Method and system for intelligently determining a condition of a water gap of a large package

Abstract

The application discloses a method and system for intelligently judging a state of a ladle nozzle and belongs to the technical field of metallurgical digitization. The state of the ladle nozzle is analyzed from multiple dimensions, not only the real-time monitoring of the ladle nozzle is realized, but also the state of the ladle nozzle can be more comprehensively analyzed, so that the abnormal state of the ladle nozzle can be found in time by workers, and property loss or safety accidents caused by the abnormal state of the ladle nozzle can be avoided.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of metallurgical digitization, and particularly relates to a method and system for intelligently determining the state of a ladle nozzle. BACKGROUND

[0002] The slide plate control of the continuous casting ladle nozzle generally does not have position detection, so it is difficult to understand the opening degree of the slide plate in production. With the progress of the production process, users have higher and higher requirements for the detection of the position of the slide plate, but it is difficult to modify the equipment that has been put into use. In the traditional slab continuous casting production process, the state of the ladle nozzle is not monitored, so it is difficult to discover abnormal states of the ladle nozzle in time. SUMMARY

[0003] In view of the above problems in the prior art, the application provides a method and system for intelligently determining the state of a ladle nozzle to solve the problems in the prior art.

[0004] In order to achieve the above-mentioned application purposes, the technical scheme adopted by the application is as follows: a method for intelligently determining the state of a ladle nozzle, comprising:

[0005] obtaining a first real-time monitoring image frame, wherein the first real-time monitoring image frame represents a photographic image of a ladle nozzle slide plate captured by a first photographic device;

[0006] adopting a first visual recognition model to perform abnormal identification on the first real-time monitoring image frame to obtain a first abnormal identification result, wherein the first abnormal identification result includes that the ladle nozzle slide plate has an abnormality or the ladle nozzle slide plate has no abnormality;

[0007] obtaining a second real-time monitoring image frame, wherein the second real-time monitoring image frame represents a photographic image of a position where a ladle bottom surface and the ladle nozzle are connected, which is captured by a second photographic device;

[0008] adopting a second visual recognition model to perform abnormal identification on the second real-time monitoring image frame to obtain a second abnormal identification result, wherein the second abnormal identification result includes that the position where the ladle bottom surface and the ladle nozzle are connected has an abnormality or the position where the ladle bottom surface and the ladle nozzle are connected has no abnormality;

[0009] obtaining a third real-time monitoring image frame, wherein the third real-time monitoring image frame is used to represent a thermal imaging image of the ladle nozzle captured by a thermal imaging device;

[0010] identifying the temperature in the third real-time monitoring image frame to obtain a plurality of temperature identification results, and performing abnormal identification on the temperature in the plurality of temperature identification results to obtain a third abnormal identification result, wherein the third abnormal identification result includes that there is a temperature identification result exceeding a set temperature threshold or there is no temperature identification result exceeding the set temperature threshold;

[0011] The first abnormality recognition result, the second abnormality recognition result, and the third abnormality recognition result are collectively taken as a large package water gap state monitoring result.

[0012] In a possible real-time manner, the first real-time monitoring image frame is subjected to abnormality recognition by using the first visual recognition model to obtain a first abnormality recognition result, including:

[0013] The first visual recognition model is constructed by using a neural network model.

[0014] A first training set is acquired, the first training set including a plurality of first historical monitoring image frames and an abnormality state corresponding to each first historical monitoring image frame, the first historical monitoring image frame representing a historical photographic image of a large package water gap slide plate captured by a first photographic device, and the abnormality state corresponding to the first historical monitoring image frame being used to represent that the large package water gap slide plate has an abnormality or does not have an abnormality in the first historical monitoring image frame.

[0015] The first visual recognition model is trained based on the first training set to obtain a trained first visual recognition model.

[0016] The first real-time monitoring image frame is taken as input data of the trained first visual recognition model, output data of the first visual recognition model is acquired, and the output data of the first visual recognition model is taken as the first abnormality recognition result.

[0017] In a possible real-time manner, the first visual recognition model is trained based on the first training set to obtain a trained first visual recognition model, including:

[0018] A1, an iteration round counter t is initialized to 1;

[0019] A2, the parameter quantity of the first visual recognition model is taken as the dimension of a first individual, and the parameters of the first visual recognition model are initialized multiple times to obtain a plurality of first individuals The i-th first individual at the t-th iteration round is represented, i=1, 2,..., I, I representing a total number of first individuals, The j-th parameter in the i-th first individual at the t-th iteration round is represented, j=1, 2,..., J, J representing a parameter quantity of the first visual recognition model.

[0020] A3, according to the first training set, a fitness corresponding to each first individual is determined, a first individual with the largest fitness is selected as an optimal first individual, and other first individuals are updated according to the optimal first individual to obtain a plurality of updated first individuals;

[0021] A4, for each first individual except the optimal first individual Generate the first mutant body And determine the first body Is the corresponding fitness greater than or equal to that of the first mutant? If the fitness level is high, then the first individual remains. Otherwise, the first entity will mutate. As the first new body

[0022] A5. Determine the fitness of each first individual, and select the first individual with the highest fitness as the optimal first individual. Use the fitness of the optimal first individual as the optimal fitness for this iteration.

