CN109396576B

CN109396576B - Deep learning-based electric spark machining stability and energy consumption state optimization decision system and decision method

Info

Publication number: CN109396576B
Application number: CN201811145126.2A
Authority: CN
Inventors: 马军; 明五一; 李晓科; 都金光; 谢欢; 王旭; 曹阳; 何文斌; 冯士浩
Original assignee: Zhengzhou University of Light Industry
Current assignee: Zhengzhou University of Light Industry
Priority date: 2018-09-29
Filing date: 2018-09-29
Publication date: 2020-02-04
Anticipated expiration: 2038-09-29
Also published as: CN109396576A

Abstract

The invention discloses an electric spark machining stability and energy consumption state optimization decision platform based on deep learning. Mining and analyzing the electric spark machining data by using a characteristic screening method to obtain an optimization index of machining stability and energy consumption state; clustering the optimization indexes by using a K-medoids algorithm to obtain the distribution conditions of the machining stability and the energy consumption state, and constructing a stable machining state database and an energy-saving discharge state database; training LSTM recurrent neural network deep learning by using historical electric spark machining data to obtain a predicted value of a real-time electric spark machining state, combining a statistical characteristic value of an optimization index, judging that the current machining state is an abnormal machining state, performing multi-objective optimization on the machining stability and the energy consumption state to obtain a comprehensive optimal target value of a stable-energy-saving machining state, and regulating and controlling the current machining parameter value according to the comprehensive optimal target value. The invention provides a stable-energy-saving comprehensive optimal electric spark machining parameter optimization decision method based on deep learning, so that electric spark machining is operated in a stable and energy-saving state.

Description

Deep learning-based electric spark machining stability and energy consumption state optimization decision system and decision method

Technical Field

The invention relates to the field of electromachining in the field of special machining, in particular to an electric spark machining stability and energy consumption state optimization decision-making system based on deep learning.

Background

The electric spark machining means that in a certain medium, through pulse discharge between a tool electrode and a workpiece electrode, instantaneous high temperature is formed to locally melt and gasify a workpiece material, so that the material is corroded and removed. The machining method does not generate cutting force, is not limited by the material of the cutter, can machine the workpiece with ultrahigh hardness, brittleness and complex shape, and is widely applied to a plurality of fields of dies, aviation industry, medical appliances and the like. Electric discharge machining is generally performed by an electric discharge machine.

The main process parameters characterizing the electric spark machining performance comprise electric parameters (peak current, peak voltage, pulse width, pulse interval and machining polarity), non-electric parameters (working fluid pressure and scouring mode) and material parameters (specific heat, density, thermal conductivity and melting point) of a workpiece. In the prior art, a plurality of researchers research the influence of the process parameters on the electric spark machining performance from different angles and try to establish an accurate process optimization model. And Wangtong and the like perform simulation analysis on the appearance of the electric erosion pit processed by the electric spark in the single pulse gas, obtain the change rule of the radius, the depth, the volume and the depth-diameter ratio of the electric erosion pit on the surface of the workpiece along with the peak current and the pulse width, and further predict the electric spark processing efficiency and the surface quality of the workpiece. Yao Zhong and the like utilize grey correlation degree analysis to optimize SKD11 electric spark processing electric parameters, and discover the influence rule of pulse discharge time, peak current, pulse gap time and gap voltage on surface roughness and material removal rate. And in the fifth step, the machining speed and the three-dimensional surface roughness are used as evaluation indexes, and the influence of duty ratio, peak current and voltage on the electric spark forming machining of the titanium alloy is analyzed. Because the degree of influence of process parameters on the electric spark machining performance is difficult to express by using an accurate mathematical model, many scholars introduce artificial neural networks, intelligent algorithms and fuzzy mathematics into the electric spark machining process optimization process. The dawn and the like establish a wire cut electrical discharge machining process index prediction model based on a BP neural network by taking pulse width, pulse interval, peak current and machining thickness as process parameters, so that the cutting speed and the surface roughness of wire cutting are combined optimally. Sunzhongfeng and the like propose a fuzzy optimization method for the parameters of the wire cut electric discharge machining process based on neural network modeling and genetic algorithm, thereby improving the probability of obtaining the global optimal solution and being difficult to be trapped into the local optimal solution.

However, in general, studies in the prior art are directed to electric spark finishing of small and medium-sized workpieces, and are being conducted with the aim of optimizing the electric spark machining efficiency and the surface quality of the workpieces. With the rapid increase of global carbon emission and the gradual depletion of energy, the development of 'green economy' becomes a global hotspot, and under the large background, the defects of low energy utilization rate, high energy consumption and unsteady operation of the traditional electric spark machining are increasingly prominent, and particularly for the electric spark rough machining of large workpieces, how to lead the electric spark machining to be operated for a long time in a stable and energy-saving state becomes very urgent. However, the electric spark machining is a complex random process of physical-chemical interaction, characteristic parameters for evaluating the machining stability and the energy consumption state of the electric spark machining have the characteristics of interdisciplinary property, strong coupling and nonlinearity, and how to accurately establish a stable-energy-saving comprehensive optimal electric spark machining process model and optimally control the characteristic parameters becomes a key problem to be solved urgently in the process of green, intelligent and sustainable transformation of the traditional electric spark machining.

Disclosure of Invention

The invention aims to provide an electric spark machining stability and energy consumption state optimization decision system based on deep learning, which has relatively high energy utilization rate and can ensure that electric spark machining can be stably operated for a long time in an energy-saving manner.

In order to achieve the purpose, the invention relates to an electric spark machining stability and energy consumption state optimization decision system based on deep learning, which is characterized in that: the system comprises a data preprocessing module, a processing stability and energy consumption state optimization index mining module, a processing stability and energy consumption state optimization index clustering analysis module, a stable processing state database and energy-saving discharge state database construction module, a processing stability and energy consumption state optimization index statistical characteristic obtaining module, a processing stability and energy consumption state characteristic parameter deep learning module and a processing stability and energy consumption state optimization decision module.

The data preprocessing module is used for extracting, cleaning, fusing and reducing the electric spark machining data acquired by an electric control device of the electric spark machine tool, and providing basic historical electric spark machining data for a subsequent machining stability and energy consumption state optimization index mining module, a machining stability and energy consumption state optimization index cluster analysis module, a stable machining state database, an energy-saving discharge state database construction module and a machining stability and energy consumption state characteristic parameter deep learning module;

the processing stability and energy consumption state optimization index mining module uses a feature screening method (combined with weight) to mine and analyze historical electric spark processing data transmitted by the data preprocessing module, obtains sensitive feature parameters influencing the processing stability and energy consumption, serves as a processing stability and energy consumption state optimization index and provides the sensitive feature parameters to the processing stability and energy consumption state optimization index cluster analysis module;

the processing stability and energy consumption state optimization index cluster analysis module carries out cluster analysis on the historical electric spark processing data transmitted by the data preprocessing module according to the processing stability and energy consumption state optimization index, and marks three processing stability state categories of stability, metastable and instability to obtain the distribution condition of each processing stability state category in the historical electric spark processing data; marking three discharging energy consumption state categories of energy conservation, energy consumption and high energy consumption to obtain the distribution condition of each discharging energy consumption state category in the historical electric spark machining data;

the stable processing state database and energy-saving discharge state database construction module is used for carrying out stable processing type screening on history electric spark processing data transmitted by the data preprocessing module according to a clustering analysis result provided by the processing stability and energy consumption state optimization index clustering analysis module to obtain a stable processing state database; and performing energy-saving discharge type screening on the historical electric spark machining data transmitted by the data preprocessing module to obtain an energy-saving discharge state database.

