Magnesium furnace abnormal condition recognition method and device based on random configuration convolutional network

Abstract

The application discloses a magnesium furnace abnormal working condition recognition method and device based on a random configuration convolutional network, and the method comprises the following steps: acquiring a magnesium furnace image training data set; setting parameters of a convolutional network and initializing; configuring the Lth convolution parameter of the Layerth layer convolution of the convolutional network, and judging whether the parameter meets the configuration condition; if yes, a convolution parameter matrix is constructed; a feature map is generated by using the convolution parameter matrix; a pooling operation is performed on the feature map to generate a down-sampling feature map; the input data is updated by using the down-sampling feature map, the error between the predicted output O of the random configuration convolutional network and the true output T is calculated, if the error is greater than the expected error, the convolutional network parameter is continuously configured, if the error is less than or equal to the expected error, the convolutional network is constructed, and the magnesium furnace to be detected is subjected to abnormal working condition detection; the application has the advantages that the magnesium furnace abnormal working condition recognition accuracy is high, and the magnesium furnace abnormal working condition recognition is convenient to deploy on site.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent control of industrial magnesium smelting, and more particularly to a magnesium furnace abnormal condition recognition method and device based on a randomly configured convolutional network. BACKGROUND

[0002] Currently, the method for producing fused magnesia in an electric fused magnesia enterprise is mainly to use a three-phase alternating current fused magnesium furnace (referred to as an electric fused magnesium furnace) to heat and smelt powder-like raw materials mainly composed of magnesite ore through an electric arc. In the smelting process of the electric fused magnesium furnace, the active power consumed is mainly affected by the smelting current, and the smelting power of the electric fused magnesium furnace can be changed by adjusting the smelting current.

[0003] Due to the high temperature of the molten liquid of the super-high-temperature electric fused magnesium furnace, the fault data of the electric fused magnesium furnace cannot be directly measured. At present, most of the under-burning abnormal conditions are diagnosed by regular observation of the characteristics of the furnace shell by workers on site, and a small amount of research uses various mature algorithms for automatic monitoring. For example, Chinese Patent Publication No. CN115033724A discloses an electric fused magnesium furnace abnormal condition recognition method based on a multi-modal modulation algorithm, which comprises: collecting historical data of the electric fused magnesium furnace and performing data screening; normalizing the data samples; constructing a multi-modal modulation neural network MNN and initializing the network parameters; training the multi-modal modulation neural network MNN; identifying the abnormal condition of the electric fused magnesium furnace based on multi-modal understanding to obtain the identification result. The patent application can realize the diagnosis of the abnormal condition of the electric fused magnesium furnace, evaluate the health status of the electric fused magnesium furnace, and provide maintenance suggestions for the electric fused magnesium furnace, to ensure the safe, efficient and stable operation of the factory, thereby improving the accuracy of abnormal condition recognition and the stability of factory production. However, the patent application uses multi-modal fault monitoring, which relies on a large amount of historical process measurement data. Due to the performance degradation, faults and external disturbances of the on-site detection instruments and systems, the measurement data is not accurate, thereby affecting the accuracy of subsequent modeling, and the model is prone to local optimization. Therefore, the model constructed has low accuracy in recognizing the abnormal condition of the magnesium furnace. Moreover, the text feature sequence of the abnormal fault of the electric fused magnesium furnace required by the multi-modal network is not easy to obtain, and therefore it is difficult to deploy on site. SUMMARY

[0004] The technical problem to be solved by the present application is that the existing magnesium furnace abnormal condition recognition method has low accuracy, the text feature sequence of the abnormal fault of the electric fused magnesium furnace is not easy to obtain, and it is difficult to deploy on site.

[0005] The present application solves the above technical problems by the following technical means: a magnesium furnace abnormal condition recognition method based on a randomly configured convolutional network, which comprises:

[0006] Step a: obtaining a magnesium furnace image training data set;

[0007] Step b: setting the parameters of the convolutional network and initializing;

[0008] Step c: configure the Lth convolution of the Layerth layer of the convolutional network, judge whether the parameters meet the configuration condition, if yes, build the convolutional parameter matrix, if not, reconfigure the parameters until the configuration condition is met;

[0009] Step d: generate a feature map using the convolutional parameter matrix;

[0010] Step e: perform a pooling operation on the feature map to generate a down-sampled feature map;

[0011] Step f: update the input data using the down-sampled feature map, calculate the error between the predicted output O and the true output T of the randomly configured convolutional network, if the error is greater than the expected error, continue to configure the convolutional network parameters, if the error is less than or equal to the expected error, the convolutional network is completed and can be used for abnormal condition detection of the magnesium furnace.

