Industrial data processing system and method for automotive interior and exterior injection molding production lines

Abstract

This invention relates to the field of intelligent control and industrial data processing technology for injection molding, specifically to an industrial data processing system and method for automotive interior and exterior injection molding production lines. The system includes: acquiring time-series data on mold temperature and cavity pressure, as well as raw material melt index and moisture content fluctuation data collected by a bottom-level sensor network; inputting this data into a reinforcement learning model to generate a first control strategy including injection speed, holding pressure, and cooling time; calculating a weighted system control vulnerability index based on the fluctuation characteristics determined by the variance within a time window and the rate of change of parameters relative to the current operating parameters; comparing this index with a trust threshold, outputting the first control strategy when the index is below the threshold, and triggering a robust control model to generate a second control strategy when the index is above or equal to the threshold, which is then output to the actuator to achieve continuous operation and risk mitigation under high production capacity.

Description

Technical Field

[0001] This invention relates to the field of intelligent control and industrial data processing technology for injection molding, specifically to an industrial data processing system and method for automotive interior and exterior injection molding production lines. Background Technology

[0002] In the daily operation of automotive interior and exterior injection molding production lines, the molding process of parts such as bumpers and dashboards is affected by a variety of factors, including mold temperature, cavity pressure, and raw material melt index and moisture content. The production line usually needs to adjust the injection speed, holding pressure, and cooling time according to real-time operating conditions to ensure the stability of production cycle and part quality. The control effect is directly related to the integrity of the underlying sensor data, the accuracy of operating condition fluctuation identification, and the timeliness of control strategy switching.

[0003] In traditional methods, injection molding production lines often rely on fixed process parameters or a single model to directly output adjustment commands. When there are delays, packet loss, drift interference in the sensor network and increased fluctuations in raw material properties, the control system's ability to comprehensively judge data disturbances and the aggressiveness of parameters is insufficient to meet the control requirements. This can easily lead to parameter overshoot, thermal instability, and continuous quality defects under compressed cycle conditions, thus limiting the production line's ability to operate continuously and stably. Summary of the Invention

[0004] The purpose of this invention is to provide an industrial data processing system and method for automotive interior and exterior injection molding production lines, addressing the following technical problems: Existing control technologies for automotive interior and exterior injection molding production lines have shortcomings in comprehensively judging data disturbances caused by sensor network latency, packet loss, drift interference, and increased fluctuations in raw material properties, as well as in suppressing overshoot and preventing continuous quality defects under compressed cycle time. There is an urgent need for an industrial data processing method and system for automotive interior and exterior injection molding production lines that can more accurately process underlying data feature reconstruction and system control vulnerability assessment, and dynamically switch robust control strategies based on unstable trends. The purpose of this invention can be achieved through the following technical solutions:

[0005] Industrial data processing methods for automotive interior and exterior injection molding production lines include:

[0006] Acquire time-series status data, including mold temperature and cavity pressure, collected by the underlying sensor network, as well as physical property fluctuation data, including melt index and moisture content, of the input raw materials;

[0007] The time-series state data and the physical property fluctuation data are input into a preset reinforcement learning model to generate a first control strategy. The reinforcement learning model is trained with minimizing the production cycle as the reward function. The state space of the reinforcement learning model is represented by a multi-dimensional state vector composed of the time-series state data and the physical property fluctuation data. The action space of the reinforcement learning model is the continuous action adjustment range of the injection speed, holding pressure, and cooling time allowed by the actuator of the injection molding production line. The first control strategy includes injection speed commands, holding pressure commands, and cooling time commands.

[0008] Based on the fluctuation characteristics determined by the variance value of the time-series state data within a preset time window, and the parameter change rate determined by the parameter deviation value of the first control strategy relative to the current operating parameters of the actuator of the injection molding production line, the system control vulnerability index is calculated by weighting through preset weighting coefficients.

[0009] The system control vulnerability index is compared with a preset trust threshold, wherein the preset trust threshold is obtained by joint distribution calibration based on the maximum state variance value and the maximum parameter safety control step size of the injection molding production line without continuous quality defects under historical operating conditions.

[0010] In response to the system control vulnerability index falling below the trust threshold, the first control strategy is output to the actuator of the injection molding production line;

[0011] In response to the system control vulnerability index being greater than or equal to the trust threshold, a preset robust control model is triggered to generate a second control strategy, which is then output to the actuator of the injection molding production line. The parameter change rate of the second control strategy is less than a preset safety change threshold. The second control strategy includes injection speed command, holding pressure command, and cooling time command.

[0012] Optionally, the step of acquiring time-series state data including mold temperature and cavity pressure collected by the underlying sensor network includes: extracting the original data collected by the underlying sensor network; identifying data delay identifiers and data loss identifiers in the original data collected; and reconstructing the features of the original data collected based on the data delay identifiers and the data loss identifiers using a preset time-series interpolation algorithm to generate the time-series state data.

[0013] Optionally, the step of weightedly calculating the system control vulnerability index based on the fluctuation characteristics determined by the variance value of the time-series state data within a preset time window, and the parameter change rate determined by the parameter deviation value of the first control strategy relative to the current operating parameters of the actuator of the injection molding production line, includes: extracting the variance value of the time-series state data within a preset time window as the fluctuation characteristics; calculating the parameter deviation value of the first control strategy relative to the current operating parameters of the actuator of the injection molding production line as the parameter change rate; multiplying the fluctuation characteristics by a first preset weighting coefficient to obtain a first characteristic value; multiplying the parameter change rate by a second preset weighting coefficient to obtain a second characteristic value; and adding the first characteristic value and the second characteristic value to calculate the system control vulnerability index.

[0014] Optionally, after calculating the system control vulnerability index, the method further includes: inputting the system control vulnerability index into a preset cascading failure prediction model; outputting a continuous failure risk probability based on the cascading failure prediction model; comparing the continuous failure risk probability with a preset failure boundary threshold; generating a forced shutdown command in response to the continuous failure risk probability being greater than or equal to the failure boundary threshold; and maintaining the output state of the current control strategy in response to the continuous failure risk probability being lower than the failure boundary threshold.

[0015] Optionally, the step of triggering a preset robust control model to generate a second control strategy includes: extracting historical thermodynamic state data of the injection molding production line; inputting the historical thermodynamic state data into the robust control model; based on the constraint condition module in the robust control model, shrinking the parameter boundary of the first control strategy to generate a candidate parameter set; selecting target parameters from the candidate parameter set that have the smallest energy consumption increment and whose parameter gradient change in adjacent time steps is less than a preset gradient threshold set based on the maximum allowable instantaneous impact limit of the equipment actuator, and generating the second control strategy.

[0016] Optionally, before inputting the time-series state data and the physical property fluctuation data into a preset reinforcement learning model, the method further includes: monitoring environmental micro-perturbation features in the time-series state data; matching the environmental micro-perturbation features with a preset pseudo-random fluctuation model; determining that sensor drift interference has occurred in response to the successful matching of the environmental micro-perturbation features with the pseudo-random fluctuation model; triggering a preset drift compensation algorithm to perform baseline calibration processing on the time-series state data in response to the occurrence of sensor drift interference, generating corrected time-series state data; inputting the corrected time-series state data into the reinforcement learning model as the time-series state data; and directly executing the step of inputting the time-series state data and the physical property fluctuation data into the preset reinforcement learning model in response to the failure of the matching of the environmental micro-perturbation features with the pseudo-random fluctuation model.

