Spring steel wire cutting command automatic control system based on industrial data analysis

Abstract

The present application relates to industrial automation control, industrial data analysis and artificial intelligence in the field of intelligent manufacturing technology, specifically to a spring steel wire cutting command automatic control system based on industrial data analysis. It comprises: a data acquisition module for acquiring production rhythm data; a first feature processing module for generating a production rhythm feature vector; a second feature processing module for generating a tool health degree feature vector; a third feature processing module for generating a physical credibility feature vector; a state prediction module for predicting the probability distribution of the future state of the system; a risk quantification module for calculating the safety distance of the failure boundary; a decision generation module for generating macro risk avoidance actions; a parameter optimization module for optimizing the generation of micro process parameter vectors; and an instruction generation module for converting the micro process parameter vectors into machine control commands sent to the spring steel wire cutting system. The present application solves the defect that traditional static parameter control cannot cope with hidden interference sources such as hydraulic oil emulsification.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of industrial automation control, industrial data analysis and artificial intelligence in intelligent manufacturing technology, in particular to a spring steel wire cutting command automatic control system based on industrial data analysis. BACKGROUND

[0002] In the field of industrial automation production, the maintenance of stable operation of equipment and the guarantee of product quality consistency are the core technical requirements in the precision machining link such as spring steel wire cutting. At present, the control strategy of spring steel wire cutting system generally depends on the static process parameters based on experience. This control method will pre-configure fixed cutting speed, feed amount and other parameters before starting production, and keep them unchanged in a production cycle.

[0003] This traditional static parameter control method has obvious limitations when dealing with dynamic changes in the production process. It is difficult to effectively handle hidden interference sources caused by the evolution of the equipment itself, such as progressive wear of the tool and emulsion contamination of hydraulic oil. These factors will continue to accumulate and cannot be covered by static parameters, which may eventually lead to batch quality deviation of products, even cause unexpected equipment downtime, and seriously affect the continuity and economy of production.

[0004] Although the prior art may contain some basic equipment state monitoring functions, it generally lacks the ability to deeply fuse and analyze multi-source sensing data and predict future state evolution. Specifically, the existing method has the following deficiencies: there is a lack of effective modeling means for how to accurately quantify the tool health degree, oil physical reliability and other key implicit states from multi-modal data such as high-frequency vibration and oil turbidity; the existing system lacks a forward-looking risk assessment mechanism that can predict the failure probability of the equipment in the future; at the decision-making level, an intelligent decision-making system that can dynamically balance production rhythm and equipment safety according to the quantified future risk, and autonomously generate optimal macro risk avoidance actions and micro process parameters, has not been established.

[0005] Therefore, how to provide a spring steel wire cutting command automatic control method based on industrial data analysis, which realizes accurate prediction and dynamic avoidance of system future risks by fusing multi-modal sensing data, so as to change from passive static control to active adaptive control, is a technical problem that needs to be solved by those skilled in the art. SUMMARY

[0006] To solve the above technical problems, the present application discloses a spring steel wire cutting command automatic control system based on industrial data analysis, in particular, the technical scheme is as follows:

[0007] A spring steel wire cutting command automatic control system based on industrial data analysis, comprising:

[0008] a data acquisition module, configured to collect multi-modal sensor data of the spring wire cutting system in real time and acquire production rhythm data;

[0009] a first feature processing module, configured to generate a production rhythm feature vector based on the production rhythm data;

[0010] a second feature processing module, configured to generate a tool health degree feature vector based on the multi-modal sensor data;

[0011] a third feature processing module, configured to generate a physical credibility feature vector based on the multi-modal sensor data;

[0012] a state prediction module, configured to combine a historical sequence of the production rhythm feature vector, the tool health degree feature vector, and the physical credibility feature vector to predict a probability distribution of a future state of the system;

[0013] a risk quantification module, configured to combine the probability distribution of the future state of the system with a preset failure boundary value to calculate a failure boundary safety distance;

[0014] a decision generation module, configured to generate a macro risk avoidance action based on the failure boundary safety distance and the three feature vectors;

[0015] a parameter optimization module, configured to generate a micro process parameter vector in response to the macro risk avoidance action;

[0016] an instruction generation module, configured to convert the micro process parameter vector into a machine control command sent to the spring wire cutting system.

[0017] Preferably, the third feature processing module is specifically configured to:

[0018] calculate an oil emulsification index based on an oil sensor reading in the multi-modal sensor data and a preset weighted normalization model;

[0019] and use the oil emulsification index as a component for generating the physical credibility feature vector.