[0012] Beneficial effects: the convolutional parameters of the convolutional network are configured, and the next step is performed only when the configuration condition is met, which ensures the reliability of the model parameters to a certain extent. The error between the predicted output O and the true output T of the convolutional network is calculated, and the convolutional network parameters are continued to be configured when the error exceeds the expected error, which avoids the prediction error of the constructed convolutional network being too large and improves the accuracy of the output results of the convolutional network, thereby improving the accuracy of the magnesium furnace abnormal condition identification. The convolutional parameter matrix is constructed based on the convolutional network, and the feature map is generated using the convolutional parameter matrix, thereby facilitating the deployment of the electric smelting magnesium furnace abnormal fault model in the magnesium furnace production site.

[0013] Further, the step a comprises:

[0014] The training data set input is X={x1, x2,..., xN}, and the output is T={t1, t2,..., tN}, N is the number of samples, xi is the i-th input sample, ti is the i-th output sample, and i=1, 2,..., N. N N i i

[0015] Further, the step b comprises:

[0016] Set the parameters of the convolutional network, the maximum number of convolutional layers Layer max , the maximum number of convolutions L max of each layer, the expected error ε, and the convolution kernel size k.

[0017] Initialize the convolutional network, set the number of convolutions L=0, the number of convolutional layers Layer=0, the input I=X, and the error

[0018] ​​​​Further, the step c comprises:

[0019] Step c01: clipping the input data I into input data I' of size k x k;

[0020] Step c02: randomly selecting the weight W L,Layer and bias b L,Layer using formula (1)

[0021]

[0022] wherein λ is a configuration fixed parameter, rand(a, b) is a random function of generating a matrix of size a x b with a uniform distribution on [0, 1), C in is the number of channels of the input data I;

[0023] Step c03: calculating h L,Layer using formula (2)

[0024] h L,Layer = g(W L,Layer ·I' + b L,Layer ) (2)

[0025] wherein g(·) is an activation function, and · is matrix multiplication;

[0026] Step c04: calculating the constraint ξ L,Layer using formula (3) L,Layer If ξ T < 0, the parameter does not meet the configuration condition, let r = r + τ and return to step c01 to reselect the convolution kernel parameter, otherwise continue step c05;

[0027]

[0028] wherein e L,Layer is the transpose of the error e, h T is the transpose of h L,Layer , r is a constraint parameter, 0 < r < 1, and τ is a configuration fixed parameter update step;

[0029] Step c05: constructing the weight parameter matrix W Layer = [W 1,Layer , W 2,Layer ,..., W L,Layer ] of the Layer layer, the bias parameter matrix b Layer = [b 1,Layer ,..., b L,Layer ], and the weight parameter matrix and the bias parameter matrix constitute a convolution parameter matrix.

[0030] Further, the step d comprises:

[0031] Using the convolution parameter matrix, the feature map F is generated by performing a convolution operation on the magnesium furnace image training dataset according to formula (4). L,Layer ;

[0032]

[0033] Among them, I i,k Let be the k-th channel matrix of the i-th input data, where i = 0, 1, ..., N, k = 1, 2, ..., C in , * is the cross-correlation operator.

[0034] Furthermore, step e includes:

[0035] The feature map F is processed using formula (5). L,Layer Perform pooling operations to generate downsampled feature maps F′ L,Layer ;

[0036]

[0037] Among them, F i L,Layer Let i = 1, ..., N be the feature sub-map of the i-th input data, h = 0, 1, ..., H-1, and H be the feature map F. L,Layer The number of rows, w = 0, 1, ..., W-1, where W is the feature map F. L,Layer The number of columns.