The processing stability and energy consumption state optimization index statistical characteristic acquisition module acquires real-time electric spark processing data through an electric control device of an electric spark machine tool and calculates a statistical characteristic value of the real-time electric spark processing data to be used as a basis for judging the current electric spark processing state.

The processing stability and energy consumption state characteristic parameter deep learning module is used for carrying out deep learning on historical electric spark processing data by utilizing an LSTM recurrent neural network to obtain a prediction model, predicting the change trend of the real-time electric spark processing state and identifying the current electric spark processing state in an auxiliary manner by using a predicted value;

and the machining stability and energy consumption state optimization decision module judges the current electric spark machining state, and when the current electric spark machining state is judged to be abnormal, the comprehensive optimal target value of the stable-energy-saving machining state is obtained through calculation, and the parameter value of the current electric spark machining parameter is regulated and controlled through an electric control device of the electric spark machine tool.

The invention also discloses a decision method of the electric spark machining stability and energy consumption state optimization decision system based on deep learning, which is sequentially carried out according to the following steps:

the first step is to perform data extraction, data cleaning, data fusion and data reduction processing through a data preprocessing module;

the second step is that a processing stability and energy consumption state optimization index mining module is used for mining the processing stability and energy consumption state optimization indexes;

the third step is to mark the processing stability state category and the discharging energy consumption state category through a processing stability and energy consumption state optimization index cluster analysis module;

the fourth step is that a stable machining state database and an energy-saving discharging state database are constructed through a stable machining state database and an energy-saving discharging state database construction module;

the fifth step is that a statistical characteristic value of real-time electric spark machining data is calculated by a machining stability and energy consumption state optimization index statistical characteristic module and is used as a basis for judging the current electric spark machining state;

the sixth step is that a deep learning module of characteristic parameters of the machining stability and the energy consumption state is utilized to carry out deep learning on historical electric spark machining data to obtain a prediction model, the change trend of the real-time electric spark machining state is predicted, and the current electric spark machining state is assisted and identified by the predicted value;

and the seventh step is that the machining stability and energy consumption state optimization decision module judges the current electric spark machining state, when the current electric spark machining state is judged to be abnormal, a stable-energy-saving machining state comprehensive optimal target value is obtained through calculation, and the parameter value of the current electric spark machining parameter is regulated and controlled through an electric control device of the electric spark machine tool.

In the first step, required historical electric spark machining data are provided for a machining stability and energy consumption state optimization index mining module, a machining stability and energy consumption state optimization index cluster analysis module, a stable machining state database, an energy-saving discharge state database construction module and a machining stability and energy consumption state characteristic parameter deep learning module through four steps of data extraction, data cleaning, data fusion and data reduction.

When data are extracted, firstly, electric spark machining data are obtained through an electric control device of an electric spark machine tool, then main factors and mutual relations influencing machining stability and energy consumption are found out in the electric spark machining data, and parameters and parameter values to be extracted are formed on the basis; the electric spark machining data comprises a plurality of parameters (including 30 parameters obtained by 30 measuring points such as high-frequency power supply equipment of an electric spark machine tool, a wire conveying mechanism, servo control, a scouring part, a machining state and the like), and 8 main parameters are extracted, wherein the main parameters comprise electrode wire moving speed, electrode wire tension, medium temperature, discharge current, discharge pulse width, discharge pulse interval, interelectrode voltage and feeding speed;

clearing fault values, missing values, repetition values and noise values of the 8 main parameters, finishing data cleaning and improving data quality;

carrying out data fusion after data cleaning; the data fusion is to extract the meta-attribute of the data to be fused, analyze the mapping relation between the data to be fused and the target data, map and integrate the data to be fused and the target data together to form a data warehouse;

carrying out data reduction after data fusion; the data reduction is to analyze the statistical characteristics of data records and remove the data records with small attribute value change for the data in the data warehouse; then, performing characteristic correlation analysis on the data records, and removing the data records with strong characteristics to obtain a historical electric spark machining data set with simple characteristics;

the quantization standard for small attribute value changes is as follows: d (x)_i)≤0.01，d(x_i) After normalizing all the data, x_iMinimum distance from other data. If d (x)_i) X is less than or equal to 0.01_iThe change in the attribute value of (c) is small.

The strongly correlated quantification criteria were:

wherein,is x_iMean of all elements in;

is to x by using least square method_i＝f(x₁,x₂,…,x_i-1,x_i+1,…,x_n) The fitting function of (a) is performed,

is x_iMean of all elements in, x_ijIs x_iThe j-th elements of (1). If R is greater than or equal to 0.01, the data x is considered_iStrongly related to other elements.

After the data preprocessing operation, obtaining historical electric spark processing data serving as the basis of subsequent operation, wherein the data set has p parameters; p is 6 or more and 12 or less.

The specific operation of the second step is as follows:

firstly, data acquisition is carried out, wherein the acquisition frequency is M, M is a natural number, and the initial value of M is 0; collecting primary electric spark machining data in a historical electric spark machining data set;

secondly, comparing the values of M and K, and if M is less than K, executing the calculation parameters to obtain the transfer work; if M is larger than or equal to K, skipping to execute the action of calculating the parameter value;

the parameter score calculating action is to screen p parameters in the electric spark machining data by respectively using nine algorithms and calculate the score of each parameter; the nine algorithms are respectively a univariate feature selection algorithm, a Pearson correlation coefficient algorithm, a decision tree algorithm, an L1 regularization algorithm, an L2 regularization algorithm, a random forest algorithm, a random LASSO algorithm, a distance correlation coefficient algorithm and a support vector machine algorithm; processing the score by using an extremum normalization algorithm, and limiting the result to a range of [0, 100] to obtain a normalized score;

then adding 1 to the value of the collection times (namely executing the self-increment operation of M + +) to obtain a value; the acquired preset total times are K, K is a natural number which is more than or equal to 3, and after the value of the acquired times is added with 1, the comparison action is executed by skipping;

when M is larger than or equal to K, skipping from the comparison action to execute the action of calculating the parameter value;

the action of calculating the parameter values is to calculate the score mean, score extreme value and score variance of each parameter (namely the characteristic parameters in the table 1); evaluating and sequencing the importance of each parameter according to the score mean, the score extreme value and the score variance of each parameter, classifying the score mean (namely the average score in the table 1) according to the order of magnitude, wherein the parameter with the score mean being a single digit is an unimportant parameter, the parameter with the score mean being a two-digit parameter is an important parameter, removing the unimportant parameter, retaining the important parameter and taking the important parameter as a processing stability and energy consumption state optimization index, and finishing the operation of mining the processing stability and energy consumption state optimization index.