[0023] A6. Determine whether the optimal fitness has converged in multiple iterations. If yes, proceed to step A7; otherwise, increment the iteration count counter t by one and return to step A3.

[0024] A7. Use the parameters corresponding to the first optimal body as the initial network parameters of the first visual recognition model, and use the gradient descent method to train the initial network parameters of the first visual recognition model to obtain the trained first visual recognition model.

[0025] In one possible real-time approach, the parameters of the first visual recognition model are:

[0026]

[0027] Where rand(0,1) represents a random number between (0,1), x max express The upper limit, x min express The lower limit.

[0028] In one possible real-time approach, the fitness of the first individual is:

[0029]

[0030] Where F represents fitness, Y m D represents the actual output of the first visual recognition model corresponding to the m-th sample; m This represents the expected output of the first visual recognition model corresponding to the m-th sample, which is the abnormal state corresponding to the m-th first historical monitoring image frame in the first training set; m = 1, 2, ..., M, where M represents the total number of first historical monitoring image frames in the first training set.

[0031] In a possible real-time manner, the first individual with the maximum fitness is selected as the optimal individual, and when the optimal individual is used to update other individuals, the updated individuals are:

[0032]

[0033] wherein, represents the i-th individual in the t-th iteration, represents the updated represents the optimal individual, A represents a first coefficient matrix of 1*J, A 1×j = 2*r1-a(t), A 1×j represents an element in the j-th column of the first coefficient matrix, r1 represents a random number between 0 and 1, a(t) represents an adjustment amount in the t-th iteration, a(t) = βa(t-1), β represents an adjustment factor, a(t-1) represents an adjustment amount in the (t-1)-th iteration, C represents a second coefficient matrix of 1*J, C = 2*r2, r2 represents a random number between 0 and 1, b represents a parameter update factor, l represents a random number between -1 and 1, π represents a circular constant, and ζ represents a random number between -1 and 1.

[0034] In a possible real-time manner, for each individual other than the optimal individual a mutated individual is generated , which includes: generating a random number R between 0 and 1, determining whether the random number R is greater than a set mutation probability, if yes, generating an individual corresponding to the mutated individual is otherwise, directly taking the individual as the mutated individual wherein, and represents three random individuals in all individuals, R1, R2 and R3 belong to [1, I], and f represents a scaling factor.

[0035] In a possible real-time manner, the parameters corresponding to the optimal individual are taken as initial network parameters of the first visual recognition model, the initial network parameters of the first visual recognition model are trained by using a gradient descent method, and a trained first visual recognition model is obtained, which includes:

[0036] B1, the parameters corresponding to the optimal individual are taken as initial network parameters x1, x2,...., x j ,...., x J of the first visual recognition model; wherein x ja jth parameter of the first visual recognition model, j = 1, 2,..., J, J representing a quantity of parameters of the first visual recognition model;

[0037] B2, taking the first historical monitoring image frame of the first training set as input data of the first visual recognition model, taking the abnormal state corresponding to the first historical monitoring image frame as expected output data of the first visual recognition model, and obtaining an error value of the first visual recognition model as:

[0038]

[0039] wherein E represents the error value, Y m represents actual output corresponding to the first visual recognition model corresponding to the mth sample; D m represents expected output corresponding to the first visual recognition model corresponding to the mth sample, that is, an abnormal state corresponding to the mth first historical monitoring image frame in the first training set; m = 1, 2,..., M, M representing a total number of first historical monitoring image frames in the first training set;

[0040] B3, judging whether the error value E is less than a set threshold value, if yes, taking the current network parameter as a final network parameter of the first visual recognition model, and obtaining the trained first visual recognition model, otherwise, entering step B4;

[0041] B4, according to the error value E, obtaining an update amount of a jth parameter x j of the first visual recognition model as and updating the jth parameter x j according to the update amount; wherein η represents a learning rate.

[0042] B5, returning to step B2.

[0043] In a possible real-time mode, a second visual recognition model is used to perform abnormal recognition on a second real-time monitoring image frame to obtain a second abnormal recognition result, including:

[0044] A neural network model is used to construct the second visual recognition model;

[0045] A second training set is obtained, the second training set including a plurality of second historical monitoring image frames and an abnormal state corresponding to each second historical monitoring image frame, the second historical monitoring image frame representing a historical photographic image of a position where a ladle bottom surface and a tundish water gap are connected, which is photographed by a second photographic device, and the abnormal state corresponding to the second historical monitoring image frame being used to represent that there is an abnormality or there is no abnormality in the position where the ladle bottom surface and the tundish water gap are connected in the second historical monitoring image frame;

[0046] Train the second visual recognition model based on the second training set to obtain the trained second visual recognition model;

[0047] Take the second real-time monitoring image frame as input data of the trained second visual recognition model, obtain output data of the second visual recognition model, and take the output data of the second visual recognition model as the second abnormality recognition result.

[0048] In a second aspect, the present application provides a system for intelligently determining the state of a ladle nozzle, comprising a first data acquisition module, a first recognition module, a second data acquisition module, a second recognition module, a third data acquisition module, a third recognition module, and a ladle nozzle state monitoring result acquisition module.

[0049] The first data acquisition module is configured to acquire a first real-time monitoring image frame, which represents a photographic image of a ladle nozzle slide plate captured by a first photographic device.