[0038] Furthermore, step f, which updates the input data using the downsampled feature map, includes:

[0039] Step f01: Update input data I to I g I g =[F′ 1,Layer ,...,F′ L,Layer ], and let L = L + 1;

[0040] Step f02: Calculate the output layer parameters β of the convolutional network using formula (6);

[0041]

[0042] in, It is a pseudo-inverse operator;

[0043] Step f03: Calculate the predicted output O of the convolutional network using formula (7);

[0044] O = I g ·β (7)

[0045] Step f04: Update the error e using formula (8);

[0046]

[0047] wherein, t i is the i-th component of the output T, o i is the i-th component of the predicted output O.

[0048] Further, the step f of calculating the error between the predicted output O and the true output T of the randomly configured convolutional network, if the error is greater than the expected error, continue to configure the convolutional network parameters, if the error is less than or equal to the expected error, the convolutional network is completed, comprising:

[0049] Step f05: determine whether the error e is greater than the expected error ε, if e>ε, continue to step f06, otherwise, the convolutional network is completed, jump to step f08;

[0050] Step f06: determine whether the convolution times L is less than or equal to the maximum convolution times L max , if L<=L max , jump to step c01 to continue to configure the convolutional parameters, otherwise, Layer=Layer+1, L=1;

[0051] Step f07: determine whether the convolution layer number Layer is less than or equal to the maximum convolution layer number Layer max , if Layer<=Layer max , jump to step c01 to continue to configure the convolutional parameters, otherwise, the convolutional network is completed, jump to step f08;

[0052] Step f08: the completed convolutional network is used for abnormal working condition detection of the magnesium furnace to be detected.

[0053] The application also provides a magnesium furnace abnormal working condition recognition device based on a randomly configured convolutional network, the device comprising:

[0054] A data set acquisition module is configured to acquire a magnesium furnace image training data set;

[0055] A network initialization module is configured to set parameters of the convolutional network and initialize;

[0056] A parameter configuration module is configured to configure the L-th convolutional parameter of the Layer-th convolutional layer of the convolutional network, and determine whether the parameter meets the configuration condition, if yes, construct a convolutional parameter matrix, if no, reconfigure the parameter until the configuration condition is met;

[0057] A feature map generation module is configured to generate a feature map by using the convolutional parameter matrix;

[0058] The pooling operation module is used to perform pooling operations on the feature map to generate a downsampled feature map;

[0059] The error judgment module is used to update the input data using the downsampled feature map, calculate the error between the predicted output O and the true output T of the randomly configured convolutional network, and continue to configure the convolutional network parameters if the error is greater than the expected error. If the error is less than or equal to the expected error, the convolutional network is completed and used to detect abnormal operating conditions of the magnesium furnace to be detected.

[0060] Furthermore, the dataset acquisition module is also used for:

[0061] The training dataset input is X = {x1, x2, ..., x...} N The output is T = {t1, t2, ..., t}. N}, where N is the number of samples, x i For the i-th input sample, t i Let be the i-th output sample, where i = 1, 2, ..., N.

[0062] Furthermore, the network initialization module is also used for:

[0063] Configure the parameters of the convolutional network, including the maximum number of convolutional layers. max Maximum number of convolutions per layer L max Expected error ε, kernel size k;

[0064] Initialize the convolutional network with L = 0 convolutions, Layer = 0 convolutional layers, input I = X, and error...

[0065] Furthermore, the parameter configuration module is also used for:

[0066] Step c01: Trim the input data I into input data I′ of size k×k;

[0067] Step c02: Randomly select the weight W using formula (1) L,Layer Bias b L,Layer ;

[0068]

[0069] Where λ is a fixed parameter, rand(a,b) is a random function that generates a matrix of size a×b using a uniformly distributed random sample on [0,1), and C in The number of channels for input data I;

[0070] Step c03: Calculate h using formula (2) L,Layer ,

[0071] h L,Layer = g(W L,Layer ·I′ + b L,Layer ) (10)

[0072] wherein g(·) is an activation function, and · is a matrix multiplication;

[0073] Step c04: calculating the constraint ξ L,Layer using formula (3), if ξ L,Layer < 0, the parameters do not meet the configuration condition, let r = r + τ and return to step c01 to reselect the convolution kernel parameters, otherwise continue step c05;

[0074]

[0075] wherein e T is the transpose of error e, h L,Layer T is the transpose of h L,Layer , r is a constraint parameter, 0 < r < 1, and τ is a configuration fixed parameter update step length;

[0076] Step c05: constructing the weight parameter matrix W Layer = [W 1,Layer , W 2,Layer ,..., W L,Layer ] of the Layer layer, and the bias parameter matrix b Layer = [b 1,Layer ,..., b L,Layer ], wherein the weight parameter matrix and the bias parameter matrix constitute a convolution parameter matrix.

