A Knowledge Graph-Based Magnetic Separation Optimization Control Method

Abstract

This invention relates to a knowledge graph-based magnetic separation optimization control method. It utilizes the knowledge graph method to construct a knowledge graph reasoning model for magnetic separation optimization control, combines reinforcement learning strategies to learn from existing control decision-making experience, and explores unknown strategies using the model. This satisfies the component and field strength optimization control of the magnetic separation process, and realizes a closed-loop model and optimization control of mineral processing indicators for multiple sets of magnetic separators.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent control of mineral processing, in particular to a magnetic separation optimization control method based on a knowledge graph. BACKGROUND

[0002] The magnetic separation process is an important link in the iron ore beneficiation production process. After grinding, the mineral particles pass through the magnetic field of the magnetic separator and are subjected to the combined action of magnetic force and mechanical force, realizing the final separation of minerals. However, most of the iron ore in China is poor, and the working condition of magnetic separation operation is time-varying, which needs to be adjusted according to the fluctuation of the nature of the incoming ore. The traditional control method is usually based on the past production experience of the on-site process personnel, which has strong subjective randomness and cannot form a relatively stable control closed loop, so it cannot effectively solve the sub-optimal problem under frequent working condition changes.

[0003] Especially when controlling the beneficiation indexes of the magnetic separation process of multiple magnetic separators, it will cause the production status of strong nonlinearity, strong coupling, time-varying and other complex dynamic characteristics. The traditional optimization control method uses artificial experience, PID, fuzzy learning and other methods, which has strong subjective randomness; the optimization control based on reinforcement learning focuses on historical control records and it is difficult to find new optimization decisions; although the optimization control algorithm using neural network has good self-learning and self-adaptability, the learning cost is high and the control cannot be effectively described.

[0004] In order to realize the optimization control of the beneficiation indexes of the magnetic separation process of multiple magnetic separators and overcome the production status of strong nonlinearity, strong coupling, time-varying and other complex dynamic characteristics, it is necessary to use the knowledge graph method, combine the existing control decision experience to build an optimization control model, and satisfy the exploration of unknown strategies, and finally realize the component and field strength optimization control of the magnetic separation process. Therefore, it is important to construct a magnetic separation control method based on a knowledge graph. SUMMARY

[0005] The purpose of the present application is to provide a magnetic separation optimization control method based on a knowledge graph, which uses the knowledge graph method to build a large knowledge network, combines the reinforcement learning strategy to learn the existing control decision experience, and combines the model to explore unknown strategies, satisfying the optimization control of the component and field strength of the magnetic separation process.

[0006] The purpose of the present application is achieved by the following technical solutions:

[0007] The magnetic separation optimization control method based on the knowledge graph comprises the following steps:

[0008] S1. Real-time acquisition of relevant data from multiple magnetic separators during the magnetic separation process, including the opening size of the feed plug, the opening size of the tailings valve, the excitation status, the number of excitation stations, the high-quality product level of the first-stage strong magnetic separator, the high-quality product level of the mixed magnetic separator, the flow rate of the large bottom flow pump before the strong magnetic separator, the pump frequency of the large bottom flow pump before the strong magnetic separator, and the concentration of the large bottom flow pump before the strong magnetic separator. At the same time, the upper and lower limits of the relevant data are set, such as the mixed magnetic high-quality product level in the range of 45-55, the strong magnetic large bottom flow pump frequency in the range of 30-44, and the tailings valve opening size in the range of 9.5-40.

[0009] S2. Clean the real-time acquired data, removing duplicate data and data exceeding production limits. For data such as mixed magnetic high-quality grade, first-stage strong magnetic high-quality grade, and mixed magnetic tailings grade, if the data is identical for three consecutive acquisition cycles, it is considered duplicate data. For data with frequent changes, such as the flow rate and concentration of the strong magnetic pre-well bottom flow pump, if the data is identical for two consecutive acquisition cycles, it is considered duplicate data. For data with set and feedback parameters, such as the opening size of the feed plug and the tailings valve, a difference exceeding 10% is considered abnormal data and cleaned. Kalman filtering is applied to all magnetic separator feed plug opening sizes, tailings valve opening sizes, excitation status, number of opening units, first-stage strong magnetic high-quality grade, and mixed magnetic high-quality grade data within one hour to stabilize time data characteristics. The cleaned real-time data and historical data within one hour are merged into a defined adjustment target set. Two months of historical control records are saved as the historical control set.