[0050] The first recognition module is configured to use a first visual recognition model to perform abnormality recognition on the first real-time monitoring image frame to obtain a first abnormality recognition result, which includes the presence or absence of abnormalities in the ladle nozzle slide plate.

[0051] The second data acquisition module is configured to acquire a second real-time monitoring image frame, which represents a photographic image of the connection position between the ladle bottom surface and the ladle nozzle captured by a second photographic device.

[0052] The second recognition module is configured to use a second visual recognition model to perform abnormality recognition on the second real-time monitoring image frame to obtain a second abnormality recognition result, which includes the presence or absence of abnormalities in the connection position between the ladle bottom surface and the ladle nozzle.

[0053] The third data acquisition module is configured to acquire a third real-time monitoring image frame, which represents a thermal imaging image of the ladle nozzle captured by a thermal imaging device.

[0054] The third data acquisition module is configured to identify the temperature in the third real-time monitoring image frame to obtain a plurality of temperature recognition results, and perform abnormality recognition on the temperatures in the plurality of temperature recognition results to obtain a third abnormality recognition result, which includes the presence or absence of temperature recognition results exceeding a set temperature threshold.

[0055] The ladle nozzle state monitoring result acquisition module is configured to take the first abnormality recognition result, the second abnormality recognition result, and the third abnormality recognition result together as the ladle nozzle state monitoring result.

[0056] The application provides a method and system for intelligently determining the state of a large ladle nozzle, which analyzes the state of the large ladle nozzle from multiple dimensions, realizes real-time monitoring of the large ladle nozzle, can more comprehensively analyze the state of the large ladle nozzle, and enables workers to timely find abnormal states of the large ladle nozzle, thereby avoiding property losses or safety accidents caused by abnormal large ladle nozzles. BRIEF DESCRIPTION OF DRAWINGS

[0057] The accompanying drawings, which are incorporated herein and constitute part of the specification, illustrate embodiments consistent with the application and, together with the description, serve to explain the principles of the application.

[0058] Figure 1 A flowchart of the method for intelligently determining the state of a large ladle nozzle provided by the application.

[0059] Figure 2 A structural diagram of the system for intelligently determining the state of a large ladle nozzle provided by the application.

[0060] 1 - first data acquisition module, 2 - first identification module, 3 - second data acquisition module, 4 - second identification module, 5 - third data acquisition module, 6 - third identification module, 7 - large ladle nozzle state monitoring result acquisition module.

[0061] The specific embodiments of the application have been described above with reference to the accompanying drawings, and will be described in more detail below. These drawings and the written description are not intended to limit the scope of the inventive concept in any way, but to illustrate the inventive concept to those skilled in the art by reference to specific embodiments. DETAILED DESCRIPTION

[0062] The specific embodiments of the application will be described below to facilitate understanding of the application by those skilled in the art, but it should be clear that the application is not limited to the scope of the specific embodiments, and that any changes that are obvious to those skilled in the art within the spirit and scope of the application as defined by the appended claims are obvious and are included in the protection of the inventive concept.

[0063] Embodiments of the application will be described in detail below with reference to the accompanying drawings.

[0064] Embodiment 1

[0065] As shown in Figure 1 A method for intelligently determining the state of a large ladle nozzle, comprising:

[0066] S1, acquiring a first real-time monitoring image frame, the first real-time monitoring image frame representing a photographic image of a large ladle nozzle slide plate taken by a first photographic device.

[0067] The installation position of the first camera should not be affected by the production of the continuous casting machine, and the photographic image for image recognition can be taken. In order to reduce the data processing amount, a shooting interval time can be set, and every time a shooting interval time passes, a first real-time monitoring image frame is taken for monitoring the state of the tundish nozzle slide plate.

[0068] S2, using the first visual recognition model to perform abnormality recognition on the first real-time monitoring image frame to obtain a first abnormality recognition result, the first abnormality recognition result including that the tundish nozzle slide plate is abnormal or the tundish nozzle slide plate is not abnormal.

[0069] The opening value of the slide plate can represent whether the slide plate is abnormal. When the opening value of the slide plate exceeds the set threshold value, it can be determined that the slide plate is abnormal. Therefore, a plurality of first historical monitoring image frames when the opening value of the slide plate is normal and the state corresponding to each first historical monitoring image frame can be used as first positive data, and a plurality of first historical monitoring image frames when the opening value of the slide plate is abnormal and the state corresponding to each first historical monitoring image frame can be used as second negative data. The first visual recognition model is trained using the first positive data and the second negative data as training data, and the trained first visual recognition model is used to recognize the first real-time monitoring image frame, so as to determine the state corresponding to the first real-time monitoring image frame. That is, by training the first visual recognition model, the first visual recognition model has the ability to identify whether the opening value of the slide plate in the first real-time monitoring image frame is abnormal, so as to obtain the first abnormality recognition result. The first historical monitoring image frame represents the historical image of the tundish nozzle slide plate, and the state corresponding to the first historical monitoring image frame represents the abnormal state of the tundish nozzle slide plate marked by the artificial, and the abnormal state includes abnormality or not abnormality.

[0070] S3, obtaining a second real-time monitoring image frame, the second real-time monitoring image frame representing a photographic image of the connection position between the bottom surface of the ladle and the tundish nozzle taken by a second camera.