In the second step, the average value of the electrode wire moving speed, the electrode wire tension and the medium temperature is low in the score condition and is a single digit, and characteristic parameters with low scores are excluded. The characteristic parameters with the highest scores are as follows according to the sequence from high to low: the parameters of the discharge current, the discharge pulse width, the discharge pulse interval, the interelectrode voltage and the feeding speed are used as optimization indexes of the machining stability and the energy consumption state. Table 1 shows data obtained by standard scoring of 8 feature parameters.

TABLE 1 evaluation chart of characteristic parameters of processing stability and energy consumption state

The third step comprises the following specific operations:

the clustering analysis adopts a K-medoids algorithm, and the machining stability state class marking and the discharging energy consumption state class marking are carried out on the historical electric spark machining data;

and marking the historical electrosparking data records as three state categories of stable, metastable and unstable according to the machining stability. Marking historical electrosparking data records as three state categories of energy saving, energy consumption and high energy consumption according to the discharge energy consumption;

for the historical electric spark machining data record containing the five machining stability and energy consumption state optimization indexes, marking the state category by adopting a K-Medoids algorithm;

the stable state is SA, the value range of the discharge current is 3-10A (including two values, the same below), the discharge pulse width is 80-300 mus, the discharge pulse interval is 20-150 mus, the interelectrode voltage is 21-26V, and the feeding speed is 80-128 mm²Between/min;

the metastable state is SB, the discharge current is in the range of 1-5A, the discharge pulse width is in the range of 30-100 mus, the discharge pulse width is in the range of 0-60 mus, the interelectrode voltage is in the range of 20-22.5V, and the feed speed is in the range of 100-145.5 mm²Between/min;

the unstable state category is SC; the value range of the discharge current is between 0 and 1A, the discharge pulse width is between 10 and 50 mu s, the discharge pulse interval is between 0 and 35 mu s,the interelectrode voltage is 18.5-22V, and the feeding speed is 116.5-150 mm²Between/min; table 2 is a processing stability clustering center table;

TABLE 2 Process stability clustering center

The class of the energy-saving state is EA, the value range of the discharge current is 6-10A, the discharge pulse width is 150-280 mus, the discharge pulse width is 50-160 mus, the interelectrode voltage is 20.5-24V, and the feeding speed is 90-135 mm²Between/min;

the type of the energy consumption state is EB, the value range of the discharge current is 3-5A, the discharge pulse width is 60-160 mus, the discharge pulse width is 0-70 mus, the interelectrode voltage is 19.5-22.5V, and the feeding speed is 75-115.5 mm²Between/min;

the high energy consumption state is EC, the value range of the discharge current is between 0 and 2A, the discharge pulse width is between 10 and 80 mu s, the discharge pulse width is between 0 and 45 mu s, the interelectrode voltage is between 18.5 and 21V, and the feeding speed is between 20 and 90mm²And/min. Table 3 is a discharge energy clustering center table;

TABLE 3 discharge energy consumption clustering center

The fourth step is to construct a stable machining state database and an energy-saving discharge state database.

Each electric spark machining data record in the stable machining state database comprises a record number, a parameter name (such as 'discharge current') and a parameter value; each electric spark machining data record in the energy-saving discharge state database comprises a record number, a parameter name and a parameter value.

The specific operation of constructing the stable processing state database is as follows:

firstly, importing and initializing actions, wherein historical electric spark machining data records of the machining stability state class are marked after the cluster analysis in the third step is imported, and the total record number n in the historical electric spark machining data is calculated, wherein n is a natural number; initializing an integer variable i equal to 1;

then, executing a reading action, and reading a parameter value of the ith historical electric discharge machining data record and a state type (namely a state label in the figure 11) corresponding to the machining data record; then judging whether the state type of the processing data record is SA or not, if not, executing the operation of skipping the processing record, and skipping to execute the self-increment operation of i;

if the judgment result is yes, the operation of calculating the Euclidean distance s between the historical electric spark machining data record and each existing machining data record in the stable machining state database is executed;

the calculation of the euclidean distance s is:

if the stable machining state database is an empty database, storing the historical electric spark machining data record into the stable machining state database;

if the machining data records are stored in the stable machining state database, calculating the Euclidean distance s between the parameter value of the historical electric spark machining data record and the corresponding parameter value in each machining data record in the stable machining state database;

if the Euclidean distance s between the historical electric spark machining data record and each machining data record existing in the stable machining state database is larger than 0.01, the historical electric spark machining data record is different from each machining data record existing in the stable machining state database, the historical electric spark machining data record is stored in the stable machining state database, and then the self-increasing operation of i is executed;

if the Euclidean distance s between the historical electric spark machining data record and each existing machining data record in the stable machining state database is smaller than 0.01, executing the operation of skipping the machining record, and skipping the self-increment operation of i;

the self-increment operation of i is to increment the value of i by 1 (namely, to execute the operation of i being i +1), then judge whether i is larger than n, if i is smaller than or equal to n, jump to execute the reading action; if i is larger than n, finishing the operation of constructing the stable machining state database;

the specific operation of constructing the energy-saving discharge state database is as follows:

firstly, importing and initializing actions, wherein historical electric spark machining data records of energy-saving state types are marked after the cluster analysis in the third step is imported, and the total record number m of the historical electric spark machining data is calculated, wherein m is a natural number; initializing an integer variable j equal to 1;

then, executing a reading action, and reading a parameter value of the jth historical electric spark machining data record and a state type (namely a type label in fig. 12) corresponding to the processing data record; then judging whether the state type of the processing data record is EA or not, if not, executing the operation of skipping the processing data record, and skipping to execute the self-increment operation of j;

if the judgment result is yes, the operation of calculating the Euclidean distance s between the machining data record and each existing machining data record in the energy-saving discharge state database is executed;

the calculation of the euclidean distance e is:

if the energy-saving discharge state database is an empty database, storing the historical electric spark machining data record into the energy-saving discharge state database;

if the energy-saving discharge state database stores machining data records, calculating the Euclidean distance e between the parameter value of the historical electric spark machining data record and the corresponding parameter value in each machining data record in the energy-saving discharge state database;

if the Euclidean distance e between the historical electric spark machining data record and each existing machining data record in the energy-saving discharge state database is larger than 0.01, the electric spark machining data record is different from each existing machining data record in the energy-saving discharge state database, the electric spark machining data record is stored in the energy-saving discharge state database, and then self-increasing operation of j is executed; (the meaning of judging the Euclidean distance lies in removing duplication, avoiding storing repeated data in the database, further avoiding increasing the load of the memory and avoiding reducing the operating efficiency by the repeated data)

If the Euclidean distance e between the historical electric spark machining data record and each machining data record in the energy-saving discharge state database is less than or equal to 0.01, executing the operation of skipping the machining record, and skipping the self-increment operation of j;

the self-increment operation of j is to increase the value of j by 1 (namely, to execute the operation of j ═ j +1), then judge whether j is greater than m, if j is less than or equal to m, then jump to execute the reading action; and if j is larger than m, finishing the operation of constructing the energy-saving discharge state database.

