CN117371561A

CN117371561A - Industrial production artificial intelligence system based on machine learning

Info

Publication number: CN117371561A
Application number: CN202311289873.4A
Authority: CN
Inventors: 向达
Original assignee: Hangzhou Asia Pacific Chemical Equipment Co ltd
Current assignee: Hangzhou Asia Pacific Chemical Equipment Co ltd
Priority date: 2023-10-08
Filing date: 2023-10-08
Publication date: 2024-01-09

Abstract

An industrial production artificial intelligence system based on machine learning comprises data training, wherein production data of preliminary screening is obtained according to distribution and classification characteristics of data; randomly dividing the machine learning model training data into a test set and a verification set, and constructing the machine learning model training data; obtaining a training result weight file after training is completed; calculating a predicted value, importing monitoring data of a certain factory, carrying out corresponding data processing, and reading trained model weights to obtain the predicted value; repeating the steps to obtain a training result; compared with the prior art, the invention combines machine learning and big data, realizes the automation to the intellectualization of industrial production, changes the current industrial production mode from a man-operated machine to a system to replace an engineer to think analysis, and provides a feasible solution for maintenance engineers to maintain the maximization of production, reduces the production cost and improves the production efficiency, and simultaneously is suitable for the production requirements of energy conservation, environmental protection, intelligence and high efficiency in all directions.

Description

Industrial production artificial intelligence system based on machine learning

Technical Field

The invention relates to the technical field of industrial production, in particular to an industrial production artificial intelligence system based on machine learning.

Background

At present, one person in the mode of industrial production controls a machine, and the machine works according to instructions sent by engineers;

the method has the advantages that the manually preset parameters in the factory control system cannot be changed along with the change of the production operation environment, so that the production is not in an optimal state, meanwhile, an alarm system of the factory can give an alarm when a problem occurs in the production process, the factory can be stopped immediately after the factory receives the alarm, and huge economic loss and huge different experience and level of engineers are brought to the factory.

For this purpose, a machine learning model is proposed to solve this problem.

Disclosure of Invention

The invention aims to provide an industrial production artificial intelligence system based on machine learning, which solves the problems that the prior system is manually preset and used inflexibly, the production is not in an optimized state, the production process has an alarm, and the emergency stop factory and the experience level of engineers are different.

In order to achieve the above purpose, the present invention provides the following technical solutions: an industrial production artificial intelligence system based on machine learning, comprising a machine learning model, comprising the following steps:

step 1: data training, namely, data received from a factory sensor are exported and stored; according to the distribution and classification characteristics of the data, preprocessing the data by using Python, eliminating noise, obtaining primarily screened production data, and storing the production data; reading the stored production data, randomly dividing the production data into a test set and a verification set, and building machine learning model training data by using a GBDT algorithm;

after training is completed, a training result weight file is obtained,

step 2: calculating predicted value, calling model at front end of system, importing monitoring data of a certain time of factory, and making correspondent data processing, using jobilib module to read trained model weight, using model weight to predict imported monitoring data so as to obtain predicted value,

step 3: along with the continuous increase of the data provided by factories, after the training data samples are also continuously increased, the training results are updated, and more data can ensure that the training results are more accurate.

Preferably, the machine learning model further comprises step 4: predictive maintenance, wherein a factory provides historical abnormal data and maintenance conditions, the abnormal data is classified according to different abnormal conditions, the abnormal data is imported into a parameter controllable interval range when each operation device normally operates, the abnormal data is randomly divided into a test set and a verification set after being integrated, and machine learning model training data is built by using a GBDT algorithm;

after training is completed, a training result weight file is obtained,

when the system is in operation, data generated on a production line is monitored and analyzed in real time, model prediction is carried out through a trained weight file, the system judges whether the current data conditions of all equipment are abnormal or not, the abnormal conditions of the production line are predicted according to the numerical values, and if the abnormal conditions are detected by the system, the system automatically gives out a solution.

Preferably, the solution includes the cause of the occurrence of the abnormal situation or the abnormal situation that may be about to occur, how to solve the problem.

Preferably, the abnormal data includes equipment condition and parameter setting.

Preferably, in step 1, the GBDT algorithm: the data shown by the process.html is calculated according to the weight model_gboost_jingbai.pkl; model weights as used herein are calculated from the GBDT regression model.