[0017] An industrial data processing system for automotive interior and exterior injection molding production lines includes: a data acquisition module, used to acquire time-series status data including mold temperature and cavity pressure collected by the underlying sensor network, as well as physical property fluctuation data of the input raw materials including melt index and moisture content;

[0018] The strategy generation module is used to input the time-series state data and the physical property fluctuation data into a preset reinforcement learning model to generate a first control strategy. The reinforcement learning model is trained using the minimization of the production cycle as the reward function. The state space of the reinforcement learning model is represented by a multi-dimensional state vector jointly composed of the time-series state data and the physical property fluctuation data. The action space of the reinforcement learning model is the continuous action adjustment range of the injection speed, holding pressure, and cooling time allowed by the actuators of the injection molding production line. The first control strategy includes injection speed commands, holding pressure commands, and cooling time commands.

[0019] The index calculation module is used to calculate the system control vulnerability index by weighting the fluctuation characteristics determined by the variance value of the time-series state data within a preset time window and the parameter change rate determined by the parameter deviation value of the first control strategy relative to the current operating parameters of the actuator of the injection molding production line.

[0020] The threshold comparison module is used to compare the system control vulnerability index with a preset trust threshold, wherein the preset trust threshold is obtained by joint distribution calibration based on the maximum state variance value and the maximum parameter safety control step size of the injection molding production line without continuous quality defects under historical operating conditions.

[0021] An adaptive execution module is configured to, in response to the system control vulnerability index being lower than the trust threshold, output the first control strategy to the actuator of the injection molding line; and in response to the system control vulnerability index being greater than or equal to the trust threshold, trigger a preset robust control model, generate a second control strategy, and output the second control strategy to the actuator of the injection molding line, wherein the parameter change rate of the second control strategy is less than a preset safety change threshold, and the second control strategy includes injection speed command, holding pressure command, and cooling time command.

[0022] Optionally, the data acquisition module includes: a data extraction unit for extracting raw acquisition data from the underlying sensor network; an identification unit for identifying data delay identifiers and data packet loss identifiers in the raw acquisition data; and a feature reconstruction unit for reconstructing features of the raw acquisition data based on the data delay identifiers and the data packet loss identifiers using a preset temporal interpolation algorithm to generate the temporal state data.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] 1. This invention generates a first control strategy by inputting acquired time-series state data and physical property fluctuation data into a reinforcement learning model with production cycle minimization as the reward function. Based on the fluctuation characteristics determined by the variance of the time-series state data and the parameter change rate corresponding to the first control strategy, a weighted solution is obtained to calculate the system control vulnerability index. The system compares this index with a trust threshold and dynamically decides whether to execute the first control strategy or trigger a robust control model to generate a second control strategy. This mechanism effectively overcomes the problem of traditional production lines having weak comprehensive judgment ability on data disturbances and parameter aggression. While shortening the production cycle, it can identify and actively converge the unstable trend of the control chain in advance, avoid parameter overshoot and thermal instability, and improve the ability of the production line to operate continuously and stably.

[0025] 2. To address the issue of communication anomalies in the underlying sensor network that can easily mislead the upper-level control system, this invention extracts the raw data acquired by the underlying sensor network, accurately identifies data delay and packet loss indicators, and uses a preset temporal interpolation algorithm for feature reconstruction to generate time-series state data. This mechanism can effectively distinguish between defects in the underlying network and real process disturbances, restoring the temporal consistency and sequence continuity of multidimensional state data, providing usable and accurate state input for subsequent model decisions, and reducing parameter tuning misjudgments caused by data delay and packet loss.

[0026] 3. When the control chain approaches the unstable boundary and triggers the robust control model, this invention extracts historical thermodynamic state data, shrinks the parameter boundary of the first control strategy based on the constraint condition module to generate a candidate parameter set, and selects the target parameter with the smallest energy consumption increment and parameter gradient change less than the preset gradient threshold to generate the second control strategy. This mechanism avoids the disruption of the forming rhythm caused by directly disabling the adaptive capability or simple rollback. Without completely abandoning the optimization of production cycle, it restricts the control action to a range that conforms to the thermal stability of the equipment, realizes the smooth transition of the control strategy, and effectively reduces the overall energy consumption of the production line.

[0027] 4. To address the problem that traditional methods easily lead to continuous quality defects, this invention inputs the system control vulnerability index into a preset cascading failure prediction model, outputs the probability of continuous failure risk, and automatically generates a forced shutdown command when the probability is greater than or equal to a preset failure boundary threshold. This mechanism elevates the vulnerability risk of a single control action to a comprehensive assessment from the perspective of multi-mode continuous production. It can complete defect warning and source blocking in advance before multi-mode cascading failures occur and the risk of large-scale downstream shutdowns is triggered, effectively avoiding the generation of continuous scrap and secondary damage to equipment. Attached Figure Description

[0028] Figure 1 This is a flowchart of the method of the present invention;

[0029] Figure 2 This is a structural diagram of the system of the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0031] Example 1:

[0032] Please see Figure 1 Industrial data processing methods for automotive interior and exterior injection molding production lines include:

[0033] Acquire time-series status data, including mold temperature and cavity pressure, collected by the underlying sensor network, as well as physical property fluctuation data, including melt index and moisture content, of the input raw materials;

[0034] The time-series state data and physical property fluctuation data are input into a preset reinforcement learning model to generate a first control strategy. The reinforcement learning model is trained with minimizing the production cycle as the reward function. The state space of the reinforcement learning model is represented by a multi-dimensional state vector composed of the time-series state data and physical property fluctuation data. The action space of the reinforcement learning model is the continuous action adjustment range of the injection speed, holding pressure, and cooling time allowed by the actuator of the injection molding production line. The first control strategy includes injection speed commands, holding pressure commands, and cooling time commands.

[0035] The system control vulnerability index is calculated by weighting the variance of the time-series state data within a preset time window and the parameter change rate determined by the parameter deviation of the first control strategy relative to the current operating parameters of the actuator of the injection molding production line.

[0036] The system control vulnerability index is compared with the preset trust threshold, which is obtained by joint distribution calibration based on the maximum state variance value and the maximum parameter safety control step size of the injection molding production line without continuous quality defects under historical operating conditions.

[0037] In response to the system control vulnerability index falling below the trust threshold, the first control strategy is output to the actuator of the injection molding production line;

[0038] In response to the system control vulnerability index being greater than or equal to the trust threshold, a preset robust control model is triggered to generate a second control strategy, which is then output to the actuator of the injection molding production line. The parameter change rate of the second control strategy is less than a preset safety change threshold. The second control strategy includes injection speed command, holding pressure command, and cooling time command.

[0039] This embodiment provides an industrial data processing mechanism for an injection molding production line for automotive bumpers and dashboards. Specifically, the production line operates in an unmanned continuous production state, using a high proportion of recycled plastics as raw materials. The underlying sensor network consists of a mold temperature sensor, a cavity pressure sensor, a raw material melt index detection unit, and a moisture content detection unit. The actuators include an injection machine injection servo assembly, a pressure holding hydraulic assembly, and a cooling circuit control assembly. This mechanism does not simply pursue the shortest cycle time, but rather, under cycle time compression, it determines in real time whether the current adaptive control has approached the instability boundary, and automatically switches from adjustment where the parameter change rate is greater than a first set value to adjustment where the parameter change rate is less than a second set value when necessary.