In the fifth step, the processing stability and energy consumption state optimization index statistical characteristic obtaining module obtains real-time electric spark processing data through an electric control device of an electric spark machine tool, calculates the mean value, the extreme value, the square difference and the abnormal times of wire breakage of five optimization indexes of discharge current, discharge pulse width, discharge pulse interval, interelectrode voltage and feeding speed in window time, and takes the statistical characteristic value as the basis for identifying the current electric spark processing state.

When the processing stability and energy consumption state optimization index statistical characteristic obtaining module works, importing real-time electric spark processing data records in a K time period (namely a certain time period), counting the total processing record number n (n is a natural number) in the K time period, initializing an integer variable i to be equal to 1, and initializing an integer variable of the abnormal filament breakage times to be equal to 0;

then executing a reading action, and reading the processing data of the ith processing record;

after reading, carrying out broken wire abnormal detection; if the broken wire occurs, adding 1 to the value of the integer variable of the abnormal times of the broken wire, and then executing the statistical action; if no wire breakage occurs, directly executing a statistical action;

the statistical action is to count the broken wire times in the processing record in the K time period, and the mean value, the extreme value and the variance of five parameters of the discharge current, the discharge pulse width, the discharge pulse interval, the interelectrode voltage and the processing speed;

after the statistical action, adding 1 to the value of the integer variable i (namely, executing i to i +1), then judging whether the processing record in the K time period is completely read, specifically judging whether the value of i is greater than the value of n, and if the value of i is less than or equal to the value of n, skipping and executing the reading action again; and if the value i is larger than the value n, transmitting all the statistical characteristic values at the moment to a processing stability and energy consumption state optimization decision module.

The operation of the sixth step is the deep learning of characteristic parameters of the processing stability and the energy consumption state:

1) reading historical electric spark machining data transmitted by the data preprocessing module, and performing characteristic standardization on the historical electric spark machining data by using a Z-score method;

2) inputting the standardized historical electric spark machining data into an LSTM recurrent neural network for training to obtain a prediction model

The core of the LSTM is the state of a neural network unit, the LSTM is transmitted backwards through the whole chain structure, and selective passing of information is realized by using a Sigmoid neural network layer of a threshold structure and dot product operation; the formula for the LSTM algorithm is as follows:

h_t＝o_t*tanh(f_t*c_t-1+i_t*c_in_t)

in the prediction model, the output value at the current time is output gate o_tForgetting door f_tAnd input gate i_tAnd last time cell state c_t-1A function of (a);

3) standardizing values of five characteristic parameters including discharge current, discharge pulse width, discharge pulse interval, interelectrode voltage and feed speed which are acquired in real time by using a Z-score method;

4) and inputting the values of five characteristic parameters of the discharge current, the discharge pulse width, the discharge pulse interval, the interelectrode voltage and the feeding speed after standardization into a prediction model for analysis, obtaining a predicted value through inverse characteristic standardization, and outputting the predicted value to a machining stability and energy consumption state optimization decision module.

The process of training a prediction model and obtaining a prediction value by adopting an LSTM recurrent neural network is explained by taking real-time acquisition data of the inter-electrode voltage and the feeding speed as an example.

First, machining stability and energy consumption state detection is performed on the inter-electrode voltage and feed speed values acquired in a continuous period of time, the counting interval is 60 seconds, and the values of the inter-electrode voltage and the feed speed are continuously 60 as shown in fig. 15.

As shown in fig. 16, the processing stability and energy consumption prediction model is trained by using the LSTM recurrent neural network, and compared with the model using the recurrent neural network RNN and the multi-layer perceptron MLP, the prediction model obtained by using the LSTM recurrent neural network has the highest accuracy.

The relationship between the predicted value and the actual value of the inter-electrode voltage and the feed speed is shown in fig. 17. The prediction error is low, the predicted value can basically reflect the variation trend of the numerical value, and the prediction effect of the model is good.

The operation of the seventh step is a process stability and energy consumption state optimization decision:

1) and the machining stability and energy consumption state optimization decision module receives the statistical characteristic value of the real-time electric spark machining data transmitted by the machining stability and energy consumption state optimization index statistical characteristic acquisition module. If the current electric spark machining state is judged to be an abnormal machining state, activating an optimization decision program;

the criteria for the abnormal processing state are: when the machining stability state is in an unstable state (the unstable state refers to a metastable or unstable state) or the discharge energy consumption state is in a non-energy-saving state (the non-energy-saving state refers to an energy consumption or high energy consumption state).

2) Performing multi-objective optimization on the processing stability and the energy consumption state by using a wolf colony algorithm to obtain a stable-energy-saving processing state multi-objective optimization solution set;

3) searching approximately matched stable machining parameters and energy-saving discharge parameters from a stable machining state database and an energy-saving discharge state database respectively to obtain a comprehensively optimal parameter target value of the stable-energy-saving machining state;

4) and regulating and controlling the parameter value of the current electric spark machining parameter through an electric control device of the electric spark machine tool according to the target value.

The invention has the following advantages:

the invention has the advantages of enabling the electric spark machining to run stably and energy-saving for a long time, improving the workpiece quality, reducing the machining cost, ensuring that the machining stability and the energy consumption index can be kept in a comprehensive optimal interval when the working condition changes, optimally regulating and controlling the characteristic parameters influencing the electric spark machining stability and the energy consumption, and achieving the purpose of green manufacturing.

The invention has simple structure and concise algorithm, finds out key parameters influencing the processing stability and energy consumption from a plurality of parameters in the electric spark processing, constructs a stable processing state database and an energy-saving discharge state database, and compactly and rapidly predicts parameter values considering the dual targets of the processing stability and the energy consumption, thereby efficiently adjusting the processing parameters of the electric spark processing machine tool and ensuring the stability and the energy saving performance of the processing process.

Whether the data mining process is successful or not depends mainly on the quality of the data. The electric spark machine tool generates a large amount of data in the running process, the data not only have multi-source isomerism and various modes, but also have abnormity, omission and repetition, and the subsequent work such as deep learning is difficult to support. The data preprocessing is to perform a series of processing operations such as necessary extraction, cleaning, fusion, reduction and the like on the acquired original electric spark machining data before data mining and use, so that the data quality and the accuracy of a data analysis result are improved, and the follow-up application is better adapted.

