt cooling rate may cause flash or warping. At this time, the compensation amount outputs a negative value to reduce the melt temperature. When the ambient temperature decreases and the heat dissipation efficiency increases, premature solidification of the melt may cause short shots or under-suppression. The compensation amount outputs a positive value to increase the melt temperature and ensure fluidity. The essence of this formula is to offset environmental interference through feedforward control and utilize the linear approximation characteristics of the heat conduction equation to stabilize the actual process state to the ideal state within the range of ±10℃ of ambient temperature change, thereby ensuring process robustness.

[0067] S7: Human-machine collaborative decision verification: Calculate the nearest distance of the parameter to the constraint boundary and generate a confidence assessment. When the confidence is lower than the threshold, push the sensitive parameter to the human-machine interaction interface, receive the fine-tuning feedback from the management terminal and update the historical defect database. The adjusted data is then fed back to the historical database of S1.

[0068] S7 automatically calculates the confidence index based on the process parameter combination optimized in S3. This index reflects the reliability of the parameter combination in actual production. In specific implementation, firstly, the safety margin value is calculated for each process parameter. The safety margin value is calculated by taking the minimum difference between the parameter value and the upper or lower limit of its allowable range, and then dividing it by the total width of the allowable range of the parameter. For example, if the allowable melt temperature range is 180℃ to 250℃, and the current optimized value is 230℃, then its difference from the lower limit is 50℃, its difference from the upper limit is 20℃, the minimum difference is 20℃, and the parameter range width is 70℃. The safety margin is 20 / 70≈0.2857. After completing the safety margin calculation for all parameters, the minimum value among all safety margin values ​​is taken as the overall confidence level of the parameter combination. Finally, the system converts the confidence level value into a percentage and displays it visually as a progress bar at the top of the solution card on the human-machine interface.

[0069] The confidence level is expressed by the following formula: This formula quantifies the reliability of the overall scheme based on the relative safety margin of each parameter value in the process parameter combination from its allowable range boundary. Its physical meaning lies in evaluating the ability of optimized parameters to resist fluctuations or deviations in an actual production environment. Specifically, for each process parameter (such as melt temperature and holding pressure), the minimum absolute distance between the parameter value and its upper or lower allowable range limit is first calculated, then divided by the total allowable range width of the parameter to obtain the single-parameter safety margin. For example, if the allowable melt temperature range is 180℃~250℃, and the optimized value is 230℃, the minimum boundary distance is 20℃, and the range width is 70℃. The full margin is approximately 20 / 70 ≈ 0.2857. Subsequently, the minimum value among all parameter safety margins is taken as the overall confidence level. This is because the reliability of the production system is constrained by the weakest parameter, and the closer any parameter is to its boundary, the greater the risk of process failure. The physical meaning of this confidence level is directly mapped to engineering practice: when the confidence level is 0.8, it indicates that all parameters retain at least 20% safety buffer space (such as the melt temperature deviating from the current value by at least 14°C within the allowable range before exceeding the limit), thereby ensuring stable production under equipment fluctuations or environmental disturbances and avoiding quality defects or equipment failures caused by the critical state of a single parameter.

[0070] When the confidence level calculated by the system falls below the 80% threshold, the interactive adjustment process is automatically triggered. The system first identifies the 1-3 key process parameters with the lowest safety margin. At this point, in the process parameter adjustment interface, the input boxes corresponding to these parameters are highlighted with a red border, and a yellow exclamation mark icon is displayed next to the parameter name. The management terminal adjusts the parameters by dragging the slider corresponding to the highlighted parameters. The slider displays the parameter value and the corresponding safety margin change curve in real time. After the management terminal confirms the modification and clicks the submit button, the system automatically records the adjusted parameter value, the employee number of the person making the modification, and the timestamp. At the same time, it generates a drop-down menu of preset options for the adjustment reason for engineers to select. The new parameter combination after submission is immediately transmitted to the MES system in the workshop for trial production. At the same time, the parameter combination and its adjustment record are written into the parameter adjustment history database as new sample data for subsequent genetic algorithm iteration training to improve the confidence of future optimization schemes.