[0040] Specifically, the time-series state data can be continuously collected according to the injection molding cycle. Taking the three most recent sampling times of a certain mold batch as an example, the mold temperature is 88℃, 91℃, and 95℃ respectively, and the cavity pressure is 72MPa, 77MPa, and 79MPa respectively. The melt index of the same batch of raw materials is 12.4 and 13.1, and the moisture content is 0.42% and 0.58% respectively. The above data is organized into data frames aligned at the same time and then input into the reinforcement learning model. During the training phase, the model uses the minimization of the production cycle as the main reward. For example, a positive reward is given when the single mold cycle is compressed from 42 seconds to 39 seconds. At the same time, scrap penalty and energy consumption penalty can be superimposed, so that the model can learn the action rules of parameter adjustment in a large and small range under different working conditions in historical samples.

[0041] Furthermore, the reinforcement learning model outputs the first control strategy. For ease of explanation, assume the current operating parameters of the actuator are: injection speed 120 mm / s, holding pressure 85 MPa, and cooling time 14 s. The new parameters output by the model are: injection speed 132 mm / s, holding pressure 92 MPa, and cooling time 11 s. Then, three deviations can be obtained: 12 mm / s, 7 MPa, and 3 s. To unify the dimensions, the above deviations can be normalized according to the rated range of their respective equipment. Assuming the normalized values ​​are 0.10, 0.07, and 0.15, the average value of 0.106 can be taken as a solution result of the parameter change rate.

[0042] Simultaneously, the system extracts fluctuation characteristics from a preset time window. Assuming the time window covers the five most recent sampling points, the mold temperature sequence is 88, 91, 95, 93, 96, and the cavity pressure sequence is 72, 77, 79, 81, 78. The system calculates the variances separately, and after normalization, obtains a temperature fluctuation value of 0.18 and a pressure fluctuation value of 0.12. A weighted average is then taken to obtain a comprehensive fluctuation characteristic of 0.15. The system integrates the fluctuation characteristics with the parameter change rate according to preset weights. For example, if the fluctuation weight is 0.6 and the parameter change weight is 0.4, the system control vulnerability index can be calculated as 0.6 × 0.15 + 0.4 × 0.106 = 0.1324.

[0043] The trust threshold is obtained by jointly distributing and calibrating the maximum state variance value and the maximum parameter safety control step size of the injection molding production line without continuous quality defects under historical operating conditions. The index is compared with the trust threshold. If the threshold is set to 0.20, then 0.1324 is lower than the threshold, indicating that the current data disturbance and parameter change are still within an acceptable range. The system directly sends the first control strategy to the actuator to execute the control actions of injection acceleration, pressure holding increase and cooling compression.

[0044] If, at another moment, the mold temperature jumps to 89, 94, 101, 97, 103, and the cavity pressure simultaneously jumps to 70, 79, 84, 76, 85, causing the overall fluctuation characteristic to rise to 0.31, and the normalized deviation of the first control strategy from the current parameters reaches 0.22, then the fragility index can rise to 0.274. Since this value is higher than the threshold, the system no longer directly executes the first control strategy, but instead triggers the robust control model to generate a second control strategy. The second control strategy still includes injection speed, holding pressure, and cooling time, but the parameter changes at adjacent moments are actively compressed, for example, 132 mm / s, 92 MPa, 11 s are corrected to 126 mm / s, 88 MPa, 12.5 s, in order to reduce the instantaneous impact on the actuator.

[0045] As an anomaly handling mechanism, when any sensor lacks key data in the current cycle, or when the raw material property data fails to establish a correspondence with the current batch, the system does not directly enter the reinforcement learning decision-making process. Instead, it maintains the parameter set that has been verified to be safe in the previous cycle and waits for the data to be supplemented in the next sampling window. If a complete data frame cannot be formed in multiple consecutive windows, the vulnerability index is directly set to a high-risk state, and the robust control path is entered first. For example, if any parameter output by the reinforcement learning model exceeds the upper or lower limits allowed by the equipment, it is first truncated according to the equipment safety boundary before participating in subsequent vulnerability calculations to avoid execution layer anomalies caused by model optimization.

[0046] During the continuous night shift operation of the automotive bumper production line, the cycle time was compressed to less than 40 seconds to meet the continuous production needs of the vehicle manufacturer. The moisture content of a batch of recycled plastic increased from 0.43% to 0.61%. Based on this, the model attempted to increase the injection speed and shorten the cooling time. Initially, the mold temperature and pressure fluctuations were small, and the system determined that the fragility index was below the threshold. Therefore, the first control strategy was adopted, reducing the single mold cycle from 40.5 seconds to 39.2 seconds. The efficiency of the workshop cooling tower decreased, and the mold temperature curve began to fluctuate amplified. Although it could still meet the basic molding requirements in the short term, the system detected that the fluctuation characteristics and parameter change rate increased together, and the fragility index exceeded the threshold. Therefore, it automatically switched to the second control strategy, adjusting the cooling time back to a smoother range to avoid continuous quality defects such as warping and missing glue after continuous molding.

[0047] The purpose of this step is to incorporate the degree of data fluctuation and the magnitude of control action adjustment into the real-time judgment logic, thereby improving the production rate, identifying the unstable trend of the control chain in advance, and achieving continuous operation of the production line and risk convergence through two-level strategy switching.

[0048] The steps for acquiring time-series state data, including mold temperature and cavity pressure, collected by the underlying sensor network include: extracting the raw data collected by the underlying sensor network; identifying data delay identifiers and data loss identifiers in the raw data; and reconstructing the features of the raw data using a preset time-series interpolation algorithm based on the data delay identifiers and data loss identifiers to generate time-series state data.

[0049] This embodiment provides a mechanism for repairing underlying sensor network data. Specifically, in the aforementioned production line scenario, the programmable logic controller network in the industrial field has fixed delays and random packet loss. If the original data is directly sent into the control model, it will cause time-displaced temperature and pressure combinations, amplifying control misjudgments. Therefore, before forming time-series state data, the original data is first subjected to delay identification, packet loss identification, and feature reconstruction.

[0050] Specifically, the system extracts raw sampling records from the mold temperature channel and cavity pressure channel. Each record contains at least a sampling timestamp, data value, and communication status bit. For ease of explanation, it is assumed that in a certain injection molding cycle, the temperature sequence 90, 92, 94, 95 and the pressure sequence 73, 76, 78, 79 should be obtained at four times: t1, t2, t3, and t4. However, due to network latency, the pressure channel fails to report on time at t2 and instead re-sends the report at t3 with a delay flag. At the same time, the temperature channel loses packets at t4, and the status bit displays a packet loss flag. If this is not handled, a misalignment will occur where the t2 temperature matches the t1 pressure or there is no temperature at t4.

[0051] To this end, the system first traces back to the theoretical time to which the record should belong based on the delay identifier, and then reattaches the resent data to position t2. For sampling points with packet loss identifiers, a time-series interpolation algorithm is used for reconstruction. The interpolation method can be linear interpolation, spline interpolation, or first-order trend interpolation combining previous and subsequent operating conditions. Taking the temperature channel as an example, if t3 is 94℃, the next moment can refer to the historical trend of the same cooling section of the same model, which is 96℃, so the missing value of t4 can be reconstructed as 95℃. If packet loss occurs between two consecutive points, the temperature curve of the same process position in the previous cycle of the same batch can also be introduced as an aid. The same applies to the pressure channel. After the recovery time is aligned, the reconstructed complete sequence is obtained for use by subsequent models.