[0071] The installation position of the second camera should not be affected by the production of the continuous casting machine, and the photographic image for image recognition can be taken. In order to reduce the data processing amount, a shooting interval time can be set, and every time a shooting interval time passes, a second real-time monitoring image frame is taken for monitoring the state of the tundish nozzle.

[0072] S4, using the second visual recognition model to perform abnormality recognition on the second real-time monitoring image frame to obtain a second abnormality recognition result, the second abnormality recognition result including that the connection position between the bottom surface of the ladle and the tundish nozzle is abnormal or the connection position between the bottom surface of the ladle and the tundish nozzle is not abnormal.

[0073] The perpendicularity between the ladle bottom surface and the ladle nozzle can represent whether there is an abnormality at the connection position between the two, and when the perpendicularity between the ladle bottom surface and the ladle nozzle exceeds a preset threshold range, it can be considered that there is an abnormality between the ladle bottom surface and the ladle nozzle. Therefore, a plurality of second historical monitoring image frames when the perpendicularity is normal and the state corresponding to each second historical monitoring image frame can be used as positive data, and a plurality of second historical monitoring image frames when the perpendicularity is abnormal and the state corresponding to each second historical monitoring image frame can be used as negative data; the positive data and the negative data are used as training data to train the second visual recognition model, and the trained second visual recognition model is used to recognize the second real-time monitoring image frame, so as to determine the state corresponding to the second real-time monitoring image frame, that is, by training the second visual recognition model, the second visual recognition model has the ability to identify whether the perpendicularity between the ladle bottom surface and the ladle nozzle in the second real-time monitoring image frame is abnormal, so as to obtain the second abnormality recognition result. Wherein, the second historical monitoring image frame represents a historical image of the connection position between the ladle bottom surface and the ladle nozzle, and the state corresponding to the second historical monitoring image frame represents the abnormal state of the perpendicularity between the ladle bottom surface and the ladle nozzle marked by the artificial, and the abnormal state includes abnormal or not abnormal.

[0074] S5, a third real-time monitoring image frame is obtained, the third real-time monitoring image frame is used to represent a thermal imaging image of the ladle nozzle captured by a thermal imaging device.

[0075] S6, the temperature in the third real-time monitoring image frame is identified to obtain a plurality of temperature identification results, and the temperature in the plurality of temperature identification results is abnormally identified to obtain a third abnormality recognition result, the third abnormality recognition result includes that there is a temperature identification result exceeding a set temperature threshold or there is no temperature identification result exceeding the set temperature threshold.

[0076] By detecting the temperature distribution of the visual area, when the temperature of a certain place is too high, it may be that the equipment has a problem, causing the risk of molten steel spilling out, causing danger to the equipment and personnel on site, and timely warning.

[0077] S7, the first abnormality recognition result, the second abnormality recognition result and the third abnormality recognition result are collectively used as a ladle nozzle state monitoring result.

[0078] Optionally, when any one of the first abnormality identification result, the second abnormality identification result and the second abnormality identification result exists an abnormality, a sound and light alarm is sent to enable the staff to timely handle the abnormality of the water gap of the large bag. A risk level can also be set, the more abnormal results in the first abnormality identification result, the second abnormality identification result and the second abnormality identification result, the higher the risk level; there is no abnormal result in the first abnormality identification result, the second abnormality identification result and the second abnormality identification result, then there is no abnormality; so that the staff can know the severity of the abnormality in the first time.

[0079] In a possible real-time manner, a first visual identification model is used to perform abnormality identification on a first real-time monitoring image frame to obtain a first abnormality identification result, including:

[0080] The first visual identification model is constructed by using a neural network model.

[0081] Optionally, the neural network model can be a recurrent neural network (RNN), a deep residual network (ResNet) or a back propagation (BP) neural network. It should be noted that in addition to the above neural network models, other neural network models with image classification capability can also be used as the first visual identification model.

[0082] A first training set is obtained, the first training set including a plurality of first historical monitoring image frames and an abnormality state corresponding to each first historical monitoring image frame, the first historical monitoring image frame representing a historical photographic image of the water gap slide of the large bag captured by the first photographic device, and the abnormality state corresponding to the first historical monitoring image frame being used to represent that the water gap slide of the large bag in the first historical monitoring image frame exists an abnormality or does not exist an abnormality.

[0083] In order to ensure the training effect of the first visual identification model, the number of the first historical monitoring image frames should be as large as possible, and the opening value of the slide corresponding to the first historical monitoring image frames should be uniformly distributed, so that the trained first visual identification model can accurately identify the first real-time monitoring image frame.

[0084] Based on the first training set, the first visual identification model is trained to obtain a trained first visual identification model.

[0085] The first real-time monitoring image frame is used as input data of the trained first visual identification model, output data of the first visual identification model is obtained, and the output data of the first visual identification model is used as the first abnormality identification result.

[0086] In a possible real-time manner, the first visual recognition model is trained based on the first training set, and a trained first visual recognition model is obtained, including:

[0087] A1, initialize an iteration round counter t = 1.

[0088] A2, taking the parameter quantity of the first visual recognition model as the dimension of the first individual, initializing the parameters of the first visual recognition model multiple times to obtain multiple first individuals denotes the i-th first individual at the t-th iteration, i = 1, 2,..., I, I represents the total number of first individuals, denotes the j-th parameter in the i-th first individual at the t-th iteration, j = 1, 2,..., J, J represents the parameter quantity of the first visual recognition model.

