A method for intelligent generation of industrial decision-making based on neural network model

Abstract

The present invention provides a method for intelligently generating industrial decisions based on a neural network model, belonging to the field of intelligent decision-making technology. The method comprises the following steps: constructing a neural network model; extracting the decision differences of each first decision sample based on the neural network model, and adding neurons to each hidden layer in the neural network model; determining the optimal variable of each first decision sample in a first industrial scenario, and optimizing the corresponding first decision sample in combination with the historical optimal probability to obtain a second decision sample; obtaining a third decision sample in a second industrial scenario related to the industrial task, and training the neural network model with the added neurons in combination with the second decision sample to obtain an industrial decision model; obtaining industrial anomalies of the new industrial task and inputting them into the industrial decision model to automatically generate intelligent decision solutions. The method can quickly and accurately generate intelligent decision solutions for various production problems, reducing decision errors.

Description

Technical Field

[0001] The present invention relates to the field of intelligent decision-making technology, and in particular to an intelligent industrial decision-making generation method based on a neural network model. Background Art

[0002] In modern industrial production systems, efficient and accurate industrial decision-making plays a critical role in ensuring stable equipment operation, optimizing production processes, and ensuring product quality meets standards. Traditional industrial decision-making methods rely heavily on manual judgment and simple statistical analysis. For example, operators set equipment operating parameters based on past production experience and execute production processes according to fixed process standards. This approach is not only inefficient but also highly susceptible to human factors, which can easily lead to problems such as delayed equipment fault diagnosis, lagging production processes, and unstable product quality, resulting in low production efficiency.

[0003] With the development of artificial intelligence (AI) technology, neural networks, owing to their powerful nonlinear fitting and data processing capabilities, are increasingly being applied to industrial decision-making. By constructing neural network models and training them with historical data, some decision-making tasks in industrial scenarios can be automated, improving decision-making efficiency and accuracy to a certain extent. However, existing neural network-based industrial decision-making solutions still have many limitations. First, traditional neural network models rely on a limited number of samples during training, resulting in insufficient accuracy in determining equipment operation, production processes, and quality inspections in industrial scenarios, making them difficult to adapt to complex and changing industrial environments. Second, existing models are relatively rigid in structure. As industrial scenarios change or new decision-making requirements arise, the models' generalization ability is poor, making them unable to adjust in a timely manner to provide reliable decision results.

[0004] Therefore, the present invention proposes an intelligent generation method for industrial decision-making based on a neural network model. Summary of the Invention

[0005] The present invention provides an intelligent generation method for industrial decision-making based on a neural network model to solve the above-mentioned technical problems.

[0006] The present invention proposes a method for intelligently generating industrial decisions based on a neural network model, comprising:

[0007] Step 1: Constructing a neural network model, wherein the neural network model is trained based on a first decision sample in a first industrial scenario related to an industrial task, and is used for determining equipment operation, production process, and quality inspection in the first industrial scenario;

[0008] Step 2: extracting the decision difference of each first decision sample based on the neural network model, and adding neurons to each hidden layer in the neural network model;

[0009] Step 3: Determine the optimal variable for each first decision sample in the first industrial scenario, and optimize the corresponding first decision sample based on the historical optimal probability to obtain the second decision sample;

[0010] Step 4: Obtain a third decision sample in a second industrial scenario related to the industrial task and train the neural network model with the added neurons in combination with the second decision sample to obtain an industrial decision model;

[0011] Step 5: Obtain industrial anomalies of new industrial tasks and input them into the industrial decision-making model to automatically generate intelligent decision-making solutions.

[0012] Preferably, constructing a neural network model includes:

[0013] Obtaining, from the first decision sample, a first abnormality for equipment operation and a first decision, a second abnormality for the production process and a second decision, and a third abnormality for quality inspection and a third decision, and constructing an input-output vector, wherein the input-output vector is: an actual input vector consisting of the first abnormality, the second abnormality, and the third abnormality, and an actual output vector consisting of the first decision, the second decision, and the third decision;

[0014] All input-output vectors in the first industrial scenario are trained to obtain a neural network model.

[0015] Preferably, adding neurons to each hidden layer in the neural network model includes:

[0016] Extracting an actual input vector from the first decision sample and inputting it into the neural network model to obtain a predicted output vector, comparing the predicted output vector with the actual output vector in the input-output vector to construct a decision difference;

[0017] According to each decision difference, the number of neurons to be added in the corresponding hidden layer is determined to obtain the additional arrays of all hidden layers, and the additional matrix is ​​constructed to obtain the additional eigenvectors and the variance of each hidden layer;

[0018] The characteristic coefficients and variances of each hidden layer in the additional characteristic vector are extracted, and the average number of neurons in each hidden layer is extracted based on the additional matrix. The additional number of neurons in the corresponding hidden layer is matched from the average-coefficient-variance-number comparison table, and the corresponding hidden layer is supplemented.