[0020] Preferably, the second feature processing module is specifically configured to:

[0021] perform wavelet packet transform on a vibration signal in the multi-modal sensor data to extract an energy proportion of each frequency band;

[0022] and use the energy proportion as a component for generating the tool health degree feature vector.

[0023] Preferably, the state prediction module is specifically configured to:

[0024] A spatiotemporal graph convolutional network is used to perform spatiotemporal dependency modeling on the historical sequences of the production cycle feature vector, the tool health feature vector, and the physical reliability feature vector, and output the hidden state of the system at the current moment.

[0025] The evolution of the hidden state is modeled using neural stochastic differential equations to calculate the probability distribution of the future state of the system.

[0026] Preferably, the risk quantification module is specifically used for:

[0027] The hidden state output by the state prediction module is mapped back to the observable feature space through the trained decoder network to determine the future prediction mean and standard deviation of the physical credibility feature.

[0028] The failure boundary safety distance is determined based on the future prediction mean, the standard deviation, and the preset failure boundary value.

[0029] Preferably, the decision generation module is specifically used for:

[0030] Determine the normalized utility term calculated from the production cycle feature vector, the tool health feature vector, and the physical reliability feature vector;

[0031] Determine the exponential penalty term calculated from the failure boundary safety distance;

[0032] A reward function is constructed by combining the normalized utility term and the exponential penalty term;

[0033] Based on the reward function, the macro risk aversion action is generated.

[0034] Preferably, the decision generation module is further configured to:

[0035] When the failure boundary safety distance is lower than the preset minimum safety threshold, the macro risk avoidance action is generated by having the exponential penalty term play a dominant role in the reward function.

[0036] When the failure boundary safety distance is not lower than the preset minimum safety threshold, the macro-risk avoidance action is generated by the normalized utility term playing a dominant role in the reward function.

[0037] Preferably, the parameter optimization module is specifically used for:

[0038] Based on the macro-risk avoidance actions generated by the decision generation module as constraints, a multi-objective optimization problem is constructed;

[0039] The multi-objective grey wolf optimization algorithm is adopted, and a preset proxy model is used to solve the multi-objective optimization problem to output the micro-process parameter vector.

[0040] Preferably, the spring steel wire cutting command automatic control system based on industrial data analysis further comprises:

[0041] A feedback calibration module is configured to calculate a long-term effective output index based on the physical credibility feature vector and the production tact feature vector.

[0042] And the weights used to construct the reward function are periodically calibrated according to the long-term effective output index.

[0043] Compared with the prior art, the present application has the following beneficial effects:

[0044] 1. The present application realizes forward-looking risk management by predicting the future state probability of the equipment, can quantify failure risks and generate avoidance actions before failure occurs, and solves the defects that traditional static parameter control cannot cope with hidden interference sources such as hydraulic oil emulsification. 2. The present application deeply quantifies the hidden state such as tool wear and oil pollution which is not easy to detect by performing wavelet packet transform on the vibration signal and calculating the oil emulsification index, realizes fine perception of the physical health condition of the equipment, and avoids batch defective products and unexpected downtime caused by deterioration of hidden faults. 3. The present application has the ability of double-level closed-loop control of tactical response and strategic adaptation, can not only optimize micro-process parameters in real time according to the current risk, but also periodically calibrate the decision model based on the long-term effective output index, so that the system decision preference evolves continuously to maximize long-term benefits. 4. The present application verifies the effectiveness of the input data, tests the core prediction model, and adds a protection mechanism in the risk quantification calculation, ensuring that the system can provide robust prediction results under non-ideal working conditions, significantly improving the robustness of the production process. BRIEF DESCRIPTION OF DRAWINGS

[0045] The present application will be further explained in conjunction with the accompanying drawings and examples:

[0046] Figure 1 The present application will be further explained in conjunction with the accompanying drawings and examples: DETAILED DESCRIPTION

[0047] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further explained in conjunction with specific examples. Example 1

[0048] Please refer to Figure 1 A spring steel wire cutting command automatic control system based on industrial data analysis comprises:

[0049] The data acquisition module is configured to acquire multi-modal sensor data of the spring wire cutting system and production rhythm data in real time.

[0050] The first feature processing module is configured to generate a production rhythm feature vector based on the production rhythm data.

[0051] The second feature processing module is configured to generate a tool health degree feature vector based on the multi-modal sensor data.