Drawings

FIG. 1 is a flow chart of a decision method of the present invention;

FIG. 2 is a flow chart of a first step data pre-processing;

FIG. 3 is a flowchart of the mining process of the optimization index of the process stability and the energy consumption status in the second step;

FIG. 4 is a flow chart of the cluster analysis of the optimization index K-medoids of the processing stability and the energy consumption state in the third step;

FIG. 5 is a graph of the probability density of the parameter distribution for the stability class (SA) in the processing stability state class;

FIG. 6 is a graph of the probability density of the parameter distribution for the metastable class (SB) in the process stability state class;

FIG. 7 is a graph of the probability density of the parameter distribution for the unstable category (SC) in the process stability state category;

FIG. 8 is a parameter distribution probability density plot for the energy saving class (EA) in the discharge energy consumption state class;

FIG. 9 is a graph of the probability density of the distribution of parameters for the energy consumption class (EB) in the discharge energy consumption state class;

FIG. 10 is a graph of the probability density of the parameter distribution for the high energy consumption class (EC) in the discharge energy consumption state class;

FIG. 11 is a flow chart of the creation of a stable processing state database;

FIG. 12 is a flow chart of the establishment of the energy-saving discharge state database;

FIG. 13 is a flowchart of the process stability and energy consumption state optimization indicator statistical characteristic acquisition module;

FIG. 14 is a flowchart of the operation of the process stability and energy consumption state feature parameter deep learning module;

fig. 15 is a continuous value graph of the interelectrode voltage and the feed speed;

FIG. 16 is a graph of the accuracy of the LSTM, RNN and MLP trained predictive models;

fig. 17 is a graph showing a relationship between predicted values and actual values of the inter-electrode voltage and the feed speed;

FIG. 18 is a workflow diagram of a process stability and energy consumption state optimization decision module.

Detailed Description

The invention will be further described with reference to the following examples (drawings):

as shown in fig. 1, the functions of the modules of the electric discharge machining stability and energy consumption state optimization decision platform based on deep learning and the logical relationship between the modules are proposed.

the processing stability and energy consumption state optimization index mining module uses a feature screening method of combined weight to mine and analyze historical electric spark processing data transmitted by the data preprocessing module, obtains sensitive feature parameters influencing the processing stability and energy consumption, serves as a processing stability and energy consumption state optimization index and provides the sensitive feature parameters to the processing stability and energy consumption state optimization index cluster analysis module;

The processing stability and energy consumption state optimization index statistical characteristic obtaining module obtains real-time electric spark processing data through an electric control device of an electric spark machine tool and calculates a statistical characteristic value of the real-time electric spark processing data, and the statistical characteristic value is used as a basis for judging the current electric spark processing state and is provided for the processing stability and energy consumption state optimization decision module.

The processing stability and energy consumption state characteristic parameter deep learning module is used for carrying out deep learning on historical electric spark processing data by utilizing an LSTM recurrent neural network to obtain a prediction model, predicting the change trend of the real-time electric spark processing state and providing the prediction value serving as an auxiliary identification for the current electric spark processing state to the processing stability and energy consumption state optimization decision module;

the machining stability and energy consumption state optimization decision module is used for judging the current electric spark machining state by combining the statistical characteristic value provided by the machining stability and energy consumption state optimization index statistical characteristic acquisition module and the predicted value provided by the machining stability and energy consumption state characteristic parameter depth learning module, calculating to obtain a stable-energy-saving machining state comprehensive optimal target value when the current electric spark machining state is judged to be abnormal, and regulating and controlling the parameter value of the current electric spark machining parameter through an electric control device of the electric spark machine tool.

Fig. 2 shows a flow chart of data preprocessing. Whether the data mining process is successful or not depends mainly on the quality of the data. The electric spark machine tool generates a large amount of data in the running process, the data not only have multi-source isomerism and various mode differences, but also have abnormity, omission and repetition, and the subsequent work such as deep learning is difficult to support. The data preprocessing is a series of processing work of necessary extraction, cleaning, fusion and reduction on the acquired original data before data mining and use, so that the data quality and the accuracy of a data analysis result are improved, and the data preprocessing is better suitable for subsequent application.

The invention is realized by the following technical scheme, and the specific steps are as follows:

in the data preprocessing module, required historical electric spark machining data are provided for the machining stability and energy consumption state optimization index mining module, the machining stability and energy consumption state optimization index cluster analysis module, the stable machining state database and energy-saving discharge state database building module and the machining stability and energy consumption state characteristic parameter deep learning module through four steps of data extraction, data cleaning, data fusion and data reduction.

When data are extracted, main factors and mutual relations influencing the processing stability and the energy consumption are firstly found out in the operation data, and parameters and parameter values to be extracted are formed on the basis. The historical operating data comprises parameters obtained by 30 measuring points such as high-frequency power supply equipment of the electric spark machine tool, a wire conveying mechanism, servo control, a flushing part, a machining state and the like, and 30 main parameters including electrode wire moving speed, electrode wire tension, medium temperature, discharge current, discharge pulse width, discharge pulse interval, interelectrode voltage, feeding speed and the like are obtained through extraction.

In the data preprocessing module, various operation data are collected by various terminals, and a great number of data quality problems exist, such as value loss, repetition, abnormality and the like. The data cleaning is to find and process dirty data in the extracted 30 parameters, and the data which does not meet the requirements is cleaned through a series of processing of removing a fault value, a missing value, a repeated value and a noise value, so that the data quality is further improved.

In the data preprocessing module, the simplification of data volume is realized by data fusion and data reduction: integrating and fusing 30 multi-source and heterogeneous parameters in a unified data warehouse through data mapping so as to facilitate subsequent unified processing; the data is reduced, namely the distribution characteristics, human factors, controllable factors and the like of the electric spark machining data are comprehensively considered, and the specific number of key parameters which can accurately describe the machining stability and the energy consumption state are selected from 30 parameter sets so as to reduce the dimensionality of the data and save the data processing time, so that the selected feature parameter set is simplified to 8 feature parameters: electrode wire moving speed, electrode wire tension, medium temperature, discharge current, discharge pulse width, discharge pulse interval, interelectrode voltage and feeding speed;

fig. 3 shows a flow chart of mining optimization indexes of processing stability and energy consumption state, and the concrete mining steps are as follows:

1) the optimization indexes of the processing stability and the energy consumption state are used as output variables Y_iTaking the characteristic parameter to be selected as an input variable X_iRespectively screening the characteristic parameters to be selected by using nine algorithms, and calculating the score of each characteristic parameter;

2) and processing each score result of the nine algorithms by using an extreme value normalization method, limiting the result to a [0, 100] interval, and then calculating the mean value, the extreme value and the variance of each characteristic parameter for multiple times. And evaluating and sequencing the importance of the characteristic parameters by using the mean value, the extreme value and the variance of the scoring result, and screening the optimization indexes of the processing stability and the energy consumption state. The results obtained after the algorithm was applied to the spark erosion data are shown in table 1 below.