[0077] Further, the feature map generation module is further configured to:

[0078] performing convolution operation on the magnesium furnace image training data set using formula (4) to generate a feature map F L,Layer ;

[0079]

[0080] wherein I i,k is the k-th channel matrix of the i-th input data, i = 0, 1,..., N, k = 1, 2,..., C in , and * is a cross-correlation operator.

[0081] Further, the pooling operation module is further configured to:

[0082] performing pooling operation on the feature map F L,Layer using formula (5) to generate a down-sampling feature map F′ L,Layer ;

[0083]

[0084] where F i L,Layer is the i-th input data, h = 0, 1,..., H-1, H is the number of rows of the feature map F L,Layer , w = 0, 1,..., W-1, W is the number of columns of the feature map F L,Layer .

[0085] Further, the error judgment module uses the down-sampling feature map to update the input data, comprising:

[0086] Step f01: update the input data I to I g , I g = [F' 1,Layer ,..., F' L,Layer ], and let L = L + 1;

[0087] Step f02: calculate the output layer parameter β of the convolutional network using formula (6);

[0088]

[0089] where is the pseudo-inverse operator;

[0090] Step f03: calculate the predicted output O of the convolutional network using formula (7);

[0091] O = I g · β (15)

[0092] Step f04: update the error e using formula (8);

[0093]

[0094] where t i is the i-th component of the output T, and o i is the i-th component of the predicted output O.

[0095] Further, the error judgment module calculates the error between the predicted output O and the true output T of the randomly configured convolutional network, and if the error is greater than the expected error, the convolutional network parameters are continued to be configured, and if the error is less than or equal to the expected error, the convolutional network is completed, comprising:

[0096] Step f05: judge whether the error e is greater than the expected error ε, if e > ε, continue to step f06, otherwise, the convolutional network is completed, and jump to step f08;

[0097] Step f06: judge whether the convolution times L is less than or equal to the maximum convolution times L max, if L <= L max , then jump to step c01 to continue configuring the convolution parameter, otherwise, let Layer = Layer + 1, L = 1;

[0098] Step f07: judge whether the convolution layer number Layer is less than or equal to the maximum convolution layer number Layer max , if Layer <= Layer max , then jump to step c01 to continue configuring the convolution parameter, otherwise, the convolution network is completed, and jump to step f08;

[0099] Step f08: the completed convolution network is used for abnormal working condition detection of the magnesium furnace to be detected.

[0100] The advantages of the present application are:

[0101] (1) The present application configures the convolution parameters of the convolution network, meets the configuration conditions to proceed to the next step, ensures the reliability of the model parameters to a certain extent, and calculates the error between the prediction output O of the convolution network and the true output T. In the case that the error exceeds the expected error, the convolution network parameters are continued to be configured, the prediction error of the constructed convolution network is avoided to be too large, the accuracy of the convolution network output result is improved, so that the accuracy of the magnesium furnace abnormal working condition recognition is high, and the present application constructs the convolution parameter matrix based on the convolution network, and generates the feature map by using the convolution parameter matrix, so that the electric melting magnesium furnace abnormal fault model is convenient to deploy in the magnesium furnace production site.