[0052] The third feature processing module is configured to generate a physical credibility feature vector based on the multi-modal sensor data.

[0053] The state prediction module is configured to predict a probability distribution of a future state of the system in combination with a historical sequence of the production rhythm feature vector, the tool health degree feature vector, and the physical credibility feature vector.

[0054] The risk quantification module is configured to calculate a failure boundary safety distance in combination with the probability distribution of the future state of the system and a preset failure boundary value.

[0055] The decision generation module is configured to generate a macro risk avoidance action in accordance with the failure boundary safety distance and the three feature vectors.

[0056] The parameter optimization module is configured to optimize a micro process parameter vector in response to the macro risk avoidance action.

[0057] The instruction generation module is configured to convert the micro process parameter vector into a machine control command sent to the spring wire cutting system. Specifically, the conversion includes formatting the optimized process parameter vector according to a communication protocol required by a target device controller, and packaging the optimized process parameter vector into a binary or text instruction code that can be directly executed.

[0058] The embodiment provides a spring wire cutting command automatic control system based on industrial data analysis. The data acquisition module is responsible for acquiring multi-modal sensor data such as high-frequency vibration and oil turbidity, and synchronously acquiring production rhythm data.

[0059] The first feature processing module generates a production rhythm feature vector based on the production rhythm data. The second feature processing module generates a tool health degree feature vector based on the multi-modal sensor data. The third feature processing module generates a physical credibility feature vector based on the multi-modal sensor data.

[0060] The state prediction module receives and combines a historical sequence of the three feature vectors to predict a probability distribution of a future state of the system. The risk quantification module calculates a failure boundary safety distance based on the probability distribution.

[0061] The decision generation module generates a macro risk avoidance action according to the failure boundary safety distance and the three characteristic vectors; the parameter optimization module generates a micro process parameter vector in response to the action; and the instruction generation module finally converts the parameter vector into a machine control command;

[0062] The system converts the original physical signal into accurate adaptive control instructions through the chain data processing flow, aims to solve the defects of traditional static parameter control that cannot cope with hidden interference sources such as hydraulic oil emulsification, and thus realizes the forward-looking risk management of the cutting process. The current model of the system mainly focuses on the internal risks caused by the device state and production rhythm. In future iterations, raw material batch information, environmental temperature and humidity sensor data, etc. can also be used as additional inputs and included in the modeling of the spatio-temporal graph convolution network, so as to further improve the adaptability and robustness of the system to external interference factors.

[0063] Before feature processing, the data acquisition module or the third feature processing module first performs validity verification on the multi-modal sensor data; the verification rules include checking whether the sensor reading is within its preset valid range , and detecting whether the data has abnormal conditions such as long-term no change or severe jitter; for the sensor data identified as invalid or abnormal, the system can use the historical mean value of the reading or an estimated value based on other related sensor readings to fill in, and generate a system alarm to ensure the robustness of the physical credibility characteristic vector. Embodiment 2

[0064] The third feature processing module is specifically configured to:

[0065] Based on the oil sensor reading in the multi-modal sensor data and a preset weighted normalization model, an oil emulsification index is calculated;

[0066] and the oil emulsification index is used as a component of the generation of the physical credibility characteristic vector;

[0067] The second feature processing module is specifically configured to:

[0068] Wavelet packet transform is performed on the vibration signal in the multi-modal sensor data to extract the energy proportion of each frequency band;

[0069] and the energy proportion is used as a component of the generation of the tool health degree characteristic vector

[0070] In this embodiment, the second feature processing module and the third feature processing module work together to quantize the system's implicit state in depth. The weight coefficients in the preset weighted normalization model are calibrated based on a large amount of historical data. To clarify the calibration process, it is necessary to distinguish between calibration data variables and model running variables. The calibration data set is composed of a group of oil samples constituted, each sample has a benchmark emulsion value determined by offline chemical analysis and a set of synchronously collected sensor reading vectors, weights are solved on the calibration dataset by least square regression analysis, targeting to minimize the error between the oil emulsion index calculated by the following formula and the benchmark emulsion value The third feature processing module calculates the oil emulsion index Ie through the model, whose formula is:

[0071] The third feature processing module calculates the oil emulsion index Ie through the model, whose formula is:

[0072]

[0073] Where Ie is the oil emulsion index, ωi is the weight, Si is the sensor reading, Sn is the lower limit of the range, Sx is the upper limit of the range, and i is the sensor index.