Preferably, in step 1, the specific algorithm of the GBDT algorithm is:

first initializing a first weak classifier:

(L loss function is a square error function in regression task)

Sequentially iterating for the remaining M-1 weak classifiers:

a. calculating a negative gradient of loss with respect to the last classifier for each sample

The negative gradient of the regression task is equal to the definition of the residual;

b. taking the negative gradient value obtained in the previous step as a new true value of the sample, and taking the data (x _i ，r _i，m ) I=1, 2, ·n is used as training data for the next tree to obtain a new regression tree f _m (x) The corresponding leaf node area is R _jm J=1, 2,., J, where J is the number of leaf nodes of the regression tree t;

c. calculate best fit value for leaf area j=1, 2, ·j

Preferably, the classifier needs to be updated, and the specific algorithm is as follows:

the functions of the model are directly updated, and the parameter additively popularized to the function space is utilized;

training F0-Fm total m basis learners to continuously update gamma along the gradient descent direction _jm And

preferably, in step 1, the GBDT algorithm needs to incorporate a LightGBM algorithm, which is optimized on the conventional GBDT algorithm as follows:

specifically, a decision tree algorithm based on a Histogram;

single-sided Gradient Sampling Gradient-based One-Side Sampling (GOSS): the GOSS can reduce a large number of data instances with small gradients, so that when the information gain is calculated, only the rest data with high gradients can be utilized, and compared with XGBoost, the method has the advantages that the time and space overhead are saved;

mutually exclusive feature bundle Exclusive Feature Bundling (EFB): the EFB can bind a plurality of mutually exclusive features into one feature, so that the purpose of dimension reduction is achieved;

leaf-growth strategy with depth limitation Leaf-wise: most GBDT tools use an inefficient level-wise decision tree growth strategy because it does not distinguish between leaves that treat the same level, resulting in much unnecessary overhead; in practice, the splitting gain of many leaves is low, and searching and splitting are not necessary; lightGBM uses a leaf-by-leaf growth (leaf-wise) algorithm with depth limiting;

direct support class feature (Categorical Feature)

Supporting efficient parallelism

And optimizing the Cache hit rate.

Compared with the prior art, the invention has the beneficial effects that:

1. according to the system, the optimal state data in the generation process can be screened out according to the distribution and classification characteristics of the data, the high-quality production data subjected to algorithm screening and classification can be used for training a model, the system can analyze the optimal state of the data in the learning historical data, the model can self-iterate and evolve, the system can perform uninterrupted operation, and an accurate prediction model is formed through more than ten thousands of iterations. When the whole system is running, the prediction model can be recommended to engineers for optimal regulation or autonomous regulation,

the system monitors and analyzes the data generated on the production line in real time, and once the system detects any abnormal condition, the system automatically provides a solution to engineers including the reason for generating the abnormal condition and how to solve the problem, so as to help the production time to keep the maximum production efficiency;

2. predictive maintenance: the system predicts the abnormal condition of the production line, so that engineers repair the problem in advance before the alarm system gives an alarm; when the system detects faults which possibly occur, the system prompts a user in advance, indicates that the specific expected fault position is easy to maintain by engineers, can help greatly reduce irregular production stopping times in production, reduces loss of raw materials caused by midway stopping, and greatly improves production efficiency;

3. the invention combines machine learning and big data, realizes the automation to intellectualization of industrial production, continuously absorbs the experience and method of engineers in use, shares the best experience in the same industry with the increase of factories, and meets the production requirements of energy conservation, environmental protection, intelligence and high efficiency in all directions.

Drawings

FIG. 1 is a diagram of an intelligent system framework of the present invention;

FIG. 2 is a schematic diagram of machine learning according to the present invention;

FIG. 3 is a diagram illustrating two machine learning techniques according to the present invention;

FIG. 4 is a schematic diagram of the energy consumption and optimization scheme detection of the present invention;

FIG. 5 is a schematic diagram of the results obtained by the actual energy consumption and optimization scheme of the present invention;

FIG. 6 is a schematic diagram of the invention before and after energy consumption adjustment;

FIG. 7 is a schematic diagram of fault detection according to the present invention;

FIG. 8 is a schematic diagram of a specific location of a fault in accordance with the present invention;

FIG. 9 is a schematic diagram of the performance of the system in actual generation of the present invention;

fig. 10 is a schematic diagram of a parameter adjustment suggestion display interface shown in fig. 9.