[0102] (2) Because the size of the magnesium furnace image is large, the parameters of the traditional full connection network are more prone to overfitting, therefore, the feature representation in the magnesium furnace image is extracted by using the convolution operation formula (4), and the maximum pooling operation formula (5) is used to compress the feature map, so as to reduce the parameter amount in the network, speed up the operation speed, and prevent the model from overfitting. BRIEF DESCRIPTION OF DRAWINGS

[0103] Figure 1 The flowchart of the magnesium furnace abnormal working condition recognition method based on the random configuration of the convolution network disclosed in the embodiment 1 of the present application. DETAILED DESCRIPTION

[0104] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, the technical scheme in the embodiments of the present application will be described clearly and completely in combination with the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0105] Embodiment 1

[0106] As Figure 1 shown in the magnesium furnace abnormal condition recognition method based on the random configuration convolutional network, the method comprises:

[0107] S1: obtaining a magnesium furnace image training data set;

[0108] The training data set input is X={x1,x2,...,x N}, the output is T={t1,t2,...,t N}, N is the number of samples x i is the i-th input sample, t i is the i-th output sample, where i=1,2,...,N.

[0109] S2: setting the parameters of the random configuration convolutional network and initializing; the specific process of this step is:

[0110] Set the parameters of the random configuration convolutional network, the maximum number of convolutional layers Layer max , the maximum number of convolutions L max of each layer, the expected error ε, and the convolution kernel size k;

[0111] Initialize the random configuration convolutional network, the number of convolutions L=0, the number of convolutional layers Layer=0, the input I=X, and the error

[0112] S3: configuring the L-th convolutional parameter of the Layer-th convolutional layer of the convolutional network, and judging whether the parameter meets the configuration condition, if it meets, constructing a convolutional parameter matrix, if it does not meet, reconfiguring the parameter until the configuration condition is met; the specific process of S3 is:

[0113] S31: cutting the input data I into input data I' with a size of k×k;

[0114] S32: randomly selecting the weight W L,Layer and the bias b L,Layer using formula (1);

[0115]

[0116] Where λ is a configuration fixed parameter, rand(a,b) is a random function that generates a matrix with a size of a×b using a uniform distribution of random samples on [0,1), C in is the number of channels of the input data I;

[0117] S33: calculating h L,Layer using formula (2)

[0118] h L,Layer = g(W L,Layer ·I′+bL,Layer ) (18)

[0119] wherein g(·) is an activation function, and · is a matrix multiplication;

[0120] S34: calculating the constraint ξ using formula (3) L,Layer , and determining whether ξ L,Layer is greater than 0, if ξ L,Layer < 0, the parameters do not meet the configuration condition, r = r + τ and returning to step S31 to reselect the convolution kernel parameters, otherwise, continuing step S35;

[0121]

[0122] wherein e T is the transpose of the error e, h L,Layer T is the transpose of h L,Layer , r is a constraint parameter, 0 < r < 1, and τ is a configuration fixed parameter update step;

[0123] S35: constructing the weight parameter matrix W Layer = [W 1,Layer , W 2,Layer ,..., W L,Layer ] of the Layer layer, and the bias parameter matrix b Layer = [b 1,Layer ,..., b L,Layer ], and the weight parameter matrix and the bias parameter matrix constitute a convolution parameter matrix.

[0124] S4: generating a feature map using the convolution parameter matrix; the specific process of this step is as follows:

[0125] using the convolution parameter matrix, performing convolution operation on the magnesium furnace image training data set using formula (4) to generate a feature map F L,Layer ;

[0126]

[0127] wherein I i,k is the k-th channel matrix of the i-th input data, i = 0, 1,..., N, k = 1, 2,..., C in , and * is a cross-correlation operator.

[0128] S5: performing a pooling operation on the feature map to generate a down-sampling feature map; the specific process of this step is as follows:

[0129] performing a pooling operation on the feature map F L,Layer using formula (5) to generate a down-sampling feature map F' L,Layer ;

[0130]

[0131] wherein F i L,Layer is the i-th input data, h = 0, 1,..., H - 1, H is the number of rows of the feature map F L,Layer , w = 0, 1,..., W - 1, W is the number of columns of the feature map F L,Layer .

[0132] S6: updating the input data using the down-sampled feature map, calculating the error between the prediction output O of the random configuration convolutional network and the true output T, if the error is greater than the expected error, continue to configure the parameters of the convolutional network, if the error is less than or equal to the expected error, the convolutional network is completed and used for abnormal condition detection of the magnesium furnace to be detected. The specific process of this step is as follows:

[0133] S61: updating the input data I to I g , I g = [F' 1,Layer ,..., F' L,Layer ], and L = L + 1;

[0134] S62: calculating the output layer parameter β of the convolutional network using formula (6);

[0135]

[0136] wherein is the pseudo-inverse operator;

[0137] S63: calculating the prediction output O of the convolutional network using formula (7);

[0138] O = I g · β (23)

[0139] Step S64: updating the error e using formula (8);

[0140]

[0141] wherein t i is the i-th component of the output T, o i is the i-th component of the prediction output O.