[0071] This invention first performs multi-source heterogeneous data fusion and enhancement processing, aligning and enhancing real-time acquired mold sensor data, material property data, and historical defect data to form a high-quality standardized dataset. This dataset is then fed into the multi-objective dynamic fitness function construction stage, dynamically configuring weight coefficients for indicators such as dimensional accuracy, energy consumption, and cycle time based on the current production order priority to generate an objective function for genetic algorithm optimization. Next, the constrained adaptive genetic algorithm uses the aforementioned dataset and fitness function to encode and optimize parameters such as melt temperature, holding pressure, and cooling time, obtaining a Pareto optimal process parameter solution set through iterative calculation. This solution set is input into the process parameter defect association mapping module, where a decision tree algorithm extracts explicit rules between parameters and defects such as flash and short shots, and analyzes the sensitivity contribution of key parameters. Simultaneously, it is used by… The data is stored in a historical case library for future transfer learning. Simultaneously, the online transfer learning optimization module calculates similarity based on the material type, cavity complexity, and gate type characteristics of the new mold. It matches optimized solutions from similar cases in the historical case library and mixes them with randomly generated solutions to construct an initial population, which is then fed back to the genetic algorithm. In the dynamic environment online correction stage, the system collects workshop temperature and humidity data in real time. It uses a long short-term memory network compensation model to predict the melt temperature compensation and holding pressure compensation required for environmental changes and corrects the parameter settings output by the genetic algorithm in real time. Finally, the human-machine collaborative decision verification module calculates the confidence level of the parameter combination. When the confidence level is insufficient, it pushes sensitive parameters to the interactive interface to request manual fine-tuning and updates the adjusted parameters and their results back to the historical defect database, thus forming a closed-loop intelligent decision-making process from data preparation, parameter optimization, environmental correction to manual verification.

[0072] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0073] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A mold forming process parameter diagnosis method based on a genetic algorithm, characterized by, Comprise: S1: multi-source heterogeneous data fusion and enhancement processing: fuse and enhance mold sensor data, material attribute data, historical defect data, including data alignment and sample enhancement, and then transmit to S2 and S3; S2: multi-objective dynamic fitness function construction: extract size precision index, energy consumption index and cycle time index, configure dynamic weight coefficient according to the priority of the current production order, and construct a multi-objective fitness function that integrates size error term, energy consumption term and cycle time term for optimization calculation in S3; S3: constraint adaptive genetic algorithm optimization: binary coding of melt temperature parameters, holding pressure parameters and cooling time parameters to generate chromosomes, repair operator is applied to handle process constraint conflicts in the iteration process, and crossover probability and mutation probability are dynamically adjusted according to the evolution generation to obtain a set of Pareto optimal parameter solutions which are transmitted to S4 for analysis and to S5 to construct a case library; S4: process parameter-defect correlation mapping: use decision tree algorithm to extract the explicit rules of parameter combination and flash defect, short shot defect, calculate the contribution of melt temperature parameters and holding pressure parameters to defect rate through parameter perturbation sensitivity analysis, and obtain parameter-defect mapping rules; S5: online transfer learning optimization: calculate mold similarity based on material type features, cavity complexity features and gate type features, match similar mold optimization solutions from the historical case library, and mix with randomly generated populations to construct an initial population, and obtain a transfer initial population as the initial input of S3; S6: dynamic environment online correction: real-time acquisition of workshop temperature and humidity sensor data sequence, prediction of melt temperature compensation and holding pressure compensation through LSTM compensation model, triggering of real-time parameter correction when the environmental temperature changes by more than a threshold, and output of environmental compensation parameters to compensate and correct the output of S3; S7: human-machine collaborative decision verification: verification and adjustment of the output of S3, calculation of the nearest distance of the parameter distance constraint boundary and generation of confidence evaluation, pushing of sensitive parameters to the human-machine interface when the confidence is lower than a threshold, receiving of fine-tuning feedback from the management personnel terminal and updating of the historical defect database, and returning of the adjusted data to the historical database of S1.