TABLE 1 feature score after screening of feature parameters to be selected by nine algorithms

3) And analyzing the comprehensive scores of the characteristic parameters, and determining the sensitive characteristic parameters which have great influence on the processing stability and the energy consumption by combining the controllability and the actual meaning of the characteristic parameters. From the score condition, the average values of the electrode wire running speed, the electrode wire tension and the medium temperature are low, and characteristic parameters with low scores are excluded. The characteristic parameters with the highest scores are as follows according to the sequence from high to low: discharge current, discharge pulse width, discharge pulse interval, interelectrode voltage, and feed rate.

4) Evaluating the screening results of the characteristic parameters according to the analysis in step 2) and step 3). Among the five characteristic parameters with higher scores, the five characteristic parameters of discharge current, discharge pulse width, discharge pulse interval, interelectrode voltage and feeding speed belong to state variables, and the values of the parameters are obtained under the comprehensive influence of other controllable variables.

And finally determining five characteristic parameters of discharge current, discharge pulse width, discharge pulse interval, interelectrode voltage and feed speed as optimization indexes of machining stability and energy consumption state by combining the analysis.

FIG. 4 shows a flow chart of the process stability and energy consumption state optimization index K-medoids clustering analysis. And (3) carrying out data preprocessing on five optimization indexes of discharge current, discharge pulse width, discharge pulse interval, interpolar voltage and feed speed by combining actual production experience and historical operation data distribution, and taking a processing result as the input of cluster analysis.

And the clustering analysis adopts a K-medoids algorithm to find the state class in the historical electric spark machining data set, and aims to carry out machining stability state class marking and discharging energy consumption state class marking on the historical electric spark machining data.

The processing stability clustering results are as follows, and when n is selected to be 3, the clustering centers and the number of data points in each class are shown in table 2, and the parameter distribution probability density maps of the clusters are shown in fig. 5, 6 and 7.

TABLE 2 processing stability clustering center Table

As can be seen from fig. 5:

the characteristics of class SA: the value range of the discharge current is between 3 and 10A (including two values, the same below), the discharge pulse width is between 80 and 300 mu s, the discharge pulse interval is between 20 and 150 mu s, the interelectrode voltage is between 21 and 26V, and the feed speed is between 80 and 128mm²And/min.

As can be seen from fig. 6:

class SB characteristics: the range of the discharge current is 1-5A, the discharge pulse width is 30-100 mus, the discharge pulse width is 0-60 mus, the interelectrode voltage is 20-22.5V, and the feeding speed is 100-145.5 mm²Between/min;

as can be seen from fig. 7:

the category SC is characterized in that: the value range of the discharge current is between 0 and 1A, the discharge pulse width is between 10 and 50 mu s, the discharge pulse interval is between 0 and 35 mu s, the interelectrode voltage is between 18.5 and 22V, and the feeding speed is between 116.5 and 150mm²Between/min;

and (3) by combining the current situation of electric spark production of a data source, calibrating the records of the class SA obtained by three clustering centers to be in a stable state, calibrating the records of the class SB to be in a metastable state, and calibrating the records in the class SC to be in an unstable state.

The energy consumption clustering result is as follows, when n is 3, the number of the clustering centers and the data points in each class is shown in table 3, and the parameter distribution probability density map of the clustering is shown in fig. 8, 9 and 10.

TABLE 3 energy consumption clustering center Table

As can be seen from fig. 8:

the classification EA characteristics: the value range of the discharge current is 6-10A, the discharge pulse width is 150-280 mus, the discharge pulse interval is 50-160 mus, the interelectrode voltage is 20.5-24V, and the feeding speed is 90-135 mm²Between/min;

as can be seen from fig. 9:

the classification EB characteristic: the value range of the discharge current is 3-5A, the discharge pulse width is 60-160 mus, the discharge pulse interval is 0-70 mus, the interelectrode voltage is 19.5-22.5V, and the feeding speed is 75-115.5 mm²Between/min;

as can be seen from fig. 10:

class EC characteristics: the value range of the discharge current is between 0 and 2A, the discharge pulse width is between 10 and 80 mu s, the discharge pulse interval is between 0 and 45 mu s, the interelectrode voltage is between 18.5 and 21V, and the feeding speed is between 20 and 90mm²And/min.

According to the data source electric spark production current situation suggestion, records of the type EA obtained when three clustering centers are taken are marked as an energy-saving state, records of the type EB are marked as an energy consumption state, and records in the type EC are marked as a high energy consumption state.

According to the calibration of the data state in the clustering, completing the class marking of the machining stability state in the existing historical electric spark machining data, marking the class of the stable state as SA, the class of the metastable state as SB and the class of the unstable state as SC; and completing the class marking of the energy consumption state in the existing historical operating data, marking the class of the energy saving state as EA, the class of the energy consumption state as EB and the class of the high energy consumption state as EC.

The steady state process database construction process is shown in fig. 11. And if the distance is less than 0.01, the machining data record exists in the stable machining state database, and the machining data record does not need to be recorded repeatedly. Otherwise, the historical electric spark machining data record is stored in a stable machining state database.

The energy saving discharge state database construction process is shown in fig. 12. And if the distance is less than 0.01, the machining data record exists in the energy-saving discharge state database, and the machining data record does not need to be recorded repeatedly. Otherwise, the historical electric spark machining data is recorded and stored in an energy-saving discharge state database.

Fig. 13 is a flowchart of the process stability and energy consumption state optimization indicator statistical characteristic obtaining module. The specific statistical process is as follows:

1) and importing real-time processing data in the K time period, and counting the processing record number n in the time period.

2) And (5) detecting the abnormal broken filament values of the n data records one by one. If yes, accumulating the abnormal yarn breaking times in the data record.

And judging the abnormal broken wire according to the value range of each parameter obtained from the stable processing mode library, and when the acquired parameter exceeds the normal range, determining that the data at the moment is an abnormal broken wire value.

3) And (3) counting the mean value, the extreme value and the variance of five parameters including discharge current, discharge pulse width, discharge pulse interval, interelectrode voltage and feeding speed in the current data record, and finally obtaining 20 values of 4 dimensions of the mean value, the extreme value and the variance and the abnormal broken wire times of each parameter in a sampling period as characteristic values for judgment so as to judge the processing stability and the energy consumption state.

4) And continuously reading the next data record for processing, and repeating the process until the last data record.

FIG. 14 is a flowchart of the deep learning module for the characteristic parameters of process stability and energy consumption status.

1) And reading historical operating data, and performing characteristic standardization on the historical operating data by using a Z-score method.