[0019] Preferably, determining the number of neurons to be supplemented in each hidden layer according to each decision difference includes:

[0020] Determining a position for adding neurons in the corresponding hidden layer based on the prediction path of each difference decision in the corresponding hidden layer and the prediction error of each neuron, wherein the position for adding neurons is a position before the first neuron in the prediction path of the corresponding hidden layer or a position after the last neuron in the prediction path of the corresponding hidden layer;

[0021] At the same time, the number of neurons to be supplemented in the corresponding hidden layer is determined according to the prediction error of each neuron;

[0022]

[0023] in, Indicates the number of hidden layers to be supplemented in the jth layer; represents the prediction error of the i1th neuron in the prediction path of the i2th first decision sample based on the jth hidden layer; represents the maximum error of the sum of the prediction errors of all neurons in the prediction path under the jth hidden layer for the m1 first decision samples; n represents the total number of neurons in the prediction path under the jth hidden layer; represents the empirical coefficient, which is 1.2; m1 represents the number of first decision samples; Indicates the rounding symbol.

[0024] Preferably, when the additional position is before the first neuron, it is fully connected with the neurons in the previous layer; when the additional position is after the last neuron, it is probabilistically connected with the neurons in the next layer;

[0025] Among them, the activation probability of the probabilistic connection is dynamically adjusted by the following formula:

[0026]

[0027] in, is the activation probability; represents the average activation value of all neurons in the prediction path of the jth hidden layer; represents the maximum activation value of all neurons in the prediction path of the jth hidden layer; The basic offset is 0.2.

[0028] Preferably, determining the optimal variable for each first decision sample in the first industrial scenario includes:

[0029] Extracting the independent variable corresponding to each anomaly under the first decision sample and the dependent variable of each anomaly based on each independent variable under the corresponding anomaly, and establishing an active influence set for each independent variable, wherein the active influence set includes a number of dependent variables that influence the corresponding independent variable, and the influence coefficient is determined based on the nonlinear relationship between the independent variable and the dependent variable;

[0030] The active influence set is input into the set analysis model to set importance labels for the independent variables under the corresponding anomalies, and a first important variable group for equipment operation, a second important variable group for production process, and a third important variable group for quality inspection are obtained, wherein the first important variable group, the second important variable group, and the third important variable group are used as the optimal variables for the first industrial scenario.

[0031] Preferably, before optimizing the first decision sample in combination with the historical optimal probability, the method includes:

[0032] Obtaining the actual participation effect of each independent variable in each first decision sample, screening first variables whose actual participation effect is greater than the preset participation effect, and if all first variables belong to the optimal variables, then determining that the first decision sample is optimal;

[0033] Count the optimal first quantity among all first decision samples, and combine the total number of samples of all first decision samples to obtain the optimal probability Yp;

[0034] According to rand(0,1-Yp), random probability screening is performed on the first decision samples that do not reach the optimal level;

[0035] When the screening result is 0, it is determined that there is no need to optimize the first decision sample that does not reach the optimal state;

[0036] Otherwise, it is determined that the first decision sample that does not reach the optimum needs to be optimized.

[0037] Preferably, after determining that the first decision sample that does not reach the optimum needs to be optimized, the method further includes:

[0038] Construct the difference set of participation effects of the first decision-making model that does not reach the optimal ,in, It represents the difference between the actual participation effect and the preset participation effect of the h1th variable; En represents the total number of variables that do not reach the optimal first decision sample;

[0039] Calculate the similarity between the first decision sample that does not reach the optimal state and each first decision sample that reaches the optimal state respectively, and use the clustering algorithm to perform cluster analysis on the first decision sample that reaches the optimal state according to the similarity to extract the clustering result under the largest cluster head;

[0040] Determine the variables in the participation effect difference set that are less than 0 and regard them as the second variables, and perform heterogeneous variable statistics on the second variables and each optimal first decision sample in the clustering result in turn to obtain a third quantity;

[0041] The second variables in the first decision samples that reach the optimum and correspond to the largest number locked from the third number are sequentially replaced with the second variables in the corresponding first decision samples that do not reach the optimum to obtain the second decision samples.

[0042] Compared with the prior art, the present invention has the following advantages:

[0043] By gradually optimizing the neural network model, the industrial decision-making model can effectively learn the decision-making experience of different production links. From initial model construction to multi-scenario data fusion training, the model continuously improves its ability to identify and handle industrial anomalies, and can quickly and accurately generate intelligent decision-making solutions for various production problems, which helps the intelligent management of industrial decision-making, reduces human intervention and decision-making errors, and improves production efficiency and product quality.

[0044] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description and the accompanying drawings.

[0045] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0047] Figure 1 This is a flow chart of an intelligent industrial decision-making method based on a neural network model in an embodiment of the present invention;

[0048] Figure 2 2 is a structural diagram of a neural network model in an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0050] The present invention proposes an intelligent generation method for industrial decision-making based on a neural network model, such as Figure 1 As shown, including:

[0051] Step 1: Constructing a neural network model, wherein the neural network model is trained based on a first decision sample in a first industrial scenario related to an industrial task, and is used for determining equipment operation, production process, and quality inspection in the first industrial scenario;

[0052] Step 2: extracting the decision difference of each first decision sample based on the neural network model, and adding neurons to each hidden layer in the neural network model;

[0053] Step 3: Determine the optimal variable for each first decision sample in the first industrial scenario, and optimize the corresponding first decision sample based on the historical optimal probability to obtain the second decision sample;

[0054] Step 4: Obtain a third decision sample in a second industrial scenario related to the industrial task and train the neural network model with the added neurons in combination with the second decision sample to obtain an industrial decision model;

[0055] Step 5: Obtain industrial anomalies of new industrial tasks and input them into the industrial decision-making model to automatically generate intelligent decision-making solutions.