[0089] A3, according to the first training set, determine the fitness corresponding to each first individual, and select the first individual with the maximum fitness as the optimal first individual, and update other first individuals according to the optimal first individual to obtain updated multiple first individuals.

[0090] A4, for each first individual except the optimal first individual generate a mutated first individual and judge whether the fitness corresponding to the first individual is greater than or equal to the fitness of the mutated first individual , if yes, keep the first individual unchanged, otherwise, take the mutated first individual as a new first individual

[0091] A5, determine the fitness corresponding to each first individual, and select the first individual with the maximum fitness as the optimal first individual, and take the fitness of the optimal first individual as the optimal fitness corresponding to the current iteration.

[0092] A6, judge whether the optimal fitness in multiple iterations converges, if yes, go to step A7, otherwise, increase the count value of the iteration round counter t by one, and return to step A3.

[0093] A7, take the parameters corresponding to the optimal first individual as the initial network parameters of the first visual recognition model, train the initial network parameters of the first visual recognition model by using the gradient descent method, and obtain the trained first visual recognition model.

[0094] The training method provided in this embodiment can first roughly find the location of the global optimum, and then use gradient descent to further train, thereby further optimizing the network parameters and making the network parameters the global optimum, thus ensuring the recognition effect of the first visual recognition model.

[0095] In one possible real-time approach, the parameters of the first visual recognition model are:

[0096]

[0097] Where rand(0,1) represents a random number between (0,1), x max express The upper limit, x min express The lower limit.

[0098] In one possible real-time approach, the fitness of the first individual is:

[0099]

[0100] Where F represents fitness, Y m D represents the actual output of the first visual recognition model corresponding to the m-th sample. m This represents the expected output of the first visual recognition model corresponding to the m-th sample, which is the abnormal state corresponding to the m-th first historical monitoring image frame in the first training set. m = 1, 2, ..., M, where M represents the total number of first historical monitoring image frames in the first training set.

[0101] In one possible real-time approach, when the first individual with the highest fitness is selected as the optimal first individual, and other first individuals are updated based on the optimal first individual, the updated first individual is:

[0102]

[0103] in, This represents the first body of the i-th iteration in the t-th iteration. Indicates the updated Let A represent the first optimal volume, and let A represent the first coefficient matrix of 1×J. 1×j =2*r1-a(t), A 1×jdenotes an element of the jth column in the first coefficient matrix, r1 denotes a random number between (0, 1), a(t) denotes an adjustment amount at the tth iteration, a(t) = βa(t-1), β denotes an adjustment factor, a(t-1) denotes an adjustment amount at the (t-1)th iteration, C denotes a 1xJ second coefficient matrix, C = 2*r2, r2 denotes a random number between (0, 1), b denotes a parameter update factor, l denotes a random number between [-1, 1], π denotes a circular constant, and ζ denotes a random number between [-1, 1].

[0104] In a possible real-time manner, for each individual except the optimal first individual generating a mutated individual comprises: generating a random number R between (0, 1), determining whether the random number R is greater than a set mutation probability, if yes, generating an individual a corresponding mutated individual is otherwise directly taking the individual as a mutated individual wherein, and denote random three individuals among all individuals, R1, R2, and R3 all belong to [1, I], and f denotes a scaling factor.

[0105] In a possible real-time manner, parameters corresponding to the optimal individual are taken as initial network parameters of the first visual recognition model, the initial network parameters of the first visual recognition model are trained by using a gradient descent method, and a trained first visual recognition model is obtained, comprising the following steps.

[0106] B1. Taking the parameters corresponding to the optimal individual as initial network parameters x1, x2,..., x j ,..., x J of the first visual recognition model; wherein x j denotes the jth parameter of the first visual recognition model, j = 1, 2,..., J, and J denotes the parameter quantity of the first visual recognition model.

[0107] B2. Taking the first historical monitoring image frame of the first training set as input data of the first visual recognition model, taking the abnormal state corresponding to the first historical monitoring image frame as expected output data of the first visual recognition model, and obtaining an error value of the first visual recognition model as follows:

[0108]

[0109] wherein E denotes the error value, Y m denotes actual output corresponding to the first visual recognition model corresponding to the mth sample, and D mrepresents the expected output corresponding to the first visual recognition model corresponding to the mth sample, that is, the abnormal state corresponding to the mth first historical monitoring image frame in the first training set. m = 1, 2, …, M, and M represents the total number of first historical monitoring image frames in the first training set.

[0110] B3, determining whether the error value E is less than a set threshold value, if yes, the current network parameter is taken as the final network parameter of the first visual recognition model, and a trained first visual recognition model is obtained, otherwise, step B4 is entered.

[0111] B4, according to the error value E, the update amount of the jth parameter x j of the first visual recognition model is obtained as and the jth parameter x j of the first visual recognition model is updated according to the update amount. Wherein, η represents the learning rate.

[0112] B5, returning to step B2.

[0113] In a possible real-time mode, the second visual recognition model is used to perform abnormal identification on the second real-time monitoring image frame to obtain a second abnormal identification result, including:

[0114] The neural network model is used to construct the second visual recognition model.

[0115] Optionally, the neural network model can be a recurrent neural network (RNN), a deep residual network (ResNet), or a BP (Back Propagation) neural network. It is worth noting that in addition to the above neural network models, other neural network models with image classification capability can also be used as the first visual recognition model.