[0010] S3. Extract triplets from historical data. When the monitoring data such as the liquid level of the magnetic separator, the opening size of the tailings valve of the magnetic separator, the magnetic separator strength, the mixed magnetic high-quality level, the first-stage strong magnetic high-quality level, the concentration and flow rate of the strong magnetic front well exceed 2% within 10 seconds, detect whether the data such as the opening size of the feed plug, the number of magnetic separators in operation, the frequency of the bottom flow pump of the strong magnetic front well, and the magnetic separator strength have changed at the same time and extract them to form a knowledge graph training dataset for magnetic separation optimization control.

[0011] S4. Using the set of triples in the knowledge graph training dataset constructed in S3 as input, model the set of triples to construct a knowledge graph reasoning model based on magnetic separation optimization control.

[0012] S5. Using the set of adjustment targets obtained in S2, find the target labels that need to be regulated;

[0013] S6. Based on the historical control set in S2, select the optimal control parameters for the target label;

[0014] Preferably, the target labels include: the size of the feed plug opening, the number of magnetic separators in operation, the frequency of the large bottom flow pump before the strong magnetic separator, and the magnetic separator strength, etc.

[0015] Preferably, the S4 constructs a knowledge graph inference model based on the optimized control of magnetic selection includes the following steps:

[0016] S4.1, any triple is expressed as (h, r, t)

[0017] For a given triple (h, r, t), it has the following phase h, r, t, h p , r p , t p ; h p , r p , t p They are the projection vectors of h, r, t, respectively, indicating the mapping of entities from entity space to relationship space, h p , r, All vectors are normalized;

[0018] S4.2, M rh , M rt are the mapping matrices of entities h, t, respectively, h ip , t ip (i = 1, 2, 3…) and r p is the projection vector, h i┴ , t i┴ are the projection vectors of the head and tail entities, and the mapping matrix is:

[0019]

[0020]

[0021] Therefore, the head entity and tail entity projection vectors are respectively:

[0022] h ⊥ = M rh h,

[0023] t ⊥ = M rt t,

[0024] Thus, for any triple, the score function is the Euclidean distance between the mapped head entity, the mapped relationship and the mapped tail entity:

[0025]

[0026] S4.3, the loss function of triple (h, r, t) is:

[0027]

[0028] Wherein △ is a correct triple set, △' is an incorrect triple set, ξ is a correct triple, ξ' is an incorrect triple, the incorrect triple is that the head entity or the tail entity of the correct triple is replaced, and the replaced triple is not in the correct triple set;

[0029] All replacements of the head entity and the tail entity in the triple are converged; if the loss function is greater than 0, no convergence is performed; if the loss function is less than 0, the random gradient descent algorithm is used to converge the incorrect triple;

[0030] S4.4, repeat S4.2 and S4.3 to sequentially converge all correct triples, and complete the construction of the knowledge graph reasoning model based on the magnetic selection optimization control;

[0031] Preferably, the obtaining of the label needing to be regulated in S5 comprises the following steps:

[0032] S5.1, obtaining a gradient according to historical data in the regulation target set, and setting a parameter k;

[0033] S5.2, for any triple (h, r, t), according to the established knowledge graph reasoning model, a new scoring function is obtained through the parameter k set in S6.1:

[0034]

[0035] At this time, a set S is obtained, wherein

[0036] S={f r (h,t)|f r (h,t)≥0},

[0037] Select the tail entity with a score greater than 0 and sort them from large to small according to the score;

[0038] S5.3, in the historical data, the same operation as the target value is obtained, and the number of regulation target labels num is selected, at this time, the first num tail entities in the set S are taken as t1, t2, t3, …t num ;

[0039] Preferably, the selecting of the optimal regulation parameter of the target label according to the historical regulation set in S2 in S6 comprises the following steps:

[0040] S6.1, from the historical regulation set, a regulation record set similar to the current production state is obtained, and the optimal regulation record is selected according to the mixed magnetic quality position, the magnetic separator liquid level and the magnetic separator tailing valve opening degree as the benefit; if there are multiple records, take the average:

[0041]

[0042] wherein S hcj , S fj , S yw are the scores given by the mixed magnetic concentrate level, the magnetic separator liquid level, and the opening size of the magnetic separator tailings valve, and the value is 1 if the relative result is good, otherwise the value is 0; S1, S2, and S3 are the weights of the mixed magnetic concentrate level, the magnetic separator liquid level, and the opening degree of the magnetic separator tailings valve, respectively; and the weights finally add up to 1; if there is no similar control record, then 1% control is performed in the control direction;

[0043] S6.2, set the parameter SC, select the highest score of the control result, if the obtained score is lower than SC, then add or subtract 10% to the control data based on the control of the parameter;

[0044] S6.3, if the obtained score is higher than SC, then consider that the parameter is the optimal result, the mediation parameter is a local optimal solution, in order to explore the optimal result, a 0-1 random number is used to judge whether to test and adjust based on this,

[0045] when the random number is in the range of 0-0.05, increase by 5%;

[0046] when the random number is in the range of 0.05-0.1, decrease by 5%;

[0047] when the random number is in the range of 0.1-1, adjust normally according to the local optimal result;

[0048] S6.4, obtain the specific parameters of the adjustment according to S6.1, S6.2, and S6.3.

[0049] The magnetic separation process optimization control system based on a knowledge graph of the application has the following advantages compared with the prior art:

[0050] The magnetic separation process optimization control system based on a knowledge graph of the application has the following advantages compared with the prior art:

[0051] The application provides a magnetic separation optimization control system. By constructing a magnetic separation optimization control knowledge graph, artificial knowledge experience can be contained and ambiguity can be eliminated, and new optimization control processes can be predicted and tried. In specific parameter adjustment control, reinforcement learning is used to effectively avoid fuzzy adjustment of artificial experience, so that parameter adjustment is more detailed. On the basis, the magnetic separator liquid level is stable and does not run or overflow, the liquid level difference of multiple magnetic separators is less than 2.5, the mixed magnetic concentrate grade is normal and stable, the mixed magnetic tail grade is lower than the process set value, so that the flotation concentrate grade can be normal and stable, and the yield is increased. The deep integration of advanced technologies such as Internet, big data and artificial intelligence and traditional magnetic separation process is realized. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 The method of the application is a structural flow chart.

[0053] Figure 2 The knowledge graph structure diagram is generated. DETAILED DESCRIPTION

[0054] The application will be further described below in combination with the drawings and specific embodiments:

[0055] As shown in Figure 1 and Figure 2 , a magnetic separation optimization control method based on a knowledge graph, characterized in that it comprises the following steps:

[0056] S1, real-time collection of related data of the magnetic separation process of multiple magnetic separators, mainly including the size of the magnetic separator ore feeding rubber plug opening, the size of the tailing valve opening, the excitation state, the number of openings, the first strong magnetic concentrate grade, the mixed magnetic concentrate grade, the strong magnetic pre strong well bottom flow pump flow, the strong magnetic strong well bottom flow pump frequency, the strong magnetic pre strong well bottom flow pump concentration, etc., while setting the upper and lower limits of the related data, the mixed magnetic concentrate grade is in the range of 45-55, the strong magnetic strong well bottom flow pump frequency is in the range of 30-44, the tailing valve opening size is in the range of 9.5-40, etc.;

[0057] S2, the real-time collected data is cleaned, the repeated data and the data exceeding the production upper and lower limits are removed, the data of the mixed magnetic concentrate grade, the one-stage strong magnetic concentrate grade, the mixed magnetic tail grade and the like are considered as repeated data if the data of three consecutive collection cycles are the same, the data of the strong magnetic front large well bottom flow pump flow, the strong magnetic front large well bottom flow pump concentration and the like are considered as repeated data if the data of two consecutive collection cycles are the same, the data of the mine feeding rubber plug opening size, the tailing valve opening size and the like have a set and feedback, and the data are considered as abnormal data if the difference is more than 10%, and the cleaning is performed; the Kalman filter is performed on the mine feeding rubber plug opening size, the tailing valve opening size, the magnetic field, the opening number, the one-stage strong magnetic concentrate grade and the mixed magnetic concentrate grade of all the magnetic separators within one hour, so as to realize the stabilization of the time data characteristics; the real-time data after the cleaning and the historical data within one hour are combined to define the adjustment target set; the historical adjustment and control records of two months are saved as the historical adjustment and control set;