2. The method of claim 1, wherein, The S1 comprises: The data alignment first extracts the timestamp sequence of the PLC signal and the timestamp sequence of the industrial camera image frame, sorts the two sequences according to the time axis; for the timestamp of each frame of industrial camera image, find the adjacent two sampling points before and after in the PLC signal sequence; if the image timestamp is between the two sampling points of the PLC signal, use linear interpolation method to calculate the process parameter value corresponding to this time; if the image timestamp coincides with the PLC signal sampling point, directly use the sampling point value; The sample enhancement first analyzes a historical defect data set using a random forest algorithm, filters out key process parameters that have a significant impact on rare defects through feature importance evaluation, including melt temperature instantaneous change amount, holding pressure gradient value, and cooling rate fluctuation range; then locates k nearest neighbor samples of the target rare sample in the dimension of the filtered key parameters; then randomly selects a nearest neighbor sample, calculates the numerical difference between the nearest neighbor sample and the target sample in each key parameter; finally, a synthetic sample is generated based on the difference, specifically by adding the key parameter value of the target sample to the result of multiplying the parameter difference by a random weight factor, to form new sample data; repeat the synthesis process until the ratio of the number of rare defect samples to the number of common defect samples reaches the preset balance threshold.

3. The method of claim 1, wherein, The S2 comprises: First, for the size accuracy target, the absolute deviation value of the actual product size from the design target size is calculated, the deviation value is divided by the design target size and multiplied by 100%, and the size error index in percentage form is obtained; second, for the energy consumption target, the total energy consumption data of the injection molding machine during actual operation is collected, and the dimensionless energy consumption efficiency ratio is obtained by dividing the rated power value marked on the equipment nameplate; finally, for the production efficiency target, the complete production cycle time data is obtained by accumulating the measured time of mold opening and closing time, melt injection time and product cooling time; after completing the original data collection, the system automatically performs normalization conversion: the size error index is directly expressed in percentage value, the energy consumption index is converted into efficiency coefficient between zero and one, and the cycle time index is converted into completion coefficient by dividing the preset maximum allowed cycle time threshold, and the conversion result is used as the standardized input item of the fitness function; The multi-objective fitness function first automatically selects a preset weight template according to the order attribute, and when the system identifies that the order is marked as high precision requirement, the weight coefficient of 0.7 is assigned to the size accuracy target, the weight coefficient of 0.2 is assigned to the energy consumption efficiency target, and the weight coefficient of 0.1 is assigned to the cycle time target; if the order is marked as normal requirement, the weight coefficient of 0.3 is assigned to the size accuracy target, the weight coefficient of 0.4 is assigned to the energy consumption efficiency target, and the weight coefficient of 0.3 is assigned to the cycle time target; the weight distribution module is integrated in the production order management system, and the weight configuration is automatically triggered when a new order is imported; the specific construction of the fitness function adopts the weighted summation method, that is, the normalized size accuracy coefficient is multiplied by the corresponding weight coefficient, the normalized energy consumption efficiency coefficient is added to the corresponding weight coefficient, and the normalized cycle time completion coefficient is added to the corresponding weight coefficient, and finally the comprehensive fitness evaluation value is obtained.

4. The method of claim 1, wherein, The S3 comprises: The chromosome uses a binary string to represent the mold forming process parameter combination, and each parameter is assigned a fixed number of binary code segments according to its physical range and accuracy requirements; The constraint conflict is processed by a repair operator to correct invalid solutions generated by the genetic algorithm in real time, and the correction logic is automatically triggered when a new individual violates the process constraint rule.

5. The method of claim 1, wherein, The S4 comprises: The rule extraction of the decision tree takes the Pareto optimal solution set output by the genetic algorithm as the input data source, each solution corresponds to a set of process parameter combinations and its defect type label, adopts the C4.5 decision tree algorithm to train a classification model, sets the information gain rate as the node splitting standard, and sets the minimum leaf node sample number to 5 to prevent overfitting, and after the training is completed, the path from the root node to the leaf node is automatically converted into readable rules in the IF-THEN form; The sensitivity analysis implements quantitative influence evaluation on the key process parameters in the decision tree rule, after the parameter to be analyzed is selected, a disturbance amplitude of ± 5% is applied on the basis of the original optimal value, other parameters remain unchanged, the mold forming simulation is executed, and the defect rate change value is recorded, then the change value is divided by the sum of the absolute values of the defect rate changes of all parameters, and the contribution percentage of this parameter to the specific defect is obtained.