[0056] In this embodiment, the new industrial task refers to the industrial task executed at the current time.

[0057] In this embodiment, the first industrial scenario is, for example, a stamping shop in automobile production, which is responsible for stamping steel plates into various parts required for the car body, such as doors, hoods, etc.

[0058] For example, the first decision sample involves: when stamping a batch of car doors, the equipment operating parameters such as the pressure, stamping speed, and mold temperature of the stamping machine, the production process parameters such as the material and thickness of the steel plate, and the quality inspection data such as the dimensional accuracy and surface flatness of the car doors after stamping, which facilitates the first decision sample to construct an input-output vector for training.

[0059] Equipment operation judgment is to determine whether the equipment operating parameters meet the set standards. For example, whether the pressure of the stamping machine is stable. If the pressure fluctuates too much, it may affect the molding quality of the car door and the pressure parameters need to be adjusted.

[0060] Production process determination is a production process flow executed by industrial tasks. For example, it determines whether the current mold temperature is appropriate. Too high or too low a temperature may cause steel plate deformation or mold wear, and the temperature control strategy needs to be adjusted.

[0061] Quality inspection and judgment is to determine whether the quality of the execution results of industrial tasks is qualified. For example, based on data such as the size deviation of the car door and whether there are scratches on the surface, it is determined whether the car door meets the production standards.

[0062] The decision difference is the difference between the actual output vector and the predicted output vector of the neural network model. The optimal variable is the variable that has always been in a stable state under the equipment operation judgment, production process judgment and quality inspection judgment.

[0063] The second decision sample is obtained based on the first decision sample after adjustment in combination with the optimal variables and historical optimal probabilities. For example, the stamping machine pressure and stamping speed in the original decision sample are adjusted to the optimal variable values ​​to form a new second decision sample.

[0064] The second industrial scenario is different from the first industrial scenario. For example, the welding workshop in automobile production is responsible for welding stamped body parts into complete body frames.

[0065] For example, the third decision sample is related to the equipment operating parameters of the welding robot such as welding current, welding speed, welding point spacing, production process parameters such as welding material type and specifications, and quality inspection data such as strength and sealing of the body frame after welding when welding the body frame.

[0066] The beneficial effects of the above technical solution are: by gradually optimizing the neural network model, the industrial decision-making model can effectively learn the decision-making experience of different production links. From the initial model construction to the multi-scenario data fusion training, the model continuously improves the ability to identify and handle industrial anomalies, and can quickly and accurately generate intelligent decision-making solutions for various production problems, which helps the intelligent management of industrial decision-making, reduces human intervention and decision-making errors, and improves production efficiency and product quality.

[0067] The present invention proposes an intelligent generation method for industrial decision-making based on a neural network model, which constructs a neural network model, including:

[0068] Obtaining, from the first decision sample, a first abnormality for equipment operation and a first decision, a second abnormality for the production process and a second decision, and a third abnormality for quality inspection and a third decision, and constructing an input-output vector, wherein the input-output vector is: an actual input vector consisting of the first abnormality, the second abnormality, and the third abnormality, and an actual output vector consisting of the first decision, the second decision, and the third decision;

[0069] All input-output vectors in the first industrial scenario are trained to obtain a neural network model.

[0070] In this embodiment, the first abnormality is, for example, in an automobile engine production workshop, when a CNC machine tool is processing a piston, the vibration frequency of the equipment exceeds the normal range. This is the first abnormality in the operation of the equipment.

[0071] First decision: For example, when an abnormal vibration frequency of a CNC machine tool is detected, the technician decides to reduce the machine speed and check the tool wear. This decision is the first decision.

[0072] The second abnormality is, for example, in the production process, during the piston processing, the cutting temperature continues to be higher than the standard temperature range required by the process, which is the second abnormality.

[0073] For example, in the case of excessively high cutting temperature, the technician decides to increase the coolant flow and adjust the cutting parameters. This is the second decision.

[0074] The third abnormality is that during quality inspection, it is found that the diameter deviation of the produced piston exceeds the allowable tolerance range. This is the third abnormality.

[0075] The third decision, for example, for the problem of unqualified piston size, the decision is to rework the unqualified piston and readjust the machine tool processing parameters. This is the third decision.

[0076] The actual input vector is a vector composed of the first abnormality (abnormal vibration frequency data of CNC machine tools), the second abnormality (excessive cutting temperature data), and the third abnormality (piston diameter size deviation data) in a certain order, for example, [vibration frequency value, cutting temperature value, piston diameter deviation value].

[0077] The actual output vector is a vector formed by the combination of the first decision (decision code for reducing the rotation speed and checking the tool), the second decision (decision code for increasing the coolant flow and adjusting the cutting parameters), and the third decision (decision code for rework and adjusting the machine tool parameters), such as [decision code 1, decision code 2, decision code 3].