[0052] As an anomaly handling mechanism, if the number of consecutive packet losses within a certain window exceeds the preset limit, such as more than 2 out of 4 key sampling points being lost, the system will no longer use the interpolation results as the basis for actual control. Instead, it will mark the window as a low-confidence window and reduce its weight in subsequent strategy calculations. If the arrival time of delayed data exceeds the decision time limit of the current injection molding cycle, the data will no longer participate in the current cycle control and will only be used for offline update of state estimation to avoid expired information from contaminating online decision-making.

[0053] During the early morning shift of the instrument panel injection molding line, fluctuations in the cooling tower caused increased electromagnetic interference on site, resulting in a 2-second delay and approximately 15% random packet loss in the programmable logic controller network. The system detected a missing point t4 in the mold temperature channel and a delayed arrival of point t2 in the pressure channel. Therefore, the pressure value was first restored to the corresponding process position, and then the t4 temperature was reconstructed based on the temperature trends before and after the process and similar mold cycles. The timing data obtained in this way still maintains the physical sequence of mold temperature rise - pressure build-up - pressure holding and release, and will not misjudge communication abnormalities as molding abnormalities.

[0054] The purpose of this step is to distinguish between the defects in the underlying network and the actual process disturbances, restore time consistency and sequence continuity first, and then provide usable state inputs to the upper-level control model, thereby reducing the incorrect parameter tuning caused by delay and packet loss.

[0055] The steps for weightedly calculating the system control vulnerability index based on the fluctuation characteristics determined by the variance value of the time-series state data within a preset time window, and the parameter change rate determined by the parameter deviation value of the first control strategy relative to the current operating parameters of the actuator of the injection molding production line, include: extracting the variance value of the time-series state data within a preset time window as the fluctuation characteristics; calculating the parameter deviation value of the first control strategy relative to the current operating parameters of the actuator of the injection molding production line as the parameter change rate; multiplying the fluctuation characteristics by a first preset weighting coefficient to obtain a first characteristic value; multiplying the parameter change rate by a second preset weighting coefficient to obtain a second characteristic value; and adding the first characteristic value and the second characteristic value to calculate the system control vulnerability index.

[0056] This embodiment provides a vulnerability index calculation mechanism. Specifically, relying solely on the parameter tuning tendency of the model output to determine whether to execute is still insufficient to identify high-risk operating conditions, because some anomalies do not come from a single parameter being too large, but from the superposition of increased fluctuations and the rate of change of action parameters exceeding a set threshold. Therefore, this embodiment quantifies state fluctuations and parameter deviations simultaneously and merges them into a single index through fixed weights, which facilitates online comparison and execution switching.

[0057] Specifically, the system extracts the variance values ​​of each state variable within a preset time window. For ease of explanation, it is assumed that the window length is the last 5 sampling points. The mold temperature sequence is 91, 92, 93, 95, 99, with a mean of 94, and the variance can be obtained by discrete calculation. The cavity pressure sequence is 75, 76, 76, 80, 83, with a mean of 78, and the variance is also obtained. Considering the different dimensions of temperature and pressure, the system normalizes each extracted variance value by dividing it by the historical maximum allowable variance of the corresponding state variable. Specifically, for mold temperature and cavity pressure, the system calculates the normalized temperature variance and pressure variance respectively, and performs linear weighted summation using preset state weights, thereby transforming the multidimensional state variance values ​​into a unified fluctuation characteristic. For example, the temperature variance is converted to 0.20, the pressure variance is converted to 0.14, and then weighted by 0.5 to obtain a comprehensive fluctuation characteristic of 0.17.

[0058] The system calculates the deviation of the first control strategy relative to the current operating parameters. Assuming the current operating parameters are injection speed 118 mm / s, holding pressure 84 MPa, and cooling time 13 s, the model provides parameters of 130 mm / s, 91 MPa, and 11.5 s. If normalized according to the equipment tolerance range, the values ​​are 0.10, 0.07, and 0.10 respectively. Taking the average value, the parameter change rate is 0.09. If we further consider that the shortened cooling time has a stronger impact on thermal balance, we can also use different internal weights for the three deviations, such as injection speed 0.3, holding pressure 0.3, and cooling time 0.4. Then, the weighted parameter change rate can be 0.093.

[0059] The system multiplies the fluctuation characteristics by a first preset weighting coefficient and the parameter change rate by a second preset weighting coefficient. Assuming the first preset weighting coefficient is 0.65 and the second preset weighting coefficient is 0.35, the first characteristic value is 0.1105 and the second characteristic value is 0.03255. The sum of the two yields the system control vulnerability index of 0.14305. The higher this index, the more easily the current operating condition can be amplified by further parameter adjustments.

[0060] As an anomaly handling mechanism, if there is only single-point data or insufficient effective sampling within the time window, making it impossible to stably calculate the variance, the system adopts the variance value of the previous complete window and marks the current window as low confidence. If the current operating parameters are completely consistent with the first control strategy in a certain aspect, making the corresponding deviation zero, the remaining aspects are retained for continued calculation instead of terminating the whole-window calculation. If the fluctuation characteristics and parameter change rate are both close to zero, the system can increase the minimum protection value to prevent the index from remaining zero for a long time and masking potential risk turning points.

[0061] In a high-compression shift on the bumper production line, to compensate for the cycle time loss of the previous two molds, the model outputs a control command to increase the injection speed from 118mm / s to 130mm / s, while reducing the cooling time from 13s to 11.5s. If the optimization goal is only to shorten the production cycle, this strategy meets the extreme value condition. However, the system finds that the recent mold temperature variance has been higher than the daily average of the same batch, indicating that the mold thermal balance is becoming sensitive. By converting the variance characteristics and parameter deviations together into a vulnerability index, the system identifies that although the strategy has not yet exceeded the limit, it is close to the trust threshold, thus reserving time margin for control adjustment in subsequent switching.

[0062] The purpose of this step is to transform the abstract control action range into a calculable and comparable unified indicator, thereby enabling real-time recording and early constraint of unstable trends in the production line.

[0063] After calculating the system control vulnerability index, the process also includes: inputting the system control vulnerability index into a preset cascading failure prediction model; outputting the probability of continuous failure risk based on the cascading failure prediction model; comparing the probability of continuous failure risk with a preset failure boundary threshold; generating a forced shutdown command in response to the probability of continuous failure risk being greater than or equal to the failure boundary threshold; and maintaining the output state of the current control strategy in response to the probability of continuous failure risk being lower than the failure boundary threshold.

[0064] This embodiment provides a continuous failure early warning and shutdown safety protection mechanism. Specifically, knowing only that the current control vulnerability is too high is not enough to decide whether production must be stopped, because some highly vulnerable states can still transition stably after robust control. Therefore, after obtaining the vulnerability index, the system further predicts its risk contribution to continuous multi-mode severe defects in order to determine whether it has approached the unacceptable failure boundary of the production line.

[0065] Specifically, the system inputs the vulnerability index into the cascade failure prediction model. This model can be trained based on historical model sequences to determine whether the current fluctuation will form continuous defects in future models. For ease of explanation, it is assumed that the model input includes not only the current vulnerability index of 0.27, but also the quality labels of the three most recent models, such as normal, slight glue deficiency, edge warping, and contextual information such as cooling tower load and batch moisture content. In order to achieve accurate mapping of the above probabilities and improve the interpretability of the model, the cascade failure prediction model can adopt a logistic regression architecture that combines multidimensional feature linear weighting with nonlinear mapping.