[0058] S3, the historical data is extracted into triplets, when the monitoring data of the magnetic separator liquid level, the magnetic separator tailing valve opening size, the magnetic field, the mixed magnetic concentrate grade, the one-stage strong magnetic concentrate grade, the strong magnetic front large well concentration and flow and the like change by more than 2% within 10 seconds, whether the data of the mine feeding rubber plug opening size, the magnetic separator opening number, the strong magnetic front large well bottom flow pump frequency and the magnetic field change is detected and extracted, so as to constitute the knowledge graph training data set of the magnetic separation optimization control;

[0059] S4, the triplets in the knowledge graph training data set constituted by S3 are taken as input, the triplets are modeled, and thus the knowledge graph reasoning model based on the magnetic separation optimization control is constructed;

[0060] The knowledge graph reasoning model based on the magnetic separation optimization control constructed in S4 includes the following steps:

[0061] S4.1, any triplet is expressed as (h, r, t)

[0062] For a given triplet (h, r, t), it has the following phase quantities h, r, t, h p , r p , t p ; h p , r p , t p They are the projection vectors of h, r, t, respectively, indicating that the entity is mapped from the entity space to the relationship space, h p , r, All vectors are normalized;

[0063] S4.2, M rh ,M rtThese are the mapping matrices of entities h and t, respectively. ip t ip (i = 1, 2, 3...) and r p h is the projection vector. i┴ t i┴ Let be the projection vectors of the head and tail entities, respectively, and the mapping matrix be:

[0064]

[0065]

[0066] Therefore, the projection vectors of the head entity and the tail entity are respectively:

[0067] h ⊥ =M rh h,

[0068] t ⊥ =M rt t,

[0069] Therefore, the score function for any triple is the mapped head entity plus the Euclidean distance between the mapped relation and the mapped tail entity:

[0070]

[0071] S4.3 The loss function for the triple (h,r,t) is:

[0072] L=∑ ξ∈Δ ∑ ξ′∈Δ′ (Υ+f r (ξ)-f r (ξ′)),

[0073] Where △ is the set of correct triples, △' is the set of incorrect triples, ξ is the correct triple, and ξ' is the incorrect triple. An incorrect triple is a correct triple that has replaced the head entity or the tail entity, and the replaced triple is not in the set of correct triples.

[0074] All replacements of the head and tail entities in the triplet converge; if the loss function is greater than 0, convergence does not occur; if the loss function is less than 0, the stochastic gradient descent algorithm is used to converge the erroneous triplet.

[0075] S4.4, repeat S4.2 and S4.3 to perform convergence operations on all correct triples in sequence, and complete the construction of the knowledge graph reasoning model based on magnetic separation optimization control;

[0076] S5. Using the set of adjustment targets obtained in S2, find the target labels that need to be regulated, which mainly include: the opening size of the feed plug, the number of magnetic separators in operation, the frequency of the large bottom flow pump before the strong magnetic separator, and the magnetic separator's field strength, etc.

[0077] S5.1 Obtain the gradient based on historical data in the target set and set the parameter k;

[0078] S5.2 For any triple (h,r,t), based on the established knowledge graph reasoning model, and using the parameter k set in S5.1, a new scoring function is obtained:

[0079]

[0080] At this point, we obtain set S, where

[0081] S={f r (h,t)|f r (h,t)≥0},

[0082] Select tail entities with scores greater than 0 and sort them from largest to smallest;

[0083] S5.3. In historical data, retrieve the operation that is the same as the target value, and simultaneously select the number of target labels to adjust, num. At this time, take the first num tail entities in set S, t1, t2, t3, ..., tn. num ;

[0084] S6. Based on the historical control set in S2, select the optimal control parameters for the target label.