[0078] In this embodiment, all input-output vectors are obtained by combining the input-output vectors of each first decision sample, and a model is built using a deep learning framework (such as TensorFlow or PyTorch). All input-output vectors are divided into a training set, a validation set, and a test set. The training set is used to train the model, the validation set is used to adjust model hyperparameters, and the test set is used to evaluate model performance. Using a backpropagation algorithm and an optimizer (such as stochastic gradient descent), the actual input vectors are fed into the model. Based on the difference between the model output and the actual output vectors, the neural network parameters are continuously adjusted. Training is repeated until the model achieves satisfactory performance on the validation set, ultimately resulting in a neural network model suitable for the first industrial scenario.

[0079] The beneficial effect of this technical solution is that it accurately extracts industrial anomalies and corresponding decisions from the first decision sample, and constructs input-output vectors to train a neural network model, enabling the model to learn the mapping relationship between anomalies and decisions in industrial scenarios. In actual production, the trained neural network model can quickly predict reasonable decision plans based on real-time industrial anomalies detected, assisting industrial production decision-making, improving the automation and intelligence level of the production process, and promptly addressing various production issues to ensure efficient and stable production.

[0080] The present invention proposes an intelligent industrial decision-making generation method based on a neural network model, which adds neurons to each hidden layer in the neural network model, including:

[0081] Extracting an actual input vector from the first decision sample and inputting it into the neural network model to obtain a predicted output vector, comparing the predicted output vector with the actual output vector in the input-output vector to construct a decision difference;

[0082] According to each decision difference, the number of neurons to be added in the corresponding hidden layer is determined to obtain the additional arrays of all hidden layers, and the additional matrix is ​​constructed to obtain the additional eigenvectors and the variance of each hidden layer;

[0083] The characteristic coefficients and variances of each hidden layer in the additional characteristic vector are extracted, and the average number of neurons in each hidden layer is extracted based on the additional matrix. The additional number of neurons in the corresponding hidden layer is matched from the average-coefficient-variance-number comparison table, and the corresponding hidden layer is supplemented.

[0084] In this embodiment, the actual input vector for the first decision sample is: {first anomaly, second anomaly, third anomaly}, the predicted output vector is: {prediction decision for the first anomaly, prediction decision for the second anomaly, prediction decision for the third anomaly}, and the actual output vector is: {actual decision for the first anomaly, actual decision for the second anomaly, actual decision for the third anomaly};

[0085] At this time, the decision difference is: {the difference between the actual decision and the predicted decision for the first anomaly, the difference between the actual decision and the predicted decision for the second anomaly, and the difference between the actual decision and the predicted decision for the third anomaly}.

[0086] In this embodiment, the number of neurons that need to be supplemented in each hidden layer can be determined by analyzing the decision differences.

[0087] In this embodiment, Figure 2As shown in the figure, the neural network model consists of an input layer (the number of nodes is related to the number of features of the industrial task, for example, 20 nodes are set), 3 hidden layers with 30 neurons in each layer, and an output layer (outputting the decision results).

[0088] In this embodiment, since there are three hidden layers, the additional array = {the data to be supplemented in hidden layer 1, the number of hidden layer 2 to be supplemented, the number of hidden layer 3 to be supplemented}, and the obtained additional matrix = .

[0089] In this embodiment, the variance of each hidden layer is the variance of the column vector in the corresponding matrix, reflecting the discrete degree of the neuron output.

[0090] In this embodiment, the additional feature vector is a vector obtained by performing feature extraction on the additional matrix, for example, [b1, b2, b3] represents the feature enhancement coefficient of each hidden layer.

[0091] In this embodiment, the average number is the average number of neurons to be supplemented involved in each column vector in the corresponding additional matrix, that is, the average number of neurons added in the corresponding hidden layer.

[0092] In this embodiment, the average-coefficient-variance-quantity comparison table contains different quantities to be supplemented, characteristic coefficients, variances, and the additional quantities that match the combinations thereof, ensuring the reliability of the neuron settings of the hidden layer. For example, the characteristic coefficient of hidden layer 1 is: 0.78, the variance is 0.3, and the average quantity is 3. At this time, the additional quantity obtained by matching according to the comparison table is 4.

[0093] The beneficial effect of the above technical solution is: supplementing hidden layer neurons according to the difference between the actual output and the preset output of the neural network model can dynamically optimize the model structure and improve decision-making accuracy.

[0094] The present invention proposes an intelligent industrial decision-making generation method based on a neural network model, which determines the number of neurons to be supplemented in each hidden layer according to each decision difference, including:

[0095] Determining a position for adding neurons in the corresponding hidden layer based on the prediction path of each difference decision in the corresponding hidden layer and the prediction error of each neuron, wherein the position for adding neurons is a position before the first neuron in the prediction path of the corresponding hidden layer or a position after the last neuron in the prediction path of the corresponding hidden layer;

[0096] At the same time, the number of neurons to be supplemented in the corresponding hidden layer is determined according to the prediction error of each neuron;

[0097]

[0098] in, Indicates the number of hidden layers to be supplemented in the jth layer; represents the prediction error of the i1th neuron in the prediction path of the i2th first decision sample based on the jth hidden layer; represents the maximum error of the sum of the prediction errors of all neurons in the prediction path under the jth hidden layer for the m1 first decision samples; n represents the total number of neurons in the prediction path under the jth hidden layer; represents the empirical coefficient, which is 1.2; m1 represents the number of first decision samples; Indicates the rounding symbol.