[0074] It should be noted that the aforementioned linear weighted model for calculating the oil emulsion index provides an efficient and effective approximation method for actual engineering applications. In future iterations, nonlinear models can also be used to capture more complex coupling relationships between features, further improving the accuracy of physically credible features.

[0075] At the same time, the second feature processing module uses wavelet packet transform technology to decompose the vibration signal and extract the energy proportion. This can accurately capture the energy anomaly of the tool due to wear or broken blade in a specific frequency band. By accurately quantifying the degree of oil emulsification and the micro-vibration characteristics of the tool, the system has a fine perception ability of the physical health status of the equipment, avoiding the production of batches of substandard products and unexpected downtime due to the continuous deterioration of hidden faults. Embodiment 3

[0076] The state prediction module is specifically configured to:

[0077] The spatio-temporal graph convolution network is used to model the spatio-temporal dependence of the production rhythm feature vector, the tool health degree feature vector, and the history sequence of the physically credible feature vector, and output the hidden state of the system at the current time.

[0078] The neural stochastic differential equation is used to model the evolution of the hidden state to solve the probability distribution of the future state of the system.

[0079] The risk quantification module is specifically configured to:

[0080] The trained decoder network is used to map the hidden state output by the state prediction module back to the observable feature space to determine the future predicted mean and standard deviation of the physically credible feature. ​

[0081] Based on the future prediction mean, standard deviation and the preset failure boundary value, the failure boundary safety distance is determined;

[0082] In this embodiment, the state prediction module and the risk quantification module jointly constitute the prediction core unit of the system, and the spatio-temporal graph convolution network is a preset model, the graph structure of which is predefined according to the association of the physical components of the cutting system, to capture the coupling effect between features. For example, the key components such as the main shaft, the tool and the hydraulic pump can be taken as the nodes of the graph, and the edges are defined according to the physical connection or functional coupling relationship between them, so as to construct a model reflecting the topological structure of the system. Specifically, the adjacency matrix A of the graph is defined asymmetrically and weighted according to the energy or information flow between components. For example, the edge weight from the hydraulic pump to the main shaft is higher than that from the main shaft to the hydraulic pump.

[0083] The network adopts a stacked structure of multi-layer graph convolution and gated recurrent unit (GRU) to capture spatial dependence and temporal dynamics at the same time. The neural stochastic differential equation is used to predict the future evolution of the hidden state in the presence of random disturbances. The specific form of the equation is where the hidden state is The Wiener process is The drift term and the diffusion term are parameterized by independent multi-layer perceptron (MLP) neural networks and trained by maximizing the evidence lower bound (ELBO) of the observation sequence. The function of the trained decoder network is to translate the abstract hidden state back to the specific physical feature prediction. The decoder network is preferably a multi-layer perceptron (MLP) structure, the input of which is the future time hidden state output by the neural stochastic differential equation The output is the predicted mean and the predicted standard deviation

[0084] The risk quantification module calculates the failure boundary safety distance based on the prediction The physical meaning is that the distance between the predicted feature mean and the failure boundary is how many times the predicted standard deviation, which is a dimensionless relative index. The preset failure boundary value is obtained from the equipment safety manual or a large number of destructive experiments, which is a hard engineering constraint; its calculation formula is:

[0085] Failure boundary safety distance Failure boundary threshold, Predicted mean Predicted standard deviation

[0086] In actual calculation, in order to prevent numerical instability caused by too small prediction standard deviation , a small positive number lower limit is set for it , that is, the actual calculation adopts , which ensures that even when the model prediction is highly certain, the safety distance is still a bounded and meaningful value;

[0087] The technical solution upgrades the system from passive response to active prediction, provides a decision basis for taking evasive action before failure occurs by quantifying the probability of touching the failure boundary within a specific future time, and significantly improves the robustness of the production process. In addition, to ensure the robustness of the core prediction model, the model will be systematically stress tested before deployment. The test content includes: injecting simulated extreme noise, long-period flat signal or sudden step signal into the historical sequence input, observing and verifying whether the hidden state output and future state probability distribution of the model are maintained within a reasonable range, and whether numerical divergence or collapse occurs, to ensure that the model can still provide robust prediction results under non-ideal working conditions. Embodiment 4

[0088] The decision generation module is specifically configured to:

[0089] determine the normalized utility term calculated by the production beat feature vector, the tool health degree feature vector and the physical credibility feature vector;

[0090] determine the exponential penalty term calculated by the failure boundary safety distance;