[0066] Specifically, this architecture pre-assigns dynamic weights to each input, for example, setting the feature vector as... ,in, , , The corresponding weight vectors are: vulnerability index, quality tag quantification value, and cooling load normalization value. ,in, , , These are the weight coefficients corresponding to the features mentioned above, combined with the bias term. The overall failure trend value of the current production system is calculated by weighted summation. The operator ⋅ represents the vector dot product operation.

[0067] The overall failure trend value Input S-type activation function ,in, Using the base of the natural logarithm, this maps an infinite range of trend values ​​to a continuous probability of failure between 0 and 1. By clearly defining the calculation rules for data flow, the system can stably quantify the cascading risks resulting from the superposition of multiple fluctuations; the model ultimately outputs the probability of consecutive failure risks, for example, 0.62, indicating that the probability of consecutive severe defects occurring within the next 3 models is 62%.

[0068] If the failure boundary threshold is set to 0.55, then 0.62 is higher than the threshold, and the system generates a forced shutdown command. This shutdown command can first enter the safe easing process, for example, after the current holding pressure is completed, stop the next mold opening and feeding, and at the same time send a shutdown alarm to the upper monitoring system. If the output risk probability is 0.38, it means that although there is fluctuation, it has not yet reached the continuous failure danger zone. The system maintains the currently selected control strategy, continues to run, and re-evaluates in the next window.

[0069] As an anomaly handling mechanism, if the cascade failure prediction model cannot output a valid probability due to missing input context, the system will not directly allow high-risk control. Instead, it will adopt a conservative approach and compare the current vulnerability index with a lower temporary threshold. If the temporary threshold has been exceeded, robust control will be switched first and manual review will be notified simultaneously. If the vulnerability index is less than or equal to the temporary threshold, the current control strategy will be maintained. If the injection molding machine is in the high-pressure clamping stage when the shutdown command is generated, the shutdown should not immediately cut off the drive. Instead, the pressure relief, retraction, and cooling should be completed according to the equipment safety procedures to prevent secondary damage to the mold and parts.

[0070] In the instrument panel production line, two consecutive molds have shown a slight trend of insufficient glue, but have not yet reached the scrapping standard. At this time, due to the continued decline in cooling tower efficiency, the fragility index rises from 0.19 to 0.29. The system sends it to the cascade failure prediction model, and combined with the quality trend of the first two molds and the current high moisture content of the batch, it obtains a risk probability of 0.67 for the next three molds with consecutive serious defects. Although the current mold batch can still meet the minimum molding quality standard, the system judges that the probability of continuing production and triggering the downstream line stop risk is greater than the preset alarm extreme value. Therefore, a forced shutdown command is generated to block the spread of defects along the time sequence.

[0071] The purpose of this step is to elevate the risk of single-event control to a continuous production perspective for assessment, so as to complete early warning and defect prevention before multi-mode cascading failures actually occur.

[0072] The steps of triggering a preset robust control model and generating a second control strategy include: extracting historical thermodynamic state data of the injection molding production line; inputting the historical thermodynamic state data into the robust control model; based on the constraint condition module in the robust control model, shrinking the parameter boundary of the first control strategy to generate a candidate parameter set; selecting target parameters from the candidate parameter set that have the smallest energy consumption increment and whose parameter gradient change in adjacent time steps is less than a preset gradient threshold set based on the maximum allowable instantaneous impact limit of the equipment actuator, and generating the second control strategy.

[0073] This embodiment provides a second control strategy generation mechanism; specifically, the aforementioned switching logic can identify when it is not appropriate to continue using a wide range of parameter tuning, but if it simply reverts to fixed conservative parameters, it will cause the cycle time to drop suddenly, or even disrupt the original forming rhythm; therefore, this embodiment does not directly disable the adaptive capability after entering the robust control path, but generates a suboptimal parameter set with a more gradual change under thermodynamic constraints.

[0074] Specifically, the system extracts historical thermodynamic state data, including at least the mold temperature trajectory, cavity pressure build-up rate, screw position changes, inlet and outlet temperature difference of the cooling circuit, and cooling recovery time of the most recent molds. For ease of explanation, it is assumed that in the past 10 molds, the peak mold temperature during stable production was distributed between 94℃ and 97℃, the peak pressure was distributed between 80MPa and 86MPa, and the temperature difference of the cooling circuit was distributed between 5℃ and 7℃. When the first control strategy proposes an injection speed of 132mm / s, a holding pressure of 92MPa, and a cooling time of 11s, the physical logic of boundary contraction is as follows: taking the current actuator operating parameters as the center, according to the dynamic compression ratio that is proportional to the degree of deviation from the system's thermal state, the allowable extreme value range of each execution action is correspondingly reduced. The constraint condition module in the robust control model will first perform boundary contraction based on the historical thermal stability range, for example, compressing the acceptable range to an injection speed of 122 to 128mm / s, a holding pressure of 86 to 89MPa, and a cooling time of 12 to 13s, to obtain a set of candidate parameters.

[0075] The system filters from a set of candidate parameters. Assume the candidate set contains three sets of parameters: Group A (124 mm / s, 87 MPa, 12.8 s), Group B (126 mm / s, 88 MPa, 12.4 s), and Group C (128 mm / s, 89 MPa, 12.1 s). The system calculates the energy consumption increment relative to the current parameter and the gradient change between adjacent time steps for each set. If the current parameter is 121 mm / s, 86 MPa, 13.0 s, the gradient changes for the three sets can be interpreted as velocity changes of 3, 5, and 7; pressure changes of 1, 2, and 3; and cooling changes of 0.2, 0.6, and 0.9, respectively. If the preset gradient threshold requires the combined value of all changes to not exceed 6, then Group C is excluded. The system then compares the energy consumption increments of Group A and Group B. If Group A's energy consumption increases by 1.8% and Group B's by 1.2%, then Group B is selected as the target parameter, generating the second control strategy.

[0076] It should be noted that the above calculation logic for energy consumption increment is to multiply the differences between the candidate parameters and the current parameters by the corresponding equipment hardware energy consumption coefficients and then sum them up. The specific evaluation process includes: extracting the change in kinetic energy work corresponding to the injection action, the change in hydraulic work during the pressure holding stage, and the change in the long-term power consumption of the circulating water pump during the cooling stage, so as to clearly define the specific contribution of each parameter change to the overall energy consumption of the machine.

[0077] In the above simulation, although Group B has a slightly higher injection speed and holding pressure than Group A, which leads to an increase in kinetic energy and hydraulic work, its cooling time is significantly shortened from 12.8s to 12.4s, thereby reducing the power consumption of the cooling circulating water pump during the single mold cycle. Since the energy consumption of the water pump is dominant at this time, the two offset each other, resulting in Group B's 1.2% increase in overall energy consumption being less than Group A's 1.8% increase in overall energy consumption. This business logic details the simulation process of energy consumption changes and eliminates the nonlinear correlation error between parameter adjustment and energy consumption reduction. Therefore, the system ultimately selects Group B, with its smaller energy consumption increase, as the target parameter to generate the second control strategy.

[0078] The boundary shrinkage here is not a fixed compression ratio, but can change dynamically with the degree of thermal deviation. When the mold temperature drift is small and only slightly exceeds the threshold, the boundary shrinkage is small in order to preserve production capacity as much as possible. When the mold temperature drift is significant and accompanied by amplified pressure fluctuations, the boundary shrinkage increases, so that the parameter search is concentrated in a more stable physical range.