[0099] In this embodiment, the value of j is 3.

[0100] In this embodiment, the number of hidden layer neurons reflects the current structure scale, and the logarithm of the number of training set samples reflects the demand of data scale for model complexity. The combination of the three can scientifically calculate the appropriate number of supplementary neurons under different data scales and model structures, avoiding overfitting or underfitting caused by blindly adding neurons. The maximum error value represents the maximum deviation between the model's prediction and actual conditions. Using this value as the denominator makes the calculated dynamic redundancy factor more sensitive to such extreme deviations. In industrial decision-making scenarios, such as equipment fault diagnosis, misjudgment can have serious consequences. Incorporating the maximum error value into the calculation allows the model to prioritize these extreme errors, allowing for timely adjustments to the number of neurons, enhancing its ability to learn from complex or unique situations, and improving the model's robustness.

[0101] In this embodiment, information transmission is carried out along a specific neuron connection as a prediction path. Assume that in the jth hidden layer, information is transmitted from neuron A → neuron B → neuron C → neuron D and participates in the calculation. This transmission path is the prediction path of the corresponding hidden layer. Each neuron undertakes different feature calculation tasks in the path. For example, neuron A is responsible for the preliminary integration of the correlation characteristics of welding current and time, and neuron B is responsible for strengthening the impact of this correlation on welding strength.

[0102] In this embodiment, each neuron in the prediction path will generate a prediction error. For example, after neuron B calculates based on the input information, it expects the output to enable subsequent neurons to more accurately predict the welding quality, but the actual output deviates from the ideal output. Assuming that the welding strength characteristic contribution corresponding to the ideal output value of neuron B is 0.8 (the value after normalization and other processing), and the actual output corresponding contribution is 0.6, the difference of 0.2 between the two is the prediction error of this neuron (the i1th neuron, i1 corresponds to the sequence number of neuron B in the path). .

[0103] In this embodiment, taking the prediction path of the j-th hidden layer, neuron A→neuron B→neuron C→neuron D as an example, either a new neuron is added before neuron A to pre-process the information such as welding process parameters input to the hidden layer earlier; or a new neuron is added after neuron D to further optimize the information processed by the hidden layer before transmitting it out. The newly added position is the location of the additional neuron.

[0104] In this example, the cumulative error contribution A1 of all prediction errors under the first neuron is determined, and the cumulative error contribution A2 of all prediction errors under the last neuron is determined. If A1 is greater than A2, it means that the first neuron has a more significant "front-end influence" on the error; otherwise, the last neuron has a more prominent "back-end influence." In this case, a new neuron is added before the first neuron; otherwise, a new neuron is added after the last neuron.

[0105] The beneficial effects of the above technical solution are: by accurately analyzing the errors in the hidden layer prediction path of differential decisions, scientifically calculating and determining the location and number of additional hidden layer neurons, it can effectively optimize the neural network structure, allowing the model to better learn the complex relationship between industrial decisions, improve prediction accuracy, and reduce the deviation between actual production and model predictions.

[0106] The present invention proposes an intelligent generation method for industrial decision-making based on a neural network model. When the additional position is before the first neuron, it is fully connected with the neurons in the previous layer; when the additional position is after the last neuron, it is probabilistically connected with the neurons in the next layer.

[0107] Among them, the activation probability of the probabilistic connection is dynamically adjusted by the following formula:

[0108]

[0109] in, is the activation probability; represents the average activation value of all neurons in the prediction path of the jth hidden layer; represents the maximum activation value of all neurons in the prediction path of the jth hidden layer; The basic offset is 0.2.

[0110] In this embodiment, due to The value range is from 0 to 1. The range of is changed to 0 to 0.8, and 0.2 is added to make the connection probability P range between 0.2 and 1. This avoids the connection probability being too low (close to 0) resulting in almost no connection between the new neuron and the next layer, and also prevents the probability being too high (close to 1) causing over-connection, ensuring that the new neuron has a moderate connection relationship with the neurons in the next layer. 0.2 is used as a basic offset to ensure that even when the average activation value is 0, there is a 0.2 connection probability between the new neuron and the neurons in the next layer, so as not to completely disconnect and maintain the basic connectivity of the network structure.

[0111] The beneficial effects of the above technical solution are: the differentiated design of full connection and probabilistic connection enables the neural network to not only grasp the basic process characteristics, but also flexibly respond to complex decisions, improve the decision-making accuracy of process optimization and quality inspection, and dynamic probability adjustment ensures that the network can still stably output decisions during production line fluctuations and process switching, reducing production stoppages and quality accidents caused by model failure.

[0112] The present invention proposes an intelligent industrial decision-making generation method based on a neural network model, which determines the optimal variables of each first decision sample in a first industrial scenario, including:

[0113] Extracting the independent variable corresponding to each anomaly under the first decision sample and the dependent variable of each anomaly based on each independent variable under the corresponding anomaly, and establishing an active influence set for each independent variable, wherein the active influence set includes a number of dependent variables that influence the corresponding independent variable, and the influence coefficient is determined based on the nonlinear relationship between the independent variable and the dependent variable;

[0114] The active influence set is input into the set analysis model to set importance labels for the independent variables under the corresponding anomalies, and a first important variable group for equipment operation, a second important variable group for production process, and a third important variable group for quality inspection are obtained, wherein the first important variable group, the second important variable group, and the third important variable group are used as the optimal variables for the first industrial scenario.