[0142] Step S65: judging whether the error e is greater than the expected error ε, if e > ε, continue to step S66, otherwise, the convolutional network is completed and jump to step S68;

[0143] Step S66: judging whether the convolutional number L is less than or equal to the maximum convolutional number L max , if L <= L max , the convolutional network is completed and jump to step S68, otherwise, continue to step S67.If yes, jump to step S31 to continue configuring the convolution parameter, otherwise, let Layer = Layer + 1, L = 1;

[0144] Step S67: judge whether the convolution layer number Layer is less than or equal to the maximum convolution layer number Layer max , if Layer <= Layer max , jump to step S31 to continue configuring the convolution parameter, otherwise, the convolution network is constructed, and jump to step S68;

[0145] Step S68: the constructed convolution network is used for detecting the abnormal working condition of the magnesium furnace to be detected.

[0146] Through the above technical scheme, the convolution parameter of the convolution network is configured, the next operation is performed only when the configuration condition is met, the reliability of the model parameter is ensured to a certain extent, the error between the prediction output O of the convolution network and the true output T is calculated, the convolution network parameter is continued to be configured when the error exceeds the expected error, the prediction error of the constructed convolution network is avoided to be too large, the accuracy of the output result of the convolution network is improved, and thus the accuracy of the magnesium furnace abnormal working condition identification is high.

[0147] Embodiment 2

[0148] Based on embodiment 1, the embodiment 2 of the present application further provides a magnesium furnace abnormal working condition identification device based on a randomly configured convolution network, which comprises:

[0149] A data set acquisition module is configured to acquire a magnesium furnace image training data set;

[0150] A network initialization module is configured to set the parameters of the convolution network and initialize;

[0151] A parameter configuration module is configured to configure the Lth convolution parameter of the Layerth convolution of the convolution network, and judge whether the parameter meets the configuration condition, if yes, construct the convolution parameter matrix, if not, reconfigure the parameter until the configuration condition is met;

[0152] A feature map generation module is configured to generate a feature map by using the convolution parameter matrix;

[0153] A pooling operation module is configured to perform a pooling operation on the feature map to generate a down-sampling feature map;

[0154] An error judgment module is configured to update the input data by using the down-sampling feature map, calculate the error between the prediction output O of the randomly configured convolution network and the true output T, if the error is greater than the expected error, continue to configure the convolution network parameter, if the error is less than or equal to the expected error, the convolution network is constructed, and is used for detecting the abnormal working condition of the magnesium furnace to be detected.

[0155] Specifically, the dataset acquisition module is also used for:

[0156] The training dataset input is X = {x1, x2, ..., x...} N The output is T = {t1, t2, ..., t}. N}, where N is the number of samples, x i For the i-th input sample, t i Let be the i-th output sample, where i = 1, 2, ..., N.

[0157] Specifically, the network initialization module is also used for:

[0158] Configure the parameters of the convolutional network, including the maximum number of convolutional layers. max Maximum number of convolutions per layer L max Expected error ε, kernel size k;

[0159] Initialize the convolutional network with L = 0 convolutions, Layer = 0 convolutional layers, input I = X, and error...

[0160] More specifically, the parameter configuration module is also used for:

[0161] Step c01: Trim the input data I into input data I′ of size k×k;

[0162] Step c02: Randomly select the weight W using formula (1) L,Layer Bias b L,Layer ;

[0163]

[0164] Where λ is a fixed parameter, rand(a,b) is a random function that generates a matrix of size a×b using a uniformly distributed random sample on [0,1), and C in The number of channels for input data I;

[0165] Step c03: Calculate h using formula (2) L,Layer ,

[0166] h L,Layer =g(W L,Layer ·I′+b L,Layer (26)

[0167] Where g(·) is the activation function, and · is matrix multiplication;