[0079] As an exception handling mechanism, if the candidate parameter set is empty, it means that there are no parameters that simultaneously satisfy thermal stability constraints and gradient thresholds near the first control strategy. The system can execute a two-level safety protection strategy: first, expand the search range of cooling time and generate the candidate set again; if it is still empty, directly call the safety baseline parameters that have been manually confirmed; if the energy consumption increment of multiple candidate groups is exactly the same, then the group with the smaller difference from the current parameter is selected first to reduce mechanical shock to the execution layer.

[0080] In the bumper production line, the reinforcement learning model aimed to compensate for the mold filling delay caused by the decrease in raw material fluidity by increasing the injection speed. However, mold temperature drift and increased equipment vibration were observed on site. After the system switched to robust control, it first read the thermodynamic trajectory of the most recent 10 molds and found that when the mold stayed above 97°C for more than the preset time threshold, the probability of subsequent warping exceeded the quality and safety threshold. Therefore, the system actively reduced the search boundary of cooling time and holding pressure. In the end, the system did not adopt the first control strategy with the largest adjustment range, but instead selected a second control strategy with a more gradual thermal change and lower energy consumption increment. This resulted in the production cycle only showing a recovery rate lower than the first preset ratio, but significantly reduced the risk of heat accumulation in the subsequent molds.

[0081] The purpose of this step is to limit the control actions to a range that is more in line with the thermal stability of the equipment, without completely abandoning the optimization of production cycle time, so as to achieve a smooth transition from wide-range adjustment control to highly robust control.

[0082] Before inputting the time-series state data and physical property fluctuation data into the preset reinforcement learning model, the method further includes: monitoring environmental micro-perturbation features in the time-series state data; matching the environmental micro-perturbation features with a preset pseudo-random fluctuation model; determining sensor drift interference in response to a successful match between the environmental micro-perturbation features and the pseudo-random fluctuation model; triggering a preset drift compensation algorithm to perform baseline calibration processing on the time-series state data in response to the occurrence of sensor drift interference, generating corrected time-series state data; inputting the corrected time-series state data into the reinforcement learning model as time-series state data; and directly executing the step of inputting the time-series state data and physical property fluctuation data into the preset reinforcement learning model in response to a failure to match the environmental micro-perturbation features with the pseudo-random fluctuation model.

[0083] This embodiment provides a sensor drift identification and compensation mechanism. Specifically, under extreme compression cycles, if the mold temperature or pressure sensor slowly shifts due to changes in the workshop microenvironment, the data curve monitored by the system will exhibit pseudo-random fluctuations that appear regular but do not actually reflect the true molding state. If such disturbances are not eliminated first, the reinforcement learning model may mistakenly believe that the process itself is unstable, thereby giving compensation actions that exceed the preset safety range. Therefore, drift identification and baseline calibration are performed before the model makes a decision.

[0084] Specifically, the system extracts environmental micro-perturbation features from time-series state data, such as low-frequency fluctuation amplitude, mean offset of adjacent windows, peak spacing distribution, and correlation with cooling tower load changes. For ease of explanation, it is assumed that the mold temperature fluctuates within ±1℃ at its peak value for each mold under normal operating conditions. However, in a certain shift, the peak values ​​of 20 consecutive molds show alternating fluctuations of +0.8℃, -0.6℃, +1.0℃, and -0.7℃, while the cavity pressure does not show a corresponding change synchronously. This phenomenon of temperature fluctuation but no synchronous pressure response may not be a true change in the process, but is closer to a sensor drift mode.

[0085] The system matches the micro-perturbation feature with a preset pseudo-random fluctuation model; this model can be established from historical drift samples, such as templates including low-frequency oscillation type, slow rise and fast fall type, and environmental coupling type.

[0086] To clearly define the matching process and improve algorithm interpretability, the matching score is calculated using the cosine similarity rule of multi-dimensional feature vectors. Specifically, the system combines the currently extracted environmental micro-perturbation features, such as amplitude variance, dominant frequency component, and mean shift, into a current feature vector. Simultaneously, it extracts the baseline feature vectors of each template. By calculating the dot product of the two vectors and dividing by the product of their magnitudes, the cosine similarity coefficient is obtained, and then linearly mapped to a matching score between 0 and 1. Assuming a matching score threshold of 0.75, if the matching score between the current feature and the environmentally coupled template is 0.81, sensor drift interference is determined to have occurred. If the matching score is less than or equal to the matching score threshold, the environmental micro-perturbation feature is determined to have failed to match the pseudo-random fluctuation model.

[0087] Trigger the drift compensation algorithm to perform baseline calibration on the time-series state data; the compensation method can be to use the peak pressure and the temperature difference of the cooling circuit in the same period as reference anchor points to deduce the range in which the temperature baseline should fall back; in specific calculations, calculate the corresponding theoretical drift-free state value based on the reference anchor point, and obtain the deviation between the current measured state value and the theoretical drift-free state value, and use the deviation as a compensation factor to inversely add to the time-series state data;

[0088] For example, if the average temperature of a certain window is determined to be 1.5℃ too high, then the window and subsequent consecutive windows will be gradually reverted to form corrected time-series state data, which will then be fed into the reinforcement learning model.

[0089] If the matching fails, for example, the score is only 0.42, it means that the current fluctuation is more likely to come from the actual process change, such as a sudden increase in the moisture content of the raw materials or insufficient filling of the mold. The system does not perform baseline calibration, but retains the original reconstructed time series state data and directly enters the subsequent control decision.

[0090] As an anomaly handling mechanism, if both temperature and pressure channels show anomalies in the same direction and are consistent with changes in raw material properties, the system will not immediately identify it as a sensor problem even if it detects some drift characteristics. Instead, it will mark it as a mixed disturbance to be verified and calibrate it in stages using a smaller compensation coefficient to avoid incorrectly compensating for real faults. If the data jumps too much after drift compensation, exceeding the physical reach of the same batch, the data will be rolled back to the data before compensation and reported for manual verification.

[0091] During the latter half of the night, when the dashboard production line was running continuously, the workshop air supply status changed, altering the heat dissipation conditions of the mold temperature sensor housing. The system observed regular fluctuations in the peak mold temperature, but the part weight and cavity pressure integral remained relatively stable. After matching with a pseudo-random fluctuation model, the system determined that the anomaly was more likely sensor drift than actual process instability. Therefore, baseline calibration was performed on the mold temperature sequence, bringing the excessively high measurement curve back to a range consistent with the pressure response. After calibration, the decision was then made by a reinforcement learning model to prevent the model from mistakenly triggering significant speed and pressure adjustments.

[0092] The purpose of this step is to first eliminate measurement biases introduced by environmental perturbations, and then optimize the parameters to reduce unnecessary parameter adjustments caused by pseudo-random fluctuations, thereby maintaining the stability and interpretability of the control chain.

[0093] Example 2:

[0094] Please see Figure 2 An industrial data processing system for automotive interior and exterior injection molding production lines includes: a data acquisition module, used to acquire time-series status data including mold temperature and cavity pressure collected by the underlying sensor network, as well as physical property fluctuation data of input raw materials including melt index and moisture content;

[0095] The strategy generation module is used to input time-series state data and physical property fluctuation data into a preset reinforcement learning model to generate a first control strategy. The reinforcement learning model is trained with minimizing the production cycle as the reward function. The state space of the reinforcement learning model is represented by a multi-dimensional state vector composed of the time-series state data and physical property fluctuation data. The action space of the reinforcement learning model is the continuous action adjustment range of the injection speed, holding pressure, and cooling time allowed by the actuator of the injection molding production line. The first control strategy includes injection speed commands, holding pressure commands, and cooling time commands.