[0115] In this embodiment, the variables are, for example, parameters that can be directly measured and controlled, such as torque value, tightening speed, and temperature during the bolt tightening process in an automobile assembly workshop.

[0116] The dependent variable is the result directly or indirectly caused by the independent variable, such as the bolt's preload, fatigue life, loosening probability and other quality indicators.

[0117] Active influence set For example, taking the torque value as an example, its active influence set may include {preload, fatigue life}, indicating that these two dependent variables are directly affected by the torque value.

[0118] The influence coefficient is an indicator that measures the strength of the nonlinear relationship between the independent variable and the dependent variable. It uses algorithms such as random forest and gradient boosting tree to establish a nonlinear mapping relationship between the independent variable and the dependent variable. For example, for the torque value T and the preload force F, the following is obtained through fitting: , when T=50N⋅m, the influence coefficient is about 0.78, which means that for every 1% change in torque value, the preload force changes by about 0.78%.

[0119] In this embodiment, the set analysis model is a feature evaluation model based on Shapley value or Gini importance, which is used to calculate the contribution of each independent variable to the decision result.

[0120] The importance label is the importance score assigned to each independent variable. For example, the importance label of the torque value is 0.85, which indicates the degree of its influence on the decision outcome.

[0121] The first important variable group is the key variables for equipment operation, such as the motor current and speed fluctuation of the tightening gun. The second important variable group is the key variables for the production process, such as the hardness of the bolt material and the lubrication status. The third important variable group is the key variables for quality inspection, such as the preload fluctuation range and torque attenuation rate.

[0122] Variable grouping is based on the business attributes of the variables and is divided into three groups: equipment operation, production process, and quality inspection. For example, the equipment operation group includes {motor current, speed fluctuation, battery voltage}, the production process group includes {bolt material hardness, lubrication status, surface roughness}, and the quality inspection group includes {preload fluctuation range, torque attenuation rate, axial force deviation}.

[0123] Threshold screening involves setting an importance threshold (e.g., 0.5) and retaining variables that exceed the threshold as optimal variables. For example, the optimal variable group ultimately determined is: First Important Variable Group: {Motor Current, Speed ​​Fluctuation}, Second Important Variable Group: {Bolt Material Hardness, Lubrication Status}, Third Important Variable Group: {Preload Fluctuation Range, Torque Attenuation Rate}.

[0124] The beneficial effects of the above technical solution are: the relationship between variables in industrial processes is usually nonlinear, which can accurately describe the differences in the impact of torque on preload in different intervals, avoid the limitations of linear models, focus on key variables, optimize sensor deployment and data collection frequency, clarify the importance level of each variable, and provide a quantitative basis for process improvement.

[0125] The present invention proposes an intelligent industrial decision-making generation method based on a neural network model. Before optimizing the first decision sample in combination with the historical optimal probability, the method includes:

[0126] Obtaining the actual participation effect of each independent variable in each first decision sample, screening first variables whose actual participation effect is greater than the preset participation effect, and if all first variables belong to the optimal variables, then determining that the first decision sample is optimal;

[0127] Count the optimal first quantity among all first decision samples, and combine the total number of samples of all first decision samples to obtain the optimal probability Yp;

[0128] According to rand(0,1-Yp), random probability screening is performed on the first decision samples that do not reach the optimal level;

[0129] When the screening result is 0, it is determined that there is no need to optimize the first decision sample that does not reach the optimal state;

[0130] Otherwise, it is determined that the first decision sample that does not reach the optimum needs to be optimized.

[0131] In this embodiment, the actual participation effect is a measurement value of the actual effect of the corresponding independent variable on the quality, and the preset participation effect is a threshold value for the corresponding independent variable pre-set based on historical experience and process standards. For example, the actual effect measurement value of the independent variable k1 is 0.7, and the corresponding threshold value is 0.6. At this time, 0.7 is greater than 0.6, and the corresponding independent variable k1 is regarded as the first variable.

[0132] For example, the optimal variables are: {k1, k2, k3, k8, k9}. At this time, the first variables for the first decision sample are k1 and k2. At this time, the first variables k1 and k2 are optimal variables, and the corresponding first decision sample is regarded as reaching the optimal state. If the first variables for the first decision sample are k1 and k4, and k4 does not belong to the optimal variable, it is regarded as the first decision sample that has not reached the optimal state.

[0133] In this embodiment, the optimal probability Yp=the first quantity / the total number of samples. Based on this probability, the overall quality level of the current industrial production can be quickly understood, providing a basic basis for subsequent decision-making, which is consistent with the basic idea of ​​frequency estimation probability in statistics.

[0134] In this embodiment, a random probability screening mechanism is introduced to determine whether to optimize suboptimal samples, rather than optimizing all suboptimal samples. This is primarily due to two considerations. First, it avoids over-optimization. In industrial production, some suboptimal samples may be caused by accidental factors (such as a brief equipment fluctuation or minor differences in raw materials), rather than fundamental problems with the process or variable settings. Blind optimization can increase production costs and complexity. Second, by setting a random range (0, 1 − Yp), the optimization probability of suboptimal samples is linked to the optimal probability. The higher the optimal probability (the more stable and high-quality the production process), the smaller the probability range for suboptimal samples to be optimized. This meets the practical requirement of "prioritizing stable production while moderately optimizing outliers" and is an effective means of balancing production stability and optimization needs.