Detailed Description

The invention will now be described in more detail by way of examples which are illustrative only and are not intended to limit the scope of the invention in any way.

The invention provides a technical scheme that: an industrial production artificial intelligence system based on machine learning, comprising a machine learning model, comprising the following steps:

after training is completed, a training result weight file is obtained,

Firstly, a decision tree is learned by a test set, a predicted value and a residual error after prediction can be obtained at leaves, the later decision tree is learned based on the residual error of the previous decision tree until the residual error of the predicted value and a true value is zero, and finally, the predicted value of a test sample, namely the predicted values of a plurality of previous decision trees are accumulated;

GBDT requires traversing the entire training data multiple times each iteration; loading the entire training data into the memory limits the size of the training data; the repeated reading and writing of training data without memory can consume a very large amount of time; for massive data in a training set, the common GBDT algorithm cannot meet the requirements of the training set, so that the GBDT can better and faster run out of the results by introducing the LightGBM algorithm;

LightGBM (Light Gradient Boosting Machine) is a framework for realizing the GBDT algorithm, and the LightGBM algorithm supports efficient parallel training;

after training is completed, the data is stored into a pkl weight file; GBDT algorithm: the data shown by the process.html is calculated according to the weight model_gboost_jingbai.pkl; ,

model weights as used herein are calculated from the GBDT regression model.

The model used was GBDT regression model (GradientBoostingReggresor) and the algorithm was GBDT (Gradient Boost Decision Tree);

the specific algorithm of the GBDT algorithm is as follows:

first initializing a first weak classifier:

(L loss function is a square error function in regression task)

Sequentially iterating for the remaining M-1 weak classifiers:

The negative gradient of the regression task is equal to the definition of the residual.

it should be noted that both in classification and regression tasks, CART regression trees are utilized, so the process of generating the tree is to generate non-leaf nodes of the tree using a least squares error criterion, note non-leaf nodes, as the output values of leaf nodes are designed to be updated in another form in GBDT;

c. calculate best fit value for leaf area j=1, 2, ·j

It should be noted that the value corresponding to each leaf node is calculated by the L function, so here, the value will change due to the difference of the L loss functions of the regression task and the classification task, which is also a place to be noted by the GBDT algorithm;

the classifier needs to be updated, and the specific algorithm is as follows:

LightGBM algorithm: before the LightGBM is proposed, the GBDT tool is XGBoost, which is a decision tree algorithm based on a pre-ordering method, but has the defects of large space consumption, large time expenditure and unfriendly cache optimization;

in order to avoid the defects of XGBoost and to accelerate the training speed of GBDT model without compromising accuracy, lightGBM is optimized on the conventional GBDT algorithm as follows:

a Histogram-based decision tree algorithm;

direct support class feature (Categorical Feature)

Supporting efficient parallelism

Optimizing the hit rate of the Cache;

step 4: obtaining a predicted value: when the front end of the system needs to call a model, importing factory detection data in a CSV format, reading trained model weights through a jobilib module, and then predicting imported monitoring data by using the model weights to obtain a required predicted value.

Embodiment one:

step 1: and (3) file preservation: exporting detection data from a factory and storing the detection data into a CSV format file;

step 2: data preprocessing: carrying out data preprocessing (removing null values and screening out needed data) on the stored data by using python, and storing the data into a database after the data are processed;

step 3: training data: reading data from a database, dividing a test set and a verification set, then training the data by using a GBDT (Gradient Boost Decision Tree) algorithm, wherein each calculation of the GBDT algorithm can reduce the residual error of the last time, and the GBDT establishes a new model in the direction of reducing the residual error (negative gradient);

wherein LightGBM (Light Gradient Boosting Machine) is a framework for implementing the GBDT algorithm;

after training is completed, the data is stored into a pkl weight file;

boosting algorithm is mainly two ideas: the combination of a stack of weak classifiers can become a strong classifier; continuously learning in errors, and iterating to reduce the error making probability;

to better understand the concept of Boosting, we first introduce from weak and strong learning algorithms: 1) Strong learning algorithm: a polynomial time learning algorithm exists to identify a set of concepts, and the accuracy of the identification is high; 2) Weak learning algorithm: the accuracy of identifying a set of concepts is only slightly better than random guessing;

kearns & Valiant prove the equivalence problem of the weak learning algorithm and the strong learning algorithm, and if the two are equivalent, the learning algorithm which is slightly better than random guess is only needed to be found, so that the learning algorithm can be improved into the strong learning algorithm;