2) And inputting the standardized historical operating data into an LSTM recurrent neural network for training to obtain a prediction model.

The core of the LSTM is the state of a neural network unit, the state is transmitted backwards through the whole chain structure, and selective passing of information is achieved by means of a Sigmoid neural network layer of a threshold structure and dot product operation. The calculation formula of the LSTM algorithm is as follows:

h_t＝o_t*tanh(f_t*c_t-1+i_t*c_in_t)

the model considers that the output value at the current moment is an output gate o_tForgetting door f_tAnd input gate i_tAnd last time cell state c_t-1As a function of (c). LSTM considers model prediction h of common influence of historical sequence values and unit state parameters_tThe value of (a).

The process of training a prediction model and obtaining a prediction value by adopting an LSTM recurrent neural network is explained by taking the collected data of the inter-electrode voltage and the feeding speed as an example.

FIG. 18 is a flow chart of the process stability and energy consumption state optimization decision module. The specific decision method is as follows:

the criteria for the abnormal processing state are: when the processing stability state is in an unstable state (metastable or unstable state) or the discharge energy consumption state is in a non-energy-saving state (energy consumption or high energy consumption state).

Although the present invention has been described in detail with reference to the above embodiments, it should be understood by those skilled in the art that: the invention can be modified and equivalents substituted for elements thereof without departing from the spirit and scope of the invention, which should be construed as limited only by the claims.

Claims

1. Electric spark machining stability and energy consumption state optimization decision-making system based on deep learning is characterized in that: the system comprises a data preprocessing module, a processing stability and energy consumption state optimization index mining module, a processing stability and energy consumption state optimization index clustering analysis module, a stable processing state database and energy-saving discharge state database construction module, a processing stability and energy consumption state optimization index statistical characteristic acquisition module, a processing stability and energy consumption state characteristic parameter deep learning module and a processing stability and energy consumption state optimization decision module;

the data preprocessing module is used for extracting, cleaning, fusing and reducing the electric spark machining data acquired by an electric control device of the electric spark machine tool, and providing basic historical electric spark machining data for a subsequent machining stability and energy consumption state optimization index mining module, a machining stability and energy consumption state optimization index cluster analysis module, a stable machining state database and energy-saving discharge state database construction module and a machining stability and energy consumption state characteristic parameter deep learning module;

the processing stability and energy consumption state optimization index mining module uses a feature screening method to mine and analyze historical electric spark processing data transmitted by the data preprocessing module, obtains sensitive feature parameters influencing the processing stability and energy consumption, serves as a processing stability and energy consumption state optimization index and provides the sensitive feature parameters to the processing stability and energy consumption state optimization index cluster analysis module;

the processing stability and energy consumption state optimization index cluster analysis module carries out cluster analysis on the historical electric spark processing data transmitted by the data preprocessing module according to the processing stability and energy consumption state optimization index, marks three processing stability state categories of stable, metastable and unstable, and obtains the distribution condition of each processing stability state category in the historical electric spark processing data; marking three discharging energy consumption state categories of energy conservation, energy consumption and high energy consumption to obtain the distribution condition of each discharging energy consumption state category in the historical electric spark machining data;

the stable processing state database and energy-saving discharge state database construction module is used for carrying out stable processing type screening on the historical electric spark processing data transmitted by the data preprocessing module according to a clustering analysis result provided by the processing stability and energy consumption state optimization index clustering analysis module to obtain a stable processing state database; performing energy-saving discharge type screening on historical electric spark machining data transmitted by the data preprocessing module to obtain an energy-saving discharge state database;

the processing stability and energy consumption state optimization index statistical characteristic acquisition module acquires real-time electric spark processing data through an electric control device of an electric spark machine tool and calculates a statistical characteristic value of the real-time electric spark processing data to be used as a basis for judging the current electric spark processing state;

2. The decision method for using the deep learning based electric discharge machining stability and energy consumption state optimization decision system as claimed in claim 1 is characterized by sequentially carrying out the following steps:

marking the type of the machining stability state and the type of the discharging energy consumption state by a machining stability and energy consumption state optimization index cluster analysis module;

the sixth step is that a deep learning module of characteristic parameters of the machining stability and the energy consumption state is used for deep learning historical electric spark machining data to obtain a prediction model, the change trend of the real-time electric spark machining state is predicted, and the current electric spark machining state is assisted and identified by the predicted value;

and the seventh step is that the machining stability and energy consumption state optimization decision module judges the current electric spark machining state, when the current electric spark machining state is judged to be abnormal, the comprehensive optimal target value of the stable-energy-saving machining state is obtained through calculation, and the parameter value of the current electric spark machining parameter is regulated and controlled through an electric control device of the electric spark machine tool.

3. The decision-making method according to claim 2, characterized in that:

in the first step, required historical electric spark machining data are provided for a machining stability and energy consumption state optimization index mining module, a machining stability and energy consumption state optimization index cluster analysis module, a stable machining state database, an energy-saving discharge state database construction module and a machining stability and energy consumption state characteristic parameter deep learning module through four steps of data extraction, data cleaning, data fusion and data reduction;

when data are extracted, firstly, electric spark machining data are obtained through an electric control device of an electric spark machine tool, and then 8 main parameters including electrode wire moving speed, electrode wire tension, medium temperature, discharge current, discharge pulse width, discharge pulse interval, interelectrode voltage and feeding speed are extracted;

clearing fault values, missing values, repeated values and noise values of the 8 main parameters, completing data cleaning and improving data quality;

carrying out data reduction after data fusion; the data reduction is to perform data record statistical characteristic analysis on data in a data warehouse and remove data records with small attribute value changes; performing characteristic correlation analysis on the data records, and removing the data records with strong characteristics to obtain a historical electric spark machining data set;

the quantization standard for small attribute value changes is as follows: d (x)_i)≤0.01，d(x_i) Is after normalizing all data, x_iMinimum distance from other data; if d (x)_i) X is less than or equal to 0.01_iThe change of the attribute value of (2) is small;

the strongly correlated quantification criteria were: r is more than or equal to 0.01,

wherein,

is x_iMean of all elements in;

is x_iAll of the elements inMean value of (1), x_ijIs x_iThe j-th element of (1); if R is greater than or equal to 0.01, the data x is considered_iStrongly related to other elements;

after the data preprocessing operation, obtaining historical electric spark machining data serving as the basis of subsequent operation, wherein the data set has p parameters; p is 6 or more and 12 or less.