[0135] The beneficial effects of the above technical solution are: by comparing the first variable and the optimal variable, samples whose processing meets the optimal process requirements and whose product quality is high can be accurately screened out; the calculation of the optimal probability Yp reflects the stability and high-quality output capacity of the current production system from a macro perspective; the random probability screening mechanism avoids the "one-size-fits-all" optimization of substandard samples, and can target samples that may indeed have room for optimization; through scientific sample analysis, probability calculation and screening mechanisms, it provides a set of intelligent and efficient decision-making optimization methods for industrial production processes.

[0136] The present invention proposes an intelligent industrial decision-making generation method based on a neural network model. After determining that a first decision sample that does not reach the optimal level needs to be optimized, the method further includes:

[0137] Construct the difference set of participation effects of the first decision-making model that does not reach the optimal ,in, It represents the difference between the actual participation effect and the preset participation effect of the h1th variable; En represents the total number of variables that do not reach the optimal first decision sample;

[0138] Calculate the similarity between the first decision sample that does not reach the optimal state and each first decision sample that reaches the optimal state respectively, and use the clustering algorithm to perform cluster analysis on the first decision sample that reaches the optimal state according to the similarity to extract the clustering result under the largest cluster head;

[0139] Determine the variables that are less than 0 in the participation effect difference set and regard them as the second variables, and perform heterogeneous variable statistics on the second variables and each optimal first decision sample in the clustering result in turn to obtain a third quantity;

[0140] The second variables in the first decision samples that reach the optimum and correspond to the largest number locked from the third number are sequentially replaced with the second variables in the corresponding first decision samples that do not reach the optimum to obtain the second decision samples.

[0141] In this embodiment, the difference effect conversion value is used to standardize and normalize the corresponding differences, eliminate the dimension problem, and limit the result to between -1 and 1, and the variables whose existence values ​​are less than 0 after the corresponding participation effect differences are centralized standardized and normalized are regarded as the second variables.

[0142] In this embodiment, the similarity is calculated by using the Euclidean distance calculation formula to calculate the similarity between the first decision sample that does not reach the optimal state and each first decision sample that reaches the optimal state, that is: , where N0 represents the total number of variables included in the corresponding decision sample; represents the value of the j0th variable in the first decision sample that does not reach the optimal value; represents the value of the j0th variable in the first decision sample that achieves the optimal result. Euclidean distance effectively measures the differences in parameter vectors, meeting the need to "find similar, high-quality cases" in process optimization.

[0143] In this embodiment, the clustering algorithm is implemented using the K-means algorithm, and the group containing the largest number of samples is extracted as the largest cluster head, which is used as a basis for subsequent analysis to avoid the randomness of a single sample.

[0144] For example, if the value of the second variable p1 is less than 0, and the value of the variable p1 corresponding to the optimal first decision sample L1 in the clustering result is greater than 0, then the variable p1 is statistically added by one, otherwise, the variable p1 is not statistically added by one, and the third quantity can be obtained.

[0145] In this embodiment, if the maximum number is 1, then the second variable in the corresponding first decision sample that has reached the optimum is sequentially replaced with the second variable in the corresponding first decision sample that has not reached the optimum;

[0146] If there are multiple maximum numbers, at this time, a corresponding second variable in the optimal first decision sample is randomly selected from the samples under the maximum number to replace the second variable in the corresponding first decision sample that does not reach the optimal state in turn. For example, the second variable p1 in the optimal first decision sample corresponding to the maximum number obtained is replaced by the second variable p2 in the first decision sample that does not reach the optimal state.

[0147] The beneficial effect of the above technical solution is: through the closed-loop logic of deviation quantification, similarity clustering, direction screening and variable replacement, industrial experience is converted into a computable and reusable optimization strategy.