boosting algorithm optimizes classification result through a series of iterations, and introduces a weak classifier once every iteration to overcome shortcutting of existing weak classifier combination: in the Adaboost algorithm, the characterization of shortturning is a sample point with high weight, and in the Gradient Boosting algorithm, the characterization of shortturning is that whether the gradient is Adaboost or Gradient Boosting, the learner is informed of how to lift the model, namely the origin of Boosting, through shortturning;

the Adaboost algorithm is proposed by Freund and Schapire in 1997, a distribution weight vector W is maintained on the whole training set, a classification hypothesis (base learner) y (x) is generated by using the training set endowed with weight through a weak classification algorithm, then an error rate is calculated, the obtained error rate is used for updating the distribution weight vector W, a larger weight is allocated to a sample of the error classification, a smaller weight is given to a sample of the correct classification, a new classification hypothesis is generated by using the same weak classification algorithm after each updating, a multi-classifier is formed by sequences of the classification hypotheses, and the multi-classifier is combined by using a weighting method, so that a decision result is finally obtained;

the weight w of the former learner is changed, then the weight w passes through the next learner, and finally all learners form the final learner together; if a sample is misclassified in the previous learner, its corresponding weight will be emphasized, and correspondingly, the weight of the correctly classified sample will be reduced;

the calculation problems of two weights are mainly related to:

1) Weight of sample: 1> under the condition that priori knowledge is not available, the initial distribution is equal-probability distribution, the number of samples is n, and the weight is 1/n;2> updating the weight value for each iteration, and improving the weight of the error-dividing sample;

2) Weight of weak learner: 1> the last strong learner is obtained by combining weights through a plurality of basic learners; and 2, reflecting the influence of different base learners through the weight, wherein the base learners with high accuracy have high weight. In fact a function of the classification error;

gradient Boosting algorithm is different from Adaboost, gradient Boosting selects the gradient descent direction at the time of iteration to ensure the best final result;

the loss function is used for describing the spectrum leaning degree of the model, and the model is assumed to be not fitted, and the larger the loss function is, the higher the error rate of the model is; if the model can make the loss function continuously drop, the model of the invention is improved continuously, and the best mode is to make the loss function drop in the gradient direction;

the model used was the GBDT regression model (Gradient BoostingRegressor) and the algorithm was GBDT (Gradient Boost Decision Tree);

the specific algorithm of the GBDT algorithm is as follows:

first initializing a first weak classifier:

(L loss function is a square error function in regression task)

Sequentially iterating for the remaining M-1 weak classifiers:

c. calculate best fit value for leaf area j=1, 2, ·j

It should be noted that the value corresponding to each leaf node is calculated through the L function, so that the value will change due to the difference of the L loss functions of the regression task and the classification task, which is also a place to be noted by the GBDT algorithm;

the classifier needs to be updated, and the specific algorithm is as follows:

first, the basic idea of this algorithm for constructing decision trees is:

pre-sequencing all the features according to the numerical values of the features; secondly, finding the best segmentation point on a feature by using the cost of O (#data) when traversing the segmentation point; finally, after finding the best division point of a feature, dividing the data into left and right child nodes; the pre-ordering algorithm has the advantages that the partitioning point can be accurately found, but the defects of large space consumption, large time expenditure and unfriendly cache optimization are obvious;

specific: firstly, the space consumption is large, such an algorithm needs to save the characteristic value of the data and also save the characteristic sorting result (for example, for the subsequent rapid calculation of the dividing points, save the sorted index), which needs to consume twice the memory of the training data;

secondly, the time cost is larger, and when each division point is traversed, the calculation of the division gain is needed, so that the consumption cost is high;

finally, the optimization of the cache is not friendly, after the pre-sequencing, the access of the features to the gradient is random access, the different feature access sequences are different, the cache cannot be optimized, meanwhile, when each layer of long tree is needed, a row index is required to be randomly accessed to an array of leaf indexes, the different feature access sequences are also different, and larger cache miss is caused;

a Histogram-based decision tree algorithm;

direct support class feature (Categorical Feature)

Supporting efficient parallelism

Optimizing the hit rate of the Cache;

the principle of the lightGBM optimization algorithm is:

decision tree algorithm based on Histogram

Histogram algorithm (Histogram algorithm), the basic idea of the histogram algorithm is: the continuous floating point eigenvalues are discretized into K K integers, and a histogram with the width of K K is constructed. Accumulating statistic in the histogram according to the discretized value as an index when traversing the data, accumulating the needed statistic in the histogram after traversing the data once, and then traversing to find the optimal segmentation point according to the discrete value of the histogram;

the histogram algorithm is simply understood as: first determining how many bins (bins) are needed for each feature and assigning an integer to each bin; then dividing the range of the floating point number into a plurality of intervals, wherein the number of the intervals is equal to that of the boxes, and updating the sample data belonging to the boxes into the values of the boxes; finally represented by a histogram (#bins).