[0085] S6.1. From the historical control set, obtain a set of control records similar to the current production state, and select the optimal control record based on the benefits of the mixed magnetic concentrate level, magnetic separator liquid level, and magnetic separator tailings valve opening degree; if there are multiple records, take the average:

[0086] score=∑ n (s1×s hcj +s2×s fj +s3×s yw ) / n,

[0087] Among them, S hcj S fj S yw The scores are given for the mixed magnetic concentrate level, the liquid level of the magnetic separator, and the opening degree of the tailings valve of the magnetic separator, respectively. If the relative result is better, the value is 1; otherwise, the value is 0. S1, S2, and S3 are the weights of the mixed magnetic concentrate level, the liquid level of the magnetic separator, and the opening degree of the tailings valve of the magnetic separator, respectively. The weights are finally totaled to 1. If there is no similar control record, the control is adjusted by 1% according to the control direction.

[0088] S6.2 Set parameter SC and select the control result with the highest score. If the obtained score is lower than SC, then the control data is increased or decreased by 10% based on the control of this parameter.

[0089] S6.3 If the obtained score is higher than SC, then the parameter is considered the optimal result, and the adjustment parameter is a local optimum. In order to explore the optimal result, a random number between 0 and 1 is used to determine whether to conduct further testing and adjustment based on this result.

[0090] If the random number is in the range of 0-0.05, increase it by 5%.

[0091] When the random number is in the range of 0.05-0.1, decrease it by 5%.

[0092] When the random number is in the range of 0.1-1, it should be adjusted according to the local optimum result.

[0093] S6.4. Obtain the specific adjustment parameters for all target labels obtained, based on S6.1, S6.2, and S6.3.

[0094] This invention discloses a knowledge graph-based optimized control system for magnetic separation processes, characterized by comprising sensor devices, a processor, and a memory; the sensors include level sensors, concentration sensors, and flow sensors, etc., used to collect real-time process data and send it to the processor; the memory stores a program, and the processor reads the program and executes the above-described method steps to achieve optimized control of the magnetic separation process.

[0095] In summary, this invention overcomes the current production challenges posed by the complex dynamic characteristics of magnetic separation processes, such as strong nonlinearity, strong coupling, and time-varying nature. It utilizes knowledge graph methods to construct a knowledge graph reasoning model based on magnetic separation optimization control, combines reinforcement learning strategies to learn from existing control decision-making experience, and simultaneously explores unknown strategies using the model. This satisfies the component and field strength optimization control of the magnetic separation process, and realizes a closed-loop model and optimized control of mineral processing indicators for multiple sets of magnetic separators.