[0148] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A method for intelligently generating industrial decisions based on a neural network model, characterized in that: include: Step 1: Constructing a neural network model, wherein the neural network model is trained based on a first decision sample in a first industrial scenario related to an industrial task, and is used for determining equipment operation, production process, and quality inspection in the first industrial scenario; Step 2: extracting the decision difference of each first decision sample based on the neural network model, and adding neurons to each hidden layer in the neural network model; Step 3: Determine the optimal variable for each first decision sample in the first industrial scenario, and optimize the corresponding first decision sample based on the historical optimal probability to obtain the second decision sample; Step 4: Obtain a third decision sample in a second industrial scenario related to the industrial task and train the neural network model with the added neurons in combination with the second decision sample to obtain an industrial decision model; Step 5: Obtain industrial anomalies of new industrial tasks and input them into the industrial decision-making model to automatically generate intelligent decision-making solutions; The process of adding neurons to each hidden layer of the neural network model includes: Extracting an actual input vector from the first decision sample and inputting it into the neural network model to obtain a predicted output vector, comparing the predicted output vector with the actual output vector in the input-output vector to construct a decision difference; According to each decision difference, the number of neurons to be added in the corresponding hidden layer is determined to obtain the additional arrays of all hidden layers, and the additional matrix is ​​constructed to obtain the additional eigenvectors and the variance of each hidden layer; Extracting the characteristic coefficients and variances of each hidden layer in the additional characteristic vector, and extracting the average number of neurons in each hidden layer based on the additional matrix, matching the additional number of neurons in the corresponding hidden layer from the average-coefficient-variance-number comparison table, and supplementing the corresponding hidden layer; The number of neurons to be supplemented in each hidden layer is determined according to each decision difference, including: Determining a position for adding neurons in the corresponding hidden layer based on the prediction path of each difference decision in the corresponding hidden layer and the prediction error of each neuron, wherein the position for adding neurons is a position before the first neuron in the prediction path of the corresponding hidden layer or a position after the last neuron in the prediction path of the corresponding hidden layer; At the same time, the number of neurons to be supplemented in the corresponding hidden layer is determined according to the prediction error of each neuron; ; in, Indicates the number of hidden layers to be supplemented in the jth layer; represents the prediction error of the i1th neuron in the prediction path of the i2th first decision sample based on the jth hidden layer; represents the maximum error of the sum of the prediction errors of all neurons in the prediction path under the jth hidden layer for the m1 first decision samples; n represents the total number of neurons in the prediction path under the jth hidden layer; represents the empirical coefficient, which is 1.2; m1 represents the number of first decision samples; Indicates the rounding symbol.

2. The method for intelligently generating industrial decisions based on a neural network model according to claim 1, characterized in that: Build a neural network model, including: Obtaining, from the first decision sample, a first abnormality for equipment operation and a first decision, a second abnormality for the production process and a second decision, and a third abnormality for quality inspection and a third decision, and constructing an input-output vector, wherein the input-output vector is: an actual input vector consisting of the first abnormality, the second abnormality, and the third abnormality, and an actual output vector consisting of the first decision, the second decision, and the third decision; All input-output vectors in the first industrial scenario are trained to obtain a neural network model.

3. The method for intelligently generating industrial decisions based on a neural network model according to claim 1, characterized in that: When the additional position is before the first neuron, it is fully connected with the neurons in the previous layer; when the additional position is after the last neuron, it is probabilistically connected with the neurons in the next layer. Among them, the activation probability of the probabilistic connection is dynamically adjusted by the following formula: ; in, is the activation probability; represents the average activation value of all neurons in the prediction path of the jth hidden layer; represents the maximum activation value of all neurons in the prediction path of the jth hidden layer; The basic offset is 0.

2.

4. The method for intelligently generating industrial decisions based on a neural network model according to claim 1, characterized in that: Determine the optimal variables for each first decision sample in the first industrial scenario, including: Extracting the independent variable corresponding to each anomaly under the first decision sample and the dependent variable of each anomaly based on each independent variable under the corresponding anomaly, and establishing an active influence set for each independent variable, wherein the active influence set includes a number of dependent variables that influence the corresponding independent variable, and the influence coefficient is determined based on the nonlinear relationship between the independent variable and the dependent variable; The active influence set is input into the set analysis model to set importance labels for the independent variables under the corresponding anomalies, and a first important variable group for equipment operation, a second important variable group for production process, and a third important variable group for quality inspection are obtained, wherein the first important variable group, the second important variable group, and the third important variable group are used as the optimal variables for the first industrial scenario.

5. The method for intelligently generating industrial decisions based on a neural network model according to claim 1, characterized in that: Before optimizing the first decision sample in combination with the historical optimal probability, the following steps are included: Obtaining the actual participation effect of each independent variable in each first decision sample, screening first variables whose actual participation effect is greater than the preset participation effect, and if all first variables belong to the optimal variables, then determining that the first decision sample is optimal; Count the optimal first quantity among all first decision samples, and combine the total number of samples of all first decision samples to obtain the optimal probability Yp; According to rand(0,1-Yp), random probability screening is performed on the first decision samples that do not reach the optimal level; When the screening result is 0, it is determined that there is no need to optimize the first decision sample that does not reach the optimal state; Otherwise, it is determined that the first decision sample that does not reach the optimum needs to be optimized.

6. The method for intelligently generating industrial decisions based on a neural network model according to claim 5, characterized in that: After determining that the first decision sample that does not reach the optimal state needs to be optimized, the following steps are also included: Construct the difference set of participation effects of the first decision-making model that does not reach the optimal ,in, It represents the difference between the actual participation effect and the preset participation effect of the h1th variable; En represents the total number of variables that do not reach the optimal first decision sample; Calculate the similarity between the first decision sample that does not reach the optimal state and each first decision sample that reaches the optimal state respectively, and use the clustering algorithm to perform cluster analysis on the first decision sample that reaches the optimal state according to the similarity to extract the clustering result under the largest cluster head; Determine the variables that are less than 0 in the participation effect difference set and regard them as the second variables, and perform heterogeneous variable statistics on the second variables and each optimal first decision sample in the clustering result in turn to obtain a third quantity; The second variables in the first decision samples that reach the optimum and correspond to the largest number locked from the third number are sequentially replaced with the second variables in the corresponding first decision samples that do not reach the optimum to obtain the second decision samples.

Images