[0096] The index calculation module is used to calculate the system control vulnerability index by weighting the fluctuation characteristics determined by the variance value of the time-series state data within a preset time window and the parameter change rate determined by the parameter deviation value of the first control strategy relative to the current operating parameters of the actuator of the injection molding production line.

[0097] The threshold comparison module is used to compare the system control vulnerability index with the preset trust threshold. The preset trust threshold is obtained by joint distribution calibration based on the maximum state variance value and the maximum parameter safety control step size of the injection molding production line without continuous quality defects under historical operating conditions.

[0098] An adaptive execution module is used to output a first control strategy to the actuator of the injection molding line in response to the system control vulnerability index being lower than the trust threshold; and to trigger a preset robust control model to generate a second control strategy in response to the system control vulnerability index being greater than or equal to the trust threshold, and output the second control strategy to the actuator of the injection molding line. The parameter change rate of the second control strategy is less than a preset safety change threshold. The second control strategy includes injection speed command, holding pressure command, and cooling time command.

[0099] This embodiment provides an industrial data processing system corresponding to the aforementioned method. Specifically, the system can be deployed in the edge control server of the injection molding workshop and communicate with the injection molding machine programmable logic controller, raw material detection unit, manufacturing execution system and upper monitoring terminal. The system realizes data acquisition, strategy generation, index calculation, threshold comparison and adaptive execution in a modular structure. Each module can run on the same industrial control host or run separately between the edge node and the upper server.

[0100] Specifically, the data acquisition module is responsible for receiving data such as mold temperature, cavity pressure, melt flow index, and moisture content from the underlying sensor network, and aligning and caching them according to the process cycle; the strategy generation module calls the reinforcement learning model to form a first control strategy based on the current cache frame; the index calculation module reads the state fluctuations in the most recent preset time window, as well as the difference between the first control strategy and the current operating parameters, and calculates the system control vulnerability index; the threshold comparison module compares the index with the trust critical threshold and outputs the judgment result of direct execution or switching robust control to the adaptive execution module; the adaptive execution module issues corresponding parameter instructions to the actuator according to the judgment result and collects execution feedback as part of the state of the next cycle;

[0101] For ease of explanation, the data processing of this system can be viewed as a continuous closed-loop data flow process. Assuming that the average temperature value sent by the data acquisition module is higher and the pressure fluctuation is greater during a certain period, the strategy generation module calculates a set of shortened-cycle parameters with a larger adjustment range. The index calculation module finds that the vulnerability index under the current operating condition has reached 0.24, which is higher than the threshold of 0.20. The threshold comparison module outputs a switching flag, and the adaptive execution module does not issue the first control strategy, but instead calls the robust control model to generate a second control strategy before sending it to the actuator. In this way, a closed loop of acquisition-judgment-switching-execution-feedback can be formed.

[0102] As an exception handling mechanism, if any module fails to output a result within a specified time, the system can perform downgrade processing according to the timeout level; for example, if the data acquisition module times out, the data snapshot that has been verified to be safe in the previous window will be used; if the strategy generation module times out, the adaptive execution module will directly call the most recent set of robust parameters; if the threshold comparison module is abnormal, a conservative execution path will be adopted by default to prevent large-scale parameter adjustments from being allowed without judgment; if the communication between the edge control server and the upper monitoring terminal is interrupted, the system can still maintain closed-loop operation locally and re-transmit logs after communication is restored.

[0103] During the cross-shift operation of the automotive bumper production line, the system in the edge control server continuously receives data from the injection molding machine and the raw material station. In the first half of the shift, due to the small fluctuations in raw materials, the fragility index remains below the threshold for a long period, and the adaptive execution module mainly outputs the first control strategy to pursue a high cycle time. In the second half of the shift, as cooling conditions deteriorate and sensor drift increases, the index calculation module frequently outputs high-risk results, and the system automatically adopts the second control strategy more often. Throughout the process, the operator terminal can see the status of each module and the reason for switching, which facilitates production management and maintenance tracking.

[0104] The purpose of this mechanism is to translate multi-stage decision-making at the methodological level into a deployable functional module structure, thereby achieving online closed-loop control for actual injection molding production lines.

[0105] The data acquisition module includes: a data extraction unit for extracting raw acquisition data from the underlying sensor network; an identification unit for identifying data delay identifiers and data packet loss identifiers in the raw acquisition data; and a feature reconstruction unit for reconstructing features of the raw acquisition data based on the data delay identifiers and data packet loss identifiers using a preset temporal interpolation algorithm to generate temporal state data.

[0106] This embodiment provides an internal implementation mechanism for a data acquisition module. Specifically, if the acquisition function is simply understood as data reception, in a scenario where traditional field-programmable logic controller networks, fixed latency, and random packet loss coexist, the data output upwards will be difficult to use directly for control. Therefore, this embodiment refines the acquisition module into a data extraction unit, an identification unit, and a feature reconstruction unit, enabling it to not only complete reception but also timing repair.

[0107] Specifically, the data extraction unit works with multiple underlying communication sources, periodically polling or subscribing to obtain raw collected data, and attaching a sampling time stamp, equipment number, process position, and status word to each record; the identification unit identifies delay and packet loss indicators from the status word, and can further correct the identification results based on time stamp anomalies; for example, theoretically, a record should arrive every 500ms. If a record arrives 2s late, even if the status word is not explicitly set, the system can automatically determine that it has a delay attribute based on the difference between the arrival time and the theoretical cycle time; the feature reconstruction unit performs time-series interpolation and process position backfilling on the identified abnormal data to form complete, continuous, and time-aligned time-series status data;

[0108] For ease of explanation, assume that the data extraction unit obtains the following four temperature records in a certain module: process point P1 is 90℃, P2 is missing, P3 is 94℃, and P4 is 96℃; the pressure records are P1 is 74MPa, P2 was supposed to arrive but was delayed until P4 to be re-issued at 76MPa, P3 is 79MPa, and P4 is 80MPa; the identification unit first reattaches the re-issued 76MPa to the P2 position, and then reconstructs the temperature P2 to 92℃ according to the trend between P1 and P3; after reconstruction, the output time sequence status data is: temperature 90, 92, 94, 96; pressure 74, 76, 79, 80; in this way, the upper-level module receives valid data arranged according to the process cycle, rather than chaotic data under the communication arrival order;

[0109] As an anomaly handling mechanism, if the identification unit finds that the same record is reported repeatedly and reported late at the same time, it will take the latest verified record as the valid value first, and other records will only be archived and will not participate in online control. If the feature reconstruction unit cannot find enough points before and after to perform interpolation within a cycle, it can call the average value of the historical same modulus position as a temporary filling value and attach a low confidence mark to the point so that the upper-layer module can reduce the weight of its use.

[0110] During shifts where the cooling system of the dashboard production line experiences increased fluctuations, the burden on the underlying network communication increases, and the original sampling sequence is frequently disrupted. After the data extraction unit accesses the data, the identification unit judges it, and the feature reconstruction unit repairs it, the output to the upper control module is still the mold temperature and pressure sequence aligned according to the injection molding process. In this way, even if there is obvious instability in the communication layer, the upper strategy generation and vulnerability judgment are still based on data that is as continuous and consistent as possible.