6. The method of claim 1, wherein, The S5 includes: The S5 selects material type, cavity complexity and gate type as the matching dimensions, the material type adopts binary judgment, that is, when completely consistent, the value is 1, otherwise, the value is 0, the cavity complexity is measured by the absolute difference value of the number of curved surfaces and is normalized to the interval of 0 to 1, the gate type is valued as 1 according to the same type matching, otherwise, the value is 0, the three features are respectively given a preset weight coefficient and the sum of the weights is 1, and finally the similarity score is the weighted sum value of each feature; The initial population construction adopts a hybrid strategy of similar case inheritance and random generation, from the highest similarity historical mold case library selected in the case feature matching link, the process parameter solution set optimized by the genetic algorithm is extracted as the basic data source, 30% of the high-quality individual solutions are extracted according to the preset proportion, and 70% of the new parameter individual solutions of the current new mold are randomly generated, the new solutions are uniformly distributed and randomly sampled within the allowable range of process parameters, and finally the two solution sets are combined to form a complete initialization population.

7. The method of claim 1, wherein, The S6 includes: The input data of the LSTM compensation model is the continuous time series data collected by the temperature and humidity sensors arranged around the mold, the data sampling frequency is 1 per minute, the temperature and humidity data of the last 10 minutes are arranged in time sequence as a sequence with a length of 10, normalization processing is first performed to make all values in the interval of 0 to 1, and then input into the two-layer LSTM network; the first layer LSTM contains 32 neurons for extracting time features, the second layer LSTM contains 16 neurons for associating feature mapping, and finally two compensation value values are output through the full connection layer; the first output value is the melt temperature compensation value, which represents the temperature value required to adjust due to the change of melt heat dissipation efficiency caused by environmental temperature rise; the second output value is the holding pressure compensation value, which represents the pressure value required to adjust due to the difference in material flowability caused by environmental humidity change, and the training data comes from the paired samples of environmental fluctuations and successfully corrected process parameter offsets recorded in the historical production process; The monitoring data stream of the ambient temperature and humidity sensor is continuously received, and when the absolute value of the difference between the current ambient temperature and the reference ambient temperature optimized by the genetic algorithm exceeds 3 DEG C, a compensation process is triggered immediately, the system calls the environmental compensation model, inputs the temperature and humidity time series data in the last 10 minutes, and the model outputs the corresponding melt temperature compensation and holding pressure compensation. Subsequently, the system automatically performs correction operation: add the melt temperature compensation to the melt temperature set value obtained by the genetic algorithm optimization, and send it to the injection molding machine temperature control unit as the new melt temperature control instruction; at the same time, add the holding pressure compensation to the optimized holding pressure set value, and send it to the hydraulic system as the new holding pressure instruction. The corrected parameters take effect continuously until the ambient temperature fluctuation returns to the range of 3 DEG C threshold, or the system starts a new round of genetic algorithm optimization process.

8. The method of claim 1, wherein, The S7 comprises: The S7 automatically calculates the confidence index according to the process parameter combination generated by the optimization of S3, which reflects the reliability of the parameter combination in actual production. In specific implementation, first, the safety margin value of each process parameter is calculated. The calculation method of the safety margin value is: taking the minimum difference between the parameter value and the upper limit or lower limit of its allowable range, and then dividing by the total width of the allowable range of the parameter. After completing the safety margin calculation of all parameters, take the minimum value of all safety margin values as the overall confidence of the parameter combination. Finally, the system converts the confidence value into percentage form and visualizes it in the form of a progress bar at the top of the scheme card on the human-machine interface; When the confidence calculated by the system is lower than the 80% threshold, the interactive adjustment process is automatically triggered. The system first locates 1-3 key process parameters with the lowest safety margin. At this time, in the process parameter adjustment interface, the input boxes corresponding to these parameters are highlighted with a red border, and a yellow exclamation mark icon is displayed next to the parameter name. The management personnel terminal adjusts the highlighted parameters by dragging the slider bar. The slider bar displays the parameter value and the corresponding safety margin change curve in real time. After the management personnel terminal confirms the modification and clicks the submit button, the system automatically records the adjusted parameter value, the modification personnel ID and the timestamp, and generates an adjustment reason preset option drop-down menu for the engineer to select. The new parameter combination after submission is immediately transmitted to the MES system of the workshop for trial production, and the parameter combination and its adjustment record are written into the parameter adjustment history database as new sample data for subsequent genetic algorithm iteration training.

Images