[0091] combine the normalized utility term and the exponential penalty term to construct a reward function;

[0092] and generate a macro risk avoidance action based on the reward function;

[0093] The decision generation module is also configured to:

[0094] When the failure boundary safety distance is lower than the preset minimum safety threshold, the macro risk avoidance action is generated in a way that the exponential penalty term plays a leading role in the reward function;

[0095] When the failure boundary safety distance is not lower than the preset minimum safety threshold, the macro risk avoidance action is generated in a way that the normalized utility term plays a leading role in the reward function

[0096] In this embodiment, the decision generation module is the decision unit of the system, and the reward function R constructed by it aims to balance production efficiency and safety risk. The preset minimum safety threshold is set according to the risk acceptance criteria and historical safety operation data analysis, and the weight are calibrated by iterative optimization in simulation environment or by inverse reinforcement learning method to maximize long-term effective output, where the normalized utility term is the quantification of different performance dimensions of the system, for example, the utility term representing production efficiency can directly take the numerical value of the normalized production rhythm feature vector; the utility term representing quality can be defined as 1 minus the normalized tool wear or oil emulsification degree;

[0097] For example, the production rhythm feature vector the tool health degree feature vector the physical credibility feature vector where f and g are preset mapping functions, for example, the utility term representing production efficiency can be defined by a specific normalization function. The utility term representing production efficiency can be calculated by the following linear mapping function f:

[0098]

[0099] where and are the maximum and minimum production rhythm allowed by the process, respectively. The utility term representing equipment state can be calculated by the weighted average function of the normalized tool health degree feature and the physical credibility feature : where the weight is determined by the importance of the two features. In this way, all the utility terms are explicitly mapped to the interval [0, 1], and the higher the value, the higher the utility; the formula is: the reward function, the weight, the normalized utility term, the exponential penalty term the utility term index);

[0100] The design of the exponential penalty term reflects the avoidance of asymmetric risk, and its value increases exponentially when the safety distance of the failure boundary is lower than the preset minimum safety threshold , and its specific form can be designed as where is a positive coefficient used to adjust the severity of the penalty, and its value can be determined by simulation experiments. Under the premise of ensuring safety, a reasonable value that does not excessively suppress normal production is selected, for example, it can be set as and decrease by one standard unit, the penalty term grows by an order of magnitude; this condition judgment logic ensures the survival priority principle of the system: within the safe region, the system is dominated by the pursuit of production utility; once entering the dangerous region, the reward function is dominated by the penalty term, and the system will be forced to perform risk-avoiding actions such as reducing the production rhythm or preventive tool change, thereby avoiding the disastrous consequences of lucky production behavior on the edge of failure. Embodiment 5

[0101] The parameter optimization module is specifically configured to:

[0102] construct a multi-objective optimization problem according to the macro risk-avoiding action generated by the decision generation module as a constraint condition;

[0103] adopt a multi-objective grey wolf optimization algorithm, which is suitable for the real-time optimization demand of the system on the process parameters due to its fast convergence speed and strong optimization ability in dealing with multi-objective optimization problems with complex constraint conditions, and solve the multi-objective optimization problem based on a preset proxy model to output a micro process parameter vector;

[0104] The feedback calibration module is configured to calculate a long-term effective output indicator based on the physical credibility feature vector and the production rhythm feature vector;

[0105] and periodically calibrate the weights used to construct the reward function according to the long-term effective output indicator;

[0106] In this embodiment, the parameter optimization module and the feedback calibration module realize double-level closed-loop control, wherein the process of converting the macro risk-avoiding action into a constraint condition is specifically as follows: if the action is to reduce the production rhythm, the upper limit of the production rhythm parameter is set to a safe value lower than the current value in the multi-objective optimization problem; if the action is preventive tool change, the system generates a maintenance instruction and suspends optimization, and then resumes after the tool change is completed; if the action is to adjust the cutting fluid ratio, the optimization range of the cutting fluid related parameters is constrained in a more conservative interval; the preset proxy model is a lightweight neural network trained based on historical data, which is used to quickly predict the production results under different process parameter combinations to support the multi-objective grey wolf optimization algorithm for real-time optimization, the input of the proxy model is a micro process parameter vector containing parameters such as cutting speed and feed rate, and the output is the predicted values of multiple performance indicators such as production rhythm, product size accuracy, and surface roughness under the parameters; the feedback calibration module works on a longer time scale, and the long-term effective output indicator L calculated by the feedback calibration module is a comprehensive evaluation of the long-term performance of the system, which has the dimension of output quantity and the physical meaning of the total number of qualified products produced in the total time, and the calculation formula is:

[0107]

[0108] wherein L is the long-term effective output indicator, for the rate of defective products, for the production rhythm, t for the time, T for the total duration;

[0109] The feedback calibration module periodically adjusts the weight in the reward function according to the result of the long-term effective output index L. The double-layer feedback mechanism enables the system to not only make the optimal tactical response to the current working condition, but also constantly optimize its decision preference according to the long-term effect, realizing the evolution from tactical optimization to strategic self-adaptation.