[0111] The purpose of this mechanism is to digest underlying communication anomalies within the acquisition module, thereby providing directly calculable timing status data to the upper-layer control logic, reducing the transmission and amplification of anomalies between modules.

[0112] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An industrial data processing method for automotive interior and exterior injection molding production lines, characterized in that, include: Acquire time-series status data, including mold temperature and cavity pressure, collected by the underlying sensor network, as well as physical property fluctuation data, including melt index and moisture content, of the input raw materials; The time-series state data and the physical property fluctuation data are input into a preset reinforcement learning model to generate a first control strategy. The reinforcement learning model is trained with minimizing the production cycle as the reward function. The state space of the reinforcement learning model is represented by a multi-dimensional state vector composed of the time-series state data and the physical property fluctuation data. The action space of the reinforcement learning model is the continuous action adjustment range of the injection speed, holding pressure, and cooling time allowed by the actuator of the injection molding production line. The first control strategy includes injection speed commands, holding pressure commands, and cooling time commands. Based on the fluctuation characteristics determined by the variance value of the time-series state data within a preset time window, and the parameter change rate determined by the parameter deviation value of the first control strategy relative to the current operating parameters of the actuator of the injection molding production line, the system control vulnerability index is calculated by weighting through preset weighting coefficients. The system control vulnerability index is compared with a preset trust threshold, wherein the preset trust threshold is obtained by joint distribution calibration based on the maximum state variance value and the maximum parameter safety control step size of the injection molding production line without continuous quality defects under historical operating conditions. In response to the system control vulnerability index falling below the trust threshold, the first control strategy is output to the actuator of the injection molding production line; In response to the system control vulnerability index being greater than or equal to the trust threshold, a preset robust control model is triggered to generate a second control strategy, which is then output to the actuator of the injection molding production line. The parameter change rate of the second control strategy is less than a preset safety change threshold. The second control strategy includes injection speed command, holding pressure command, and cooling time command.

2. The industrial data processing method for automotive interior and exterior injection molding production lines as described in claim 1, characterized in that, The steps for acquiring time-series state data, including mold temperature and cavity pressure, collected by the underlying sensor network include: extracting the original data collected by the underlying sensor network; identifying data delay identifiers and data packet loss identifiers in the original data collected; and reconstructing the features of the original data collected based on the data delay identifiers and the data packet loss identifiers using a preset time-series interpolation algorithm to generate the time-series state data.

3. The industrial data processing method for automotive interior and exterior injection molding production lines as described in claim 1, characterized in that, The step of weightedly calculating the system control vulnerability index based on the fluctuation characteristics determined by the variance value of the time-series state data within a preset time window, and the parameter change rate determined by the parameter deviation value of the first control strategy relative to the current operating parameters of the actuator of the injection molding production line, includes: extracting the variance value of the time-series state data within a preset time window as the fluctuation characteristics; calculating the parameter deviation value of the first control strategy relative to the current operating parameters of the actuator of the injection molding production line as the parameter change rate; multiplying the fluctuation characteristics by a first preset weighting coefficient to obtain a first characteristic value; multiplying the parameter change rate by a second preset weighting coefficient to obtain a second characteristic value; and adding the first characteristic value and the second characteristic value to calculate the system control vulnerability index.

4. The industrial data processing method for automotive interior and exterior injection molding production lines as described in claim 1, characterized in that, After calculating the system control vulnerability index, the method further includes: inputting the system control vulnerability index into a preset cascading failure prediction model; outputting a continuous failure risk probability based on the cascading failure prediction model; comparing the continuous failure risk probability with a preset failure boundary threshold; generating a forced shutdown command in response to the continuous failure risk probability being greater than or equal to the failure boundary threshold; and maintaining the output state of the current control strategy in response to the continuous failure risk probability being lower than the failure boundary threshold.

5. The industrial data processing method for automotive interior and exterior injection molding production lines as described in claim 1, characterized in that, The steps of triggering a preset robust control model and generating a second control strategy include: extracting historical thermodynamic state data of the injection molding production line; inputting the historical thermodynamic state data into the robust control model; based on the constraint condition module in the robust control model, shrinking the parameter boundary of the first control strategy to generate a candidate parameter set; selecting target parameters from the candidate parameter set that have the smallest energy consumption increment and whose parameter gradient change in adjacent time steps is less than a preset gradient threshold set based on the maximum allowable instantaneous impact limit of the equipment actuator, and generating the second control strategy.

6. The industrial data processing method for automotive interior and exterior injection molding production lines as described in claim 1, characterized in that, Before inputting the time-series state data and the physical property fluctuation data into the preset reinforcement learning model, the method further includes: monitoring environmental micro-perturbation features in the time-series state data; matching the environmental micro-perturbation features with a preset pseudo-random fluctuation model; determining that sensor drift interference has occurred in response to the successful matching of the environmental micro-perturbation features with the pseudo-random fluctuation model; triggering a preset drift compensation algorithm to perform baseline calibration processing on the time-series state data in response to the occurrence of sensor drift interference, generating corrected time-series state data; inputting the corrected time-series state data into the reinforcement learning model as the time-series state data; and directly executing the step of inputting the time-series state data and the physical property fluctuation data into the preset reinforcement learning model in response to the failure of the matching of the environmental micro-perturbation features with the pseudo-random fluctuation model.

7. An industrial data processing system for automotive interior and exterior injection molding production lines, characterized in that: include: The data acquisition module is used to acquire time-series status data, including mold temperature and cavity pressure, collected by the underlying sensor network, as well as physical property fluctuation data, including melt index and moisture content, of the input raw materials. The strategy generation module is used to input the time-series state data and the physical property fluctuation data into a preset reinforcement learning model to generate a first control strategy. The reinforcement learning model is trained using the minimization of the production cycle as the reward function. The state space of the reinforcement learning model is represented by a multi-dimensional state vector jointly composed of the time-series state data and the physical property fluctuation data. The action space of the reinforcement learning model is the continuous action adjustment range of the injection speed, holding pressure, and cooling time allowed by the actuators of the injection molding production line. The first control strategy includes injection speed commands, holding pressure commands, and cooling time commands. The index calculation module is used to calculate the system control vulnerability index by weighting the fluctuation characteristics determined by the variance value of the time-series state data within a preset time window and the parameter change rate determined by the parameter deviation value of the first control strategy relative to the current operating parameters of the actuator of the injection molding production line. The threshold comparison module is used to compare the system control vulnerability index with a preset trust threshold, wherein the preset trust threshold is obtained by joint distribution calibration based on the maximum state variance value and the maximum parameter safety control step size of the injection molding production line without continuous quality defects under historical operating conditions. An adaptive execution module is configured to, in response to the system control vulnerability index being lower than the trust threshold, output the first control strategy to the actuator of the injection molding line; and in response to the system control vulnerability index being greater than or equal to the trust threshold, trigger a preset robust control model, generate a second control strategy, and output the second control strategy to the actuator of the injection molding line, wherein the parameter change rate of the second control strategy is less than a preset safety change threshold, and the second control strategy includes injection speed command, holding pressure command, and cooling time command.

8. The industrial data processing system for automotive interior and exterior injection molding production lines as described in claim 7, characterized in that, The data acquisition module includes: a data extraction unit for extracting raw acquisition data from the underlying sensor network; an identification unit for identifying data delay identifiers and data packet loss identifiers in the raw acquisition data; and a feature reconstruction unit for reconstructing features of the raw acquisition data based on the data delay identifiers and the data packet loss identifiers using a preset temporal interpolation algorithm to generate the temporal state data.

Images