[0096] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1.A method for optimizing control of magnetic separation based on a knowledge graph, characterized in that, It comprises the following steps: S1, real-time collection of relevant data of the magnetic separation process of multiple groups of magnetic separators, mainly including the opening size of the ore feeding rubber plug of the magnetic separator, the opening size of the tailings valve, the excitation state, the number of open stations, the concentrate grade of the first-stage high-intensity magnetic field, the concentrate grade of the mixed magnetic field, the flow rate of the large well bottom flow pump before the high-intensity magnetic field, the frequency of the large well bottom flow pump before the high-intensity magnetic field, the concentration of the large well bottom flow pump before the high-intensity magnetic field, and the setting of the upper and lower limits of the relevant data in production; S2, cleaning of the real-time collected data, removing repeated data and data exceeding the upper and lower limits of production, and regarding the same data of the concentrate grade of the mixed magnetic field, the concentrate grade of the first-stage high-intensity magnetic field, and the concentrate grade of the mixed magnetic field for three consecutive collection periods as repeated data, and regarding the same data of the flow rate of the large well bottom flow pump before the high-intensity magnetic field and the concentration of the large well bottom flow pump before the high-intensity magnetic field for two consecutive collection periods as repeated data; regarding the data of the opening size of the ore feeding rubber plug and the opening size of the tailings valve as set and feedback data, regarding the difference of more than 10% as abnormal data for cleaning; performing Kalman filtering on the opening size of the ore feeding rubber plug, the opening size of the tailings valve, the excitation state, the number of open stations, the concentrate grade of the first-stage high-intensity magnetic field, and the concentrate grade of the mixed magnetic field of all magnetic separators within one hour to realize the stabilization of the time data characteristics; merging the cleaned real-time data and the historical data within one hour into an adjustment target set; saving the historical control records for two months as a historical control set; S3, extracting triplets from the historical data, detecting whether the opening size of the ore feeding rubber plug, the number of open stations of the magnetic separator, the frequency of the large well bottom flow pump before the high-intensity magnetic field, and the magnetic field data of the magnetic field change at the same time and extracting them to constitute the knowledge graph training data set of the magnetic separation optimization control when the change amplitude of the magnetic separator liquid level, the opening size of the tailings valve of the magnetic separator, the magnetic field of the magnetic field, the concentrate grade of the first-stage high-intensity magnetic field, the concentrate grade of the mixed magnetic field, and the concentration and flow rate monitoring data of the large well bottom flow pump before the high-intensity magnetic field within 10 seconds exceeds 2%; S4, taking the triplet set in the knowledge graph training data set constituted by S3 as input, modeling the triplet set, and thereby constructing a knowledge graph reasoning model based on the magnetic separation optimization control; S5, finding the target label that needs to be controlled through the adjustment target set obtained in S2; S6, selecting the optimal control parameters of the target label according to the historical control set in S2. 2.The knowledge graph-based magnetic separation optimization control method according to claim 1, characterized in that The target label includes the opening size of the ore feeding rubber plug, the number of open stations of the magnetic separator, the frequency of the large well bottom flow pump before the high-intensity magnetic field, and the magnetic field of the magnetic field. The relevant collection label includes the opening size of the ore feeding rubber plug of the magnetic separator, the opening size of the tailings valve, the excitation state, the number of open stations, the concentrate grade of the first-stage high-intensity magnetic field, the concentrate grade of the mixed magnetic field, the flow rate of the large well bottom flow pump before the high-intensity magnetic field, the frequency of the large well bottom flow pump before the high-intensity magnetic field, and the concentration of the large well bottom flow pump before the high-intensity magnetic field. 3.The knowledge graph-based magnetic separation optimization control method according to claim 1, characterized in that The S5 finds the target label that needs to be controlled, which comprises the following steps: S5.1, obtaining the gradient according to the historical data in the adjustment target set and setting the parameter k; S5.2, for any triplet (h, r, t), obtaining a new scoring function according to the established knowledge graph reasoning model and the parameter k set by S5.1: ， wherein, is a squared norm, is a scoring function for triples, a head entity projection vector, a tail entity projection vector; At this time, the set S is obtained, wherein ， selecting the tail entity with a score greater than 0 and sorting them from large to small according to the score; S5.3, in the historical data, the same operation is obtained as the adjustment target set, while selecting the number of adjustment target labels num, at this time, the first num tail entities in the set S are respectively taken as t 1, , t 2, , t 3, …t num . 4.The knowledge graph-based magnetic separation optimization control method according to claim 1, characterized in that The S6 includes the following steps of selecting the optimal regulation parameter of the target label according to the historical regulation set in S2: S6.1, obtaining a regulation record set similar to the current production state from the historical regulation set, and selecting the optimal regulation record according to the magnetic field quality position, the magnetic separator liquid level and the magnetic separator tailing valve opening degree; if there are multiple records, taking the average; ， wherein S hcj , S fj、 , and S yw are the scores given by the mixed magnetic concentrate level, the magnetic separator liquid level, and the opening size of the magnetic separator tailings valve, respectively. If the relative results of the mixed magnetic concentrate level, the magnetic separator liquid level, and the opening size of the magnetic separator tailings valve are good, the corresponding S hcj , S fj , and S yw are 1; otherwise, S hcj , S fj , and S yw are 0; S1, S2, and S3 are the weights of the mixed magnetic concentrate level, the magnetic separator liquid level, and the opening degree of the magnetic separator tailings valve, respectively, and the weights ultimately add up to 1; if there is no similar regulation record, then 1% regulation is performed according to the regulation direction; S6.2, setting a parameter SC, and selecting the regulation result with the highest score; if the obtained score is lower than SC, then adding or reducing 10% to the regulation data on the basis of the SC parameter; S6.3, if the obtained score is higher than SC, then the optimal result is the local optimal solution; in order to explore the optimal result, a 0-1 random number is used to judge whether to test and adjust on this basis, when the random number is in the range of 0-0.05, increase by 5%; when the random number is in the range of 0.05-0.1, decrease by 5%; when the random number is in the range of 0.1-1, adjust according to the local optimal result; S6.4, obtaining all target labels, and obtaining the specific parameters of the adjustment according to S6.1, S6.2 and S6.3.

Images