Feature discretization has many advantages, such as convenient storage, faster operation, strong robustness, more stable model, etc., and the most direct for histogram algorithm has the following two advantages:

the memory occupation is smaller: the histogram algorithm not only does not need to additionally store the pre-ordered result, but also can only store the value after feature discretization, and the value is generally enough to store with 888 bits integer, so that the memory consumption can be reduced to 1/81/81/8 of the original value. That is to say, XGBoost needs to store the characteristic value by using a 323232-bit floating point number and index by using 323232-bit shaping, while LightGBM only needs to store the histogram by using 323232 bits, and the memory is reduced to 1/81/81/8;

the calculation cost is smaller: the pre-ordering algorithm XGBoost needs to calculate a gain of one division every time a feature value is traversed, while the histogram algorithm LightGBM only needs to calculate kk times (kkk can be regarded as a constant), and directly reduces the time complexity from O (#data#feature) O (\text { data } \text { feature }) to O (k#feature), #da > >, k#da > >, k#data > >;

since the features are discretized, the found points are not very accurate, so the results are affected, but the results on different data sets show that the discretized points have better influence on the final accuracy; the reason is that the decision tree is originally a weak model, and whether the segmentation points are accurate or not is not important; thicker division points also have regularization effect, and can effectively prevent overfitting; even though the training error of a single tree is slightly larger than the algorithm of accurate segmentation, there is no significant impact under the framework of gradient lifting (Gradient Boosting);

histogram difference acceleration, another optimization of LightGBM is Histogram (Histogram) difference acceleration;

the histogram of a leaf can be obtained by taking the difference between the histogram of its parent node and the histogram of its sibling, which can be doubled in speed, and when the histogram is usually constructed, all data on the leaf need to be traversed, but the histogram difference only needs to be traversed by k barrels of the histogram, in the process of actually constructing the tree, the LightGBM can also calculate the leaf node with small histogram first, then use the histogram difference to obtain the leaf node with large histogram, so that the histogram of its sibling leaf can be obtained with very tiny cost

Note that: XGBoost accelerates with only non-zero values considered when pre-ordering, while LightGBM also employs a similar strategy: constructing a histogram using only non-zero features;

the Leaf-wise algorithm with depth limitation, the LightGBM is further optimized over the Histogram algorithm;

firstly it discards the layer-wise (level-wise) decision tree growth strategy used by most GBDT tools, and uses a leaf-wise (leaf-wise) algorithm with depth limitation;

XGBoost adopts a Level-wise growth strategy, and the strategy traverses once data to simultaneously split leaves of the same layer, so that multithread optimization is easy to perform, the complexity of a model is well controlled, and fitting is not easy to be performed;

however, in practice, level-wise is an inefficient algorithm, because it treats leaves of the same layer indiscriminately, in practice, the splitting gain of many leaves is low, and searching and splitting are unnecessary, thus bringing about much unnecessary calculation overhead;

the LightGBM adopts a Leaf-wise growth strategy, the strategy finds one Leaf with the maximum splitting gain from all the current leaves at a time, then splits, and the cycle is repeated;

thus, compared with the Level-wise, the advantages of the Leaf-wise are: under the condition that the splitting times are the same, the Leaf-wise can reduce more errors and obtain better precision; the disadvantages of Leaf-wise are: a deeper decision tree may be grown, resulting in an overfitting, so the LightGBM increases a limit on maximum depth above Leaf-wise, preventing overfitting while ensuring high efficiency;

the single-Side Gradient Sampling algorithm, namely Gradient-based One-Side Sampling (GOSS), excludes most of samples with small gradients from the perspective of reducing the samples, and calculates information gain by using the rest samples, wherein the algorithm is balanced in terms of reducing data quantity and guaranteeing precision;

in AdaBoost, the sample weight is an index of data importance, however, in GBDT, no original sample weight is available, and weight sampling cannot be applied, so that each data in GBDT has different gradient values, which is very useful for sampling, i.e. samples with small gradients have small training errors, so that the data are well learned by a model, and the direct idea is to discard the data with small gradients, however, in doing so, the distribution of the data is changed, the accuracy of the training model is affected, and in order to avoid the problem, a GOSS algorithm is proposed;