[0116] A second training set is obtained, the second training set including a plurality of second historical monitoring image frames and an abnormal state corresponding to each second historical monitoring image frame, the second historical monitoring image frame representing a historical photographic image of the connection position between the ladle bottom surface and the ladle nozzle taken by a second photographic device, and the abnormal state corresponding to the second historical monitoring image frame being used to represent whether there is an abnormality in the connection position between the ladle bottom surface and the ladle nozzle in the second historical monitoring image frame or whether there is no abnormality in the connection position between the ladle bottom surface and the ladle nozzle.

[0117] In order to ensure the training effect of the second visual recognition model, the number of second historical monitoring image frames should be as large as possible, and the corresponding perpendicularity of the second historical monitoring image frames should be uniformly distributed, so that the trained second visual recognition model can accurately identify the second real-time monitoring image frame.

[0118] Train the second visual recognition model based on the second training set to obtain the trained second visual recognition model.

[0119] Take the second real-time monitoring image frame as input data of the trained second visual recognition model, obtain output data of the second visual recognition model, and take the output data of the second visual recognition model as the second abnormality recognition result.

[0120] Optionally, the second visual recognition model is trained based on the second training set to obtain the trained second visual recognition model, including:

[0121] C1, initialize an iteration round counter T = 1.

[0122] C2, take the parameter quantity of the second visual recognition model as the dimension of the second individual, initialize the parameters of the second visual recognition model multiple times to obtain multiple second individuals denotes the i-th second individual at the T-th iteration, i = 1, 2,..., I, I represents the total number of second individuals, denotes the j-th parameter in the i-th second individual at the t-th iteration, j = 1, 2,..., J, J represents the parameter quantity of the second visual recognition model.

[0123] C3, determine the fitness corresponding to each second individual according to the second training set, select the second individual with the maximum fitness as the optimal second individual, and update other second individuals according to the optimal second individual to obtain updated multiple second individuals.

[0124] C4, for each second individual except the optimal second individual generate a mutated second individual and judge whether the fitness corresponding to the second individual is greater than or equal to the fitness of the mutated second individual , if yes, keep the second individual unchanged, otherwise take the mutated second individual as a new second individual

[0125] C5, determine the fitness corresponding to each second individual, select the second individual with the maximum fitness as the optimal second individual, and take the fitness of the optimal second individual as the optimal fitness corresponding to the current iteration.

[0126] C6, judge whether the optimal fitness converges in multiple iterations, if yes, go to step C7, otherwise increase the count value of the iteration round counter T by one and return to step C3.

[0127] C7, taking the parameters corresponding to the optimal second individual as initial network parameters of the second visual recognition model, training the initial network parameters of the second visual recognition model by using a gradient descent method, and obtaining the trained second visual recognition model.

[0128] The formulas involved in the training of the second visual recognition model are the same as the formulas involved in the training of the first visual recognition model, which will not be described here.

[0129] The method and system for intelligently determining the state of the ladle nozzle provided in the embodiment analyze the state of the ladle nozzle from multiple dimensions, not only realize real-time monitoring of the ladle nozzle, but also more comprehensively analyze the state of the ladle nozzle, so that the staff can timely find the abnormal state of the ladle nozzle, and avoid property loss or safety accidents caused by the abnormality of the ladle nozzle.

[0130] Embodiment 2

[0131] As shown in Figure 2 The system for intelligently determining the state of the ladle nozzle provided in the embodiment includes a first data acquisition module 1, a first identification module 2, a second data acquisition module 3, a second identification module 4, a third data acquisition module 5, a third identification module 6, and a ladle nozzle state monitoring result acquisition module 7.

[0132] The first data acquisition module 1 is configured to acquire a first real-time monitoring image frame, which represents a photographic image of the ladle nozzle slide plate captured by a first photographic device.

[0133] The first identification module 2 is configured to perform abnormality identification on the first real-time monitoring image frame by using a first visual recognition model to obtain a first abnormality identification result, which includes whether the ladle nozzle slide plate is abnormal or not.

[0134] The second data acquisition module 3 is configured to acquire a second real-time monitoring image frame, which represents a photographic image of the connection position between the ladle bottom surface and the ladle nozzle captured by a second photographic device.

[0135] The second identification module 4 is configured to perform abnormality identification on the second real-time monitoring image frame by using a second visual recognition model to obtain a second abnormality identification result, which includes whether the connection position between the ladle bottom surface and the ladle nozzle is abnormal or not.

[0136] The third data acquisition module 5 is configured to acquire a third real-time monitoring image frame, which is used to represent a thermal imaging image of the ladle nozzle captured by a thermal imaging device.

[0137] The third data acquisition module 6 is configured to identify the temperature in the third real-time monitoring image frame to obtain a plurality of temperature identification results, and perform abnormality identification on the temperature in the plurality of temperature identification results to obtain a third abnormality identification result, the third abnormality identification result including whether there is a temperature identification result exceeding a set temperature threshold or whether there is no temperature identification result exceeding the set temperature threshold.

[0138] The ladle nozzle state monitoring result acquisition module 7 is configured to collectively take the first abnormality identification result, the second abnormality identification result and the third abnormality identification result as a ladle nozzle state monitoring result.