[0110] The above merely illustrates the preferred embodiments of the present application and is not intended to limit the protection scope of the present application; any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0111] It should be noted that the above embodiments are merely used to illustrate the technical solutions of the present application but not to limit the protection scope of the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the present application.

Claims

1. A spring steel wire cutting command automatic control system based on industrial data analysis, characterized by, The method comprises the following steps: a data acquisition module for real-time acquisition of multi-modal sensor data of a spring wire cutting system and acquisition of production rhythm data; a first feature processing module for generating a production rhythm feature vector based on the production rhythm data; a second feature processing module for generating a tool health degree feature vector based on the multi-modal sensor data; a third feature processing module for generating a physical credibility feature vector based on the multi-modal sensor data; a state prediction module for predicting the probability distribution of the future state of the system in combination with the historical sequence of the production rhythm feature vector, the tool health degree feature vector, and the physical credibility feature vector; a risk quantification module for calculating the failure boundary safety distance in combination with the probability distribution of the future state of the system and the preset failure boundary value; a decision generation module for generating a macro risk avoidance action according to the failure boundary safety distance and the three feature vectors; a parameter optimization module for optimizing the generation of a micro process parameter vector in response to the macro risk avoidance action; an instruction generation module for converting the micro process parameter vector into a machine control command sent to the spring wire cutting system; The third feature processing module is specifically configured to: calculate an oil emulsion index based on the oil sensor readings in the multi-modal sensor data and a preset weighted normalization model; and the oil emulsion index is used as a component of the physical credibility feature vector; The second feature processing module is specifically configured to: perform wavelet packet transform on the vibration signal in the multi-modal sensor data to extract the energy proportion of each frequency band; and the energy proportion is used as a component of the tool health degree feature vector; The state prediction module is specifically configured to: use a spatio-temporal graph convolution network to perform spatio-temporal dependence modeling on the historical sequence of the production rhythm feature vector, the tool health degree feature vector, and the physical credibility feature vector, and output the hidden state of the system at the current time; use a neural stochastic differential equation to model the evolution of the hidden state to solve the probability distribution of the future state of the system; The risk quantification module is specifically configured to: map the hidden state output by the state prediction module back to the observable feature space through a trained decoder network to determine the future prediction mean and standard deviation of the physical credibility feature; determine the failure boundary safety distance based on the future prediction mean, the standard deviation, and the preset failure boundary value; The decision generation module is specifically configured to: determine a normalized utility term calculated from the production rhythm feature vector, the tool health degree feature vector, and the physical credibility feature vector; determine an exponential penalty term calculated from the failure boundary safety distance; construct a reward function in combination with the normalized utility term and the exponential penalty term; and generate the macro risk avoidance action based on the reward function.

2. The spring wire cutting command automatic control system based on industrial data analysis according to claim 1, characterized in that, The decision generation module is also configured to: when the failure boundary safety distance is lower than a preset minimum safety threshold, generating the macro risk-averse action in a manner that the exponential penalty term plays a leading role in the reward function; when the failure boundary safety distance is not lower than the preset minimum safety threshold, generating the macro risk-averse action in a manner that the normalized utility term plays a leading role in the reward function.

3. The spring wire cutting command automatic control system based on industrial data analysis according to claim 1, characterized in that, The parameter optimization module is specifically configured to: construct a multi-objective optimization problem according to the macro risk-averse action generated by the decision generation module as a constraint condition; solve the multi-objective optimization problem by using a multi-objective grey wolf optimization algorithm and based on a preset surrogate model, to output the micro process parameter vector.

4. The spring wire cutting command automatic control system based on industrial data analysis of claim 1, wherein, Further comprising: a feedback calibration module configured to calculate a long-term effective output index based on the physical credibility feature vector and the production tact feature vector; and periodically calibrate weights used for constructing the reward function according to the long-term effective output index. ​

Images