GOSS is a sampling algorithm of samples, with the aim of losing samples that do not contribute to the calculation of the information gain, leaving behind a contribution;

according to the definition of the calculated information gain, the samples with large gradients have larger influence on the information gain, so that the GOSS only retain the data with large gradients when sampling the data, but if all the data with small gradients are directly discarded and tend to influence the overall distribution of the data, so that all the values of the features to be split are firstly ordered in descending order of absolute value (XGBoost is also ordered, but the result after ordering is not saved by the LightGBM), the a 100% of data with the largest absolute value is selected, then b 100% of data is randomly selected from the rest of the smaller gradient data, b 100% of data is multiplied by a constant 1-a b, frac {1-a } { b1-a, so that the algorithm is more concerned with the rest of data and the data is not more than the original distribution of b 100% of data is calculated (the data is not excessively concerned with the data of the whole data is calculated);

mutually exclusive feature binding algorithms, high-dimensional data tend to be sparse, so a lossless method needs to be designed to reduce the feature dimensions;

features which are usually erased are mutually exclusive (i.e. features cannot be non-zero values at the same time, like one-hot), so that information cannot be lost when two features are erased, if two features are not completely mutually exclusive (in some cases, two features are all non-zero values), an index can be used for carrying out image quantity on the degree of non-mutual exclusion of the features, namely a conflict ratio, and when the value is smaller, two incompletely mutually exclusive features can be selected to be bound to the lung without affecting the final precision;

the mutually exclusive feature lung binding algorithm (Exclusive Feature Bundling, EFB) indicates that the number of features can be reduced if some features are fusion-bound. Thus, the time complexity in constructing the histogram is changed from O (#data #feature) to O (#data #bundle), where #bundle refers to the number of feature packets after feature fusion binding, and #bundle is far smaller than #feature;

the optimized LightGBM has the following advantages over XGBoost:

the speed is faster, the LightGBM adopts a histogram algorithm to convert the traversal samples into traversal histograms, so that the time complexity is greatly reduced; the LightGBM adopts a unilateral gradient algorithm to filter out samples with small gradients in the training process, so that a large number of calculations are reduced; the LightGBM adopts a growth strategy based on a Leaf-wise algorithm to construct a tree, so that a lot of unnecessary calculation amount is reduced; the LightGBM adopts an optimized feature parallel and data parallel method to accelerate calculation, and a voting parallel strategy can be adopted when the data volume is very large; the LightGBM optimizes the cache, so that the cache hit rate is increased;

the memory is smaller, the XGBoost uses the index of the statistical value of the feature value and the corresponding sample thereof to be recorded after the pre-ordering, the LightGBM uses the histogram algorithm to convert the feature value into the bin value, and the index of the feature to the sample is not required to be recorded, so that the space complexity is reduced from O (2 x#data) to O (#bin), and the memory consumption is greatly reduced; the LightGBM adopts a histogram algorithm to convert the storage characteristic value into a storage bin value, so that the memory consumption is reduced; the LightGBM adopts a mutual exclusion feature binding algorithm in the training process, so that the feature quantity is reduced, and the memory consumption is reduced;

to avoid overfitting by growing deeper decision trees, the LightGBM adds a maximum depth limit above Leaf-wise, preventing overfitting while ensuring high efficiency;

the Boosting family is an iterative algorithm, and each iteration carries out weight adjustment on a sample according to the prediction result of the last iteration, so that as the iteration is continuously carried out, the error is smaller and smaller, and the deviation (bias) of the model is continuously reduced;