[0139] The system for intelligently determining the state of a ladle nozzle provided in the embodiment can execute the method technical solution as described in Embodiment 1, and has similar principles and beneficial effects, which will not be described one by one here.

[0140] It should be noted that any method using the concept of the present application should be within the scope of the present application. Other embodiments of the present application will be readily apparent to those skilled in the art upon considering the description herein and the disclosure of the application contained herein. The present application is intended to cover any variations, uses, or adaptations of the application following, in general, the principles of the application and including such variations, uses, or adaptations of the present application as come within the true scope of the application. It should be understood that the application is not limited to the precise structures as already described and illustrated in the drawings, and that various modifications and changes can be made without departing from the scope of the present application.

Claims

1. A method for intelligently determining the condition of a water gap of a large package, characterized in that, The method comprises the following steps: obtaining a first real-time monitoring image frame, the first real-time monitoring image frame representing a photographic image of a ladle slide gate taken by a first photographic device; performing abnormality identification on the first real-time monitoring image frame by using a first visual recognition model to obtain a first abnormality identification result, the first abnormality identification result including that the ladle slide gate is abnormal or the ladle slide gate is not abnormal; obtaining a second real-time monitoring image frame, the second real-time monitoring image frame representing a photographic image of a position where a ladle bottom surface and a ladle slide gate are connected taken by a second photographic device; performing abnormality identification on the second real-time monitoring image frame by using a second visual recognition model to obtain a second abnormality identification result, the second abnormality identification result including that the position where the ladle bottom surface and the ladle slide gate are connected is abnormal or the position where the ladle bottom surface and the ladle slide gate are connected is not abnormal; obtaining a third real-time monitoring image frame, the third real-time monitoring image frame representing a thermal imaging image of the ladle slide gate taken by a thermal imaging device; identifying temperatures in the third real-time monitoring image frame to obtain a plurality of temperature identification results, and performing abnormality identification on the temperatures in the plurality of temperature identification results to obtain a third abnormality identification result, the third abnormality identification result including that there is a temperature identification result exceeding a set temperature threshold or there is no temperature identification result exceeding the set temperature threshold; collecting the first abnormality identification result, the second abnormality identification result and the third abnormality identification result as a ladle slide gate state monitoring result.

2. The method of claim 1, wherein, The method comprises the following steps: performing abnormality identification on the first real-time monitoring image frame by using a first visual recognition model to obtain a first abnormality identification result, including: constructing the first visual recognition model by using a neural network model; obtaining a first training set, the first training set including a plurality of first historical monitoring image frames and an abnormality state corresponding to each first historical monitoring image frame, the first historical monitoring image frame representing a historical photographic image of a ladle slide gate taken by a first photographic device, and the abnormality state corresponding to the first historical monitoring image frame representing that the ladle slide gate in the first historical monitoring image frame is abnormal or the ladle slide gate in the first historical monitoring image frame is not abnormal; training the first visual recognition model based on the first training set to obtain a trained first visual recognition model; 3. The method of claim 1, wherein, taking the first real-time monitoring image frame as input data of the trained first visual recognition model, obtaining output data of the first visual recognition model, and taking the output data of the first visual recognition model as the first abnormality identification result. A1, initialize iteration round counter ; A2, taking the parameter quantity of the first visual recognition model as the dimension of the first individual, initializing the parameters of the first visual recognition model multiple times to obtain multiple first individuals , represents the total number of first individuals, represents the first individual in the i-th iteration, , represents the total number of first individuals, represents the i-th parameter in the j-th iteration of the i-th individual, represents the i-th parameter in the j-th iteration of the i-th individual, , represents the parameter quantity of the first visual recognition model;​​​ The method comprises the following steps: A4. For each first body except the optimal first body. Generate the first mutant body And determine the first body Is the corresponding fitness greater than or equal to that of the first mutant? If the fitness level is high, then the first individual remains. Otherwise, the first entity will mutate. As the first new body ; A3, determining an adaptability corresponding to each first individual according to the first training set, selecting a first individual with the largest adaptability as an optimal first individual, and updating other first individuals according to the optimal first individual to obtain updated first individuals; A5, determining an adaptability corresponding to each first individual, selecting a first individual with the largest adaptability as an optimal first individual, and taking the adaptability of the optimal first individual as an optimal adaptability corresponding to the current iteration. A6, judge whether the optimal fitness converges in the multi-round iteration, if yes, go to step A7, otherwise, let the count value of the iteration round counter be added by one and return to step A3; A7, taking the parameters corresponding to the optimal individual as initial network parameters of the first visual recognition model, training the initial network parameters of the first visual recognition model by using a gradient descent method, and obtaining the trained first visual recognition model.

4. The method of claim 3, wherein, The parameters of the first visual recognition model are: ， wherein, represents a random number between (0, 1), represents an upper limit of represents a lower limit of 5. The method of claim 3, wherein, The fitness of the first individual is: ， wherein, represents fitness, represents actual output corresponding to the first visual recognition model corresponding to the mth sample; represents expected output corresponding to the first visual recognition model corresponding to the mth sample, i.e. abnormal state corresponding to the mth first historical monitoring image frame in the first training set; , represents total number of first historical monitoring image frames in the first training set.