4. The decision-making method according to claim 3, characterized in that: the specific operation of the second step is as follows:

secondly, comparing the values of M and K, and if M is less than K, executing a parameter scoring action; if M is larger than or equal to K, skipping to execute the action of calculating the parameter value;

then adding 1 to the value of the acquisition times; the preset total number of times of collection is K, the K is a natural number which is more than or equal to 3, and after the value of the number of times of collection is added with 1, the jump is carried out to compare the actions;

the action of calculating the parameter values is to calculate and obtain the score mean value, the score extreme value and the score variance of each parameter; and evaluating and sequencing the importance of each parameter according to the score mean, the score extreme value and the score variance of each parameter, classifying the score mean according to the order of magnitude, wherein the parameter with the score mean being a single digit is an unimportant parameter, the parameter with the score mean being a two-digit parameter is an important parameter, removing the unimportant parameter, reserving the important parameter and taking the important parameter as a processing stability and energy consumption state optimization index, and finishing the operation of mining the processing stability and energy consumption state optimization index.

5. The decision-making method according to claim 2, characterized in that: in the second step, five parameters of discharge current, discharge pulse width, discharge pulse interval, interelectrode voltage and feed speed are used as optimization indexes of machining stability and energy consumption state in the order of grading average value from high to low.

6. The decision-making method according to claim 2, characterized in that: the third step comprises the following specific operations:

marking the historical electrosparking data records as stable, metastable and unstable state categories according to the machining stability; marking historical electric spark machining data records as three state categories of energy saving, energy consumption and high energy consumption according to discharge energy consumption;

for the historical electric spark machining data record containing the five machining stability and energy consumption state optimization indexes, marking the state type by adopting a K-medoids algorithm;

the stable state is SA, the value range of the discharge current is between 3 and 10A, the discharge pulse width is between 80 and 300 mu s, the discharge pulse width is between 20 and 150 mu s, the interelectrode voltage is between 21 and 26V, and the feeding speed is between 80 and 128mm²Between/min;

the unstable state category is SC; the value range of the discharge current is between 0 and 1A, the discharge pulse width is between 10 and 50 mu s, and the discharge pulses are separated0 to 35 mu s, an interelectrode voltage of 18.5 to 22V, and a feeding speed of 116.5 to 150mm²Between/min;

the high energy consumption state is EC, the value range of the discharge current is between 0 and 2A, the discharge pulse width is between 10 and 80 mu s, the discharge pulse width is between 0 and 45 mu s, the interelectrode voltage is between 18.5 and 21V, and the feeding speed is between 20 and 90mm²And/min.

7. The decision-making method according to claim 2, characterized in that: the fourth step is to construct a stable machining state database and an energy-saving discharge state database;

each electric spark machining data record in the stable machining state database comprises a record number, a parameter name and a parameter value; each electric spark machining data record in the energy-saving discharge state database comprises a record number, a parameter name and a parameter value;

firstly, importing and initializing actions, wherein historical electric spark machining data records of the type of the machining stability state are marked after the cluster analysis in the third step is imported, and the total record number n in the historical electric spark machining data is calculated, wherein n is a natural number; initializing an integer variable i equal to 1;

then executing a reading action, and reading the parameter value of the ith historical electric spark machining data record and the state type corresponding to the machining data record; then judging whether the state type of the processing data record is SA or not, if not, executing the operation of skipping the processing record, and skipping to execute the self-increment operation of i;

the calculation of the euclidean distance s is:

if the Euclidean distance s between the historical electric spark machining data record and each existing machining data record in the stable machining state database is larger than 0.01, the historical electric spark machining data record is different from each existing machining data record in the stable machining state database, the historical electric spark machining data record is stored in the stable machining state database, and then self-increasing operation of i is executed;

the self-increment operation of i is to increase the value of i by 1, then judge whether i is greater than n, if i is less than or equal to n, jump to execute the reading action; if i is larger than n, finishing the operation of constructing the stable machining state database;

firstly, importing and initializing actions, wherein after importing the cluster analysis of the third step, historical electric spark machining data records with energy-saving state categories are marked, and the total record number m of the historical electric spark machining data is calculated, wherein m is a natural number; initializing an integer variable j equal to 1;

then executing a reading action, and reading a parameter value of the jth historical electric spark machining data record and a state type corresponding to the machining data record; then judging whether the state type of the processing data record is EA or not, if not, executing the operation of skipping the processing data record, and skipping to execute the self-increment operation of j;

the calculation of the euclidean distance e is:

if the Euclidean distance e between the historical electric spark machining data record and each machining data record existing in the energy-saving discharge state database is larger than 0.01, the electric spark machining data record is different from each machining data record existing in the energy-saving discharge state database, the electric spark machining data record is stored in the energy-saving discharge state database, and then self-increasing operation of j is executed;

the self-increment operation of j is to increase the value of j by 1, then judge whether j is greater than m, if j is less than or equal to m, jump to execute the reading action; and if j is larger than m, finishing the operation of constructing the energy-saving discharge state database.

8. A decision method as claimed in claim 2, characterized in that: in the fifth step, the processing stability and energy consumption state optimization index statistical characteristic obtaining module obtains real-time electric spark processing data through an electric control device of an electric spark machine tool, calculates the mean value, the extreme value, the variance and the abnormal broken wire times of five optimization indexes of discharge current, discharge pulse width, discharge pulse interval, interelectrode voltage and feeding speed in window time, and takes the statistical characteristic value as the basis for identifying the current electric spark processing state;

when the processing stability and energy consumption state optimization index statistical characteristic obtaining module works, importing real-time electric spark processing data records in a K time period, counting the total processing record number n in the K time period, initializing an integer variable i to be equal to 1, and initializing an integer variable of the abnormal broken wire times to be equal to 0;

after the statistical action, adding 1 to the value of an integer variable i, then judging whether the reading of the processing record in the K time period is finished, specifically judging whether the value of i is greater than the value of n, and if the value of i is less than or equal to the value of n, skipping and executing the reading action again; and if the value i is larger than the value n, transmitting all the statistical characteristic values at the moment to a processing stability and energy consumption state optimization decision module.

9. The decision-making method according to claim 2, characterized in that: the operation of the sixth step is the deep learning of characteristic parameters of the processing stability and the energy consumption state:

The core of the LSTM is the state of a neural network unit, the LSTM is transmitted backwards through the whole chain structure, and selective passing of information is realized by using a Sigmoid neural network layer of a threshold structure and dot product operation; the calculation formula of the LSTM algorithm is as follows:

h_t＝o_t*tanh(f_t*c_t-1+i_t*c_in_t)

10. The decision-making method according to claim 2, characterized in that: the operation of the seventh step is a process stability and energy consumption state optimization decision:

1) the processing stability and energy consumption state optimization decision module receives a statistical characteristic value of real-time electric spark processing data transmitted by the processing stability and energy consumption state optimization index statistical characteristic acquisition module; if the current electric spark machining state is judged to be an abnormal machining state, activating an optimization decision program;

the criteria for the abnormal processing state are: when the machining stability state is in an unstable state or the discharge energy consumption state is in a non-energy-saving state;

3) searching a stable machining parameter and an energy-saving discharge parameter which are approximately matched from a stable machining state database and an energy-saving discharge state database respectively to obtain a parameter target value which is comprehensively optimal in a stable-energy-saving machining state;