The system is embodied as follows:

referring to fig. 2, the high-quality production data after being filtered and classified by the algorithm is used for training a model, the system analyzes the best state of the data in the learning history data, and the model is subject to self-iterative evolution;

referring to fig. 3, the system can operate continuously, and an accurate prediction model is formed through 1 ten thousand iterations;

after noise elimination and over-fitting, the prediction model is more accurate; in fig. 3, the dark color represents the actual value, and the light color represents the model predicted value;

referring to fig. 4, the system main interface compares the acquired data with the iterated model to analyze the current production state process during the automatic detection of the current state of the system, so as to display the energy consumption and the generation efficiency state through the main interface, prompt the user to perform tuning when the production energy consumption is too high, and display the normal state if the production energy consumption is in the normal range;

referring to fig. 5, if the energy consumption is too high, the system will provide a solution for adjusting the production efficiency to an optimal state, and the engineer can directly adjust the production line according to the solution given by the system, so as to help the factory save energy and reduce the raw material consumption;

referring to fig. 6, the system generates a real-time energy consumption table, and according to the energy consumption result obtained by the actual test, the system helps to reduce the production cost by 10-15%;

referring to fig. 7, when the system detects a possible fault to be occurred, the system provides the cause of the problem, the treatment opinion and the position of the fault to the user on the intelligent analysis page, so as to facilitate the predictive maintenance of engineers, and the predictive maintenance suggestion of the fault of the spray gun atomizing system is shown in fig. 8.

Referring to fig. 9, for the performance of the system in actual generation, when the user checks the system display to be adjusted, clicks a button to be adjusted, checks the optimization scheme provided by the system and performs corresponding operation, as shown in fig. 10, the suggested natural gas flow is adjusted, so as to help the enterprise to greatly reduce the irregular production stopping times, reduce the raw material consumption, and improve the management efficiency.

Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made therein without departing from the principles and spirit of the invention, the scope of which is defined in the appended claims and their equivalents.

Claims

1. An industrial production artificial intelligence system based on machine learning, which comprises a machine learning model and is characterized by comprising the following steps:

after training is completed, a training result weight file is obtained,

step 3: step 1 is repeated continuously, and after training data samples are increased continuously along with the increase of data provided by factories, training results are updated.

2. The machine learning based industrial production artificial intelligence system of claim 1, wherein the machine learning model further comprises step 4: predictive maintenance, wherein a factory provides historical abnormal data and maintenance conditions, the abnormal data is classified according to different abnormal conditions, the abnormal data is imported into a parameter controllable interval range when each operation device normally operates, the abnormal data is randomly divided into a test set and a verification set after being integrated, and machine learning model training data is built by using a GBDT algorithm;

after training is completed, a training result weight file is obtained,

3. The machine learning based industrial production artificial intelligence system of claim 2, wherein: the solutions include the cause of the occurrence of the abnormal situation or the abnormal situation that may be about to occur, how to solve the problem.

4. The machine learning based industrial production artificial intelligence system of claim 2, wherein: the abnormal data comprise equipment conditions and parameter settings.

5. The machine learning based industrial production artificial intelligence system of claim 1, wherein: in step 1, GBDT algorithm: the data shown by the process.html is calculated according to the weight model_gboost_jingbai.pkl; model weights as used herein are calculated from the GBDT regression model.

6. The machine learning based industrial production artificial intelligence system of claim 1, wherein: in step 1, the specific algorithm of the GBDT algorithm is:

first initializing a first weak classifier:

(L loss function is a square error function in regression task)

Sequentially iterating for the remaining M-1 weak classifiers:

b. taking the negative gradient value obtained in the previous step as a new true value of the sample, and taking the data (x _i ，r _i，m ) I=1, 2, ·n is used as training data for the next tree to obtain a new regression tree f _m (x) The corresponding leaf node area is R _jm J=1, 2,..j, where J is the number of leaf nodes of the regression tree t;

c. calculate best fit value for leaf area j=1, 2, ·j

7. The machine learning based industrial production artificial intelligence system of claim 3, wherein: the classifier needs to be updated, and the specific algorithm is as follows:

8. the machine learning based industrial production artificial intelligence system of claim 1, wherein: in step 1, the GBDT algorithm needs to introduce a LightGBM algorithm, on which the LightGBM is optimized as follows:

specifically, a decision tree algorithm based on a Histogram;

single-sided gradient sampling GOSS: using GOSS to reduce a large number of data instances with only small gradients, and using the remaining data with high gradients to calculate information gain;

mutually exclusive feature binding EFB: binding a number of mutually exclusive features into one feature using EFB;

leaf-growth strategy with depth limitation Leaf-wise: the GBDT algorithm uses a decision tree growth strategy for layer-wise growth; lightGBM uses a per-leaf growth algorithm with depth limiting.