6. The method of claim 3, wherein, The first individual with the maximum fitness is selected as the optimal individual, and when the optimal individual is updated according to other individuals, the updated first individual is: ， wherein, denotes the denotes the denotes the denotes the updated denotes the , denotes the denotes a random number between (0,1), denotes the denotes the , denotes the denotes the denotes the denotes the , denotes a random number between (0,1), b denotes a parameter update factor, denotes a random number between [-1,1], denotes the denotes a random number between [-1,1].

7. The method of claim 3, wherein, For each individual except the optimal one , a mutated individual is generated, including: generating a random number R between (0, 1), judging whether the random number R is greater than a set mutation probability, if yes, generating an individual corresponding to the mutated individual is , otherwise directly taking the individual as the mutated individual ; wherein, , and represent random three individuals in all individuals, R1, R2 and R3 all belong to [1, I], and f represents a scaling factor.

8. The method of claim 3, wherein, The parameters corresponding to the optimal individual are taken as the initial network parameters of the first visual recognition model, the initial network parameters of the first visual recognition model are trained by using a gradient descent method, and the trained first visual recognition model is obtained, comprising: B1, taking the parameter corresponding to the optimal first individual as the initial network parameter of the first visual recognition model ; wherein, denotes the jth parameter of the first visual recognition model, , denotes the parameter quantity of the first visual recognition model; B2, taking the first historical monitoring image frame of the first training set as the input data of the first visual recognition model, taking the abnormal state corresponding to the first historical monitoring image frame as the expected output data of the first visual recognition model, and obtaining the error value of the first visual recognition model: ， wherein E represents an error value, represents an actual output corresponding to the first visual recognition model corresponding to the mth sample; represents an expected output corresponding to the first visual recognition model corresponding to the mth sample, that is, an abnormal state corresponding to the mth first historical monitoring image frame in the first training set; , represents the total number of first historical monitoring image frames in the first training set; B3, determining whether the error value E is less than a set threshold value, if yes, taking the current network parameters as the final network parameters of the first visual recognition model, and obtaining the trained first visual recognition model, otherwise, entering step B4; B4. Based on the error value E, obtain the j-th parameter of the first visual recognition model. The update volume is And adjust the j-th parameter according to the update amount. Update; among them, Indicates the learning rate; B5, returning to step B2.

9. The method of claim 1-8, wherein, The second visual recognition model is used to identify the abnormality of the second real-time monitoring image frame, and the second abnormality recognition result is obtained, comprising: The second visual recognition model is constructed by using a neural network model; A second training set is obtained, the second training set comprises a plurality of second historical monitoring image frames and an abnormal state corresponding to each second historical monitoring image frame, the second historical monitoring image frame represents a historical photographic image of the connection position of the ladle bottom surface and the ladle nozzle taken by a second photographic device, and the abnormal state corresponding to the second historical monitoring image frame is used to represent that the connection position of the ladle bottom surface and the ladle nozzle in the second historical monitoring image frame has an abnormality or the connection position of the ladle bottom surface and the ladle nozzle has no abnormality; The second visual recognition model is trained based on the second training set, and a trained second visual recognition model is obtained; The output data of the second visual recognition model is obtained by taking the second real-time monitoring image frame as the input data of the trained second visual recognition model, and the output data of the second visual recognition model is taken as the second abnormality recognition result.

10. A system for intelligently determining a state of a water nozzle of a large package, the system being capable of performing the method for intelligently determining a state of a water nozzle of a large package according to any one of claims 1 to 9, characterized in that The first data acquisition module, the first recognition module, the second data acquisition module, the second recognition module, the third data acquisition module, the third recognition module, and the ladle nozzle state monitoring result acquisition module are included. The first data acquisition module is used to obtain a first real-time monitoring image frame, and the first real-time monitoring image frame represents a photographic image of a ladle nozzle sliding plate taken by a first photographic device; The first recognition module is used to identify the abnormality of the first real-time monitoring image frame by using a first visual recognition model, and obtain a first abnormality recognition result, the first abnormality recognition result comprises that the ladle nozzle sliding plate has an abnormality or the ladle nozzle sliding plate has no abnormality; The second data acquisition module is configured to acquire a second real-time monitoring image frame, the second real-time monitoring image frame representing a photographic image of a position where a ladle bottom surface is connected with a tundish nozzle, captured by a second camera device; The second identification module is configured to perform abnormality identification on the second real-time monitoring image frame by using a second visual identification model, to obtain a second abnormality identification result, the second abnormality identification result including that the position where the ladle bottom surface is connected with the tundish nozzle is abnormal or that the position where the ladle bottom surface is connected with the tundish nozzle is normal; The third data acquisition module is configured to acquire a third real-time monitoring image frame, the third real-time monitoring image frame representing a thermal imaging image of the tundish nozzle, captured by a thermal imaging device; The third data acquisition module is configured to identify temperatures in the third real-time monitoring image frame, to obtain a plurality of temperature identification results, and to perform abnormality identification on the temperatures in the plurality of temperature identification results, to obtain a third abnormality identification result, the third abnormality identification result including that there is a temperature identification result exceeding a set temperature threshold or that there is no temperature identification result exceeding the set temperature threshold; The tundish nozzle state monitoring result acquisition module is configured to jointly use the first abnormality identification result, the second abnormality identification result, and the third abnormality identification result as a tundish nozzle state monitoring result.

Images