[0168] Step c04: Calculate constraint ξ using formula (3) L,Layer If ξ L,LayerIf <0, the parameter does not meet the configuration condition, let r=r+τ and return to step c01 to reselect the convolution kernel parameter, otherwise continue to step c05;

[0169]

[0170] wherein e T is the transpose of error e, h L,Layer T is the transpose of h L,Layer , r is a constraint parameter, 0

[0171] Step c05: construct the weight parameter matrix W Layer =[W 1,Layer ,W 2,Layer ,...,W L,Layer ] of the Layer layer, the bias parameter matrix b Layer =[b 1,Layer ,...,b L,Layer ], and the weight parameter matrix and the bias parameter matrix constitute a convolution parameter matrix.

[0172] More specifically, the feature map generation module is further configured to:

[0173] perform convolution operation on the magnesium furnace image training data set by using formula (4) to generate a feature map F L,Layer .

[0174]

[0175] wherein I i,k is the k-th channel matrix of the i-th input data, i=0,1,...,N, k=1,2,...,C in , and * is a cross-correlation operator.

[0176] More specifically, the pooling operation module is further configured to:

[0177] perform pooling operation on the feature map F L,Layer by using formula (5) to generate a down-sampling feature map F' L,Layer .

[0178]

[0179] wherein F i L,Layer , i=1,...N is a feature sub-map of the i-th input data, h=0,1,...,H-1, H is the number of rows of the feature map F L,Layer , w=0,1,...,W-1, W is the number of columns of the feature map F L,Layer .

[0180] More specifically, the error judgment module updates the input data using the down-sampled feature map, including:

[0181] Step f01: update the input data I to I g , g = [F' 1,Layer ,...,F' L,Layer ], and let L = L + 1;

[0182] Step f02: calculate the output layer parameter β of the convolutional network using formula (6);

[0183]

[0184] wherein, is the pseudo-inverse operator;

[0185] Step f03: calculate the predicted output O of the convolutional network using formula (7);

[0186] O = I g · β (31)

[0187] Step f04: update the error e using formula (8);

[0188]

[0189] wherein, t i is the i-th component of the output T, and o i is the i-th component of the predicted output O.

[0190] More specifically, the error judgment module calculates the error between the predicted output O of the randomly configured convolutional network and the true output T, and if the error is greater than the expected error, the convolutional network parameters are continuously configured, and if the error is less than or equal to the expected error, the convolutional network is completed, including:

[0191] Step f05: judge whether the error e is greater than the expected error ε, if e > ε, continue to step f06, otherwise, the convolutional network is completed, and jump to step f08;

[0192] Step f06: judge whether the convolution times L is less than or equal to the maximum convolution times L max , if L <= L max , jump to step c01 to continue to configure the convolutional parameters, otherwise, let Layer = Layer + 1, L = 1;

[0193] Step f07: judge whether the convolution layer number Layer is less than or equal to the maximum convolution layer number Layer max , if Layer <= Layer max , jump to step c01 to continue to configure the convolutional parameters, otherwise, let Layer = Layer + 1, L = 1;If yes, the step c01 continues to configure the convolution parameter, otherwise, the convolution network is built, and the step f08 is jumped to;

[0194] The step f08 is used for detecting the abnormal working condition of the magnesium furnace to be detected by using the built convolution network.

[0195] The above examples are only used to illustrate the technical solutions of the present application, but not limit the same; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can be modified, or some technical features can be replaced by the equivalent ones; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A magnesium furnace abnormal condition recognition method based on a randomly configured convolutional network, characterized in that, The method comprises: Step a: obtaining a magnesium furnace image training data set; Step b: setting parameters of a convolutional network and initializing; the step b comprises: Setting the parameters of the convolutional network, the maximum number of convolutional layers Layer max , the maximum number of convolutions per layer L max , the expected error ε, the convolution kernel size k; Initialize the convolutional network, the number of convolutions L = 0, the number of convolutional layers Layer = 0, the input I = X, the error Step c: configuring the Lth convolutional parameter of the Layerth convolutional layer of the convolutional network, and judging whether the parameter meets the configuration condition; if yes, a convolutional parameter matrix is constructed; if not, the parameter is reconfigured until the configuration condition is met; the step c comprises: Step c01: cropping the input data I into input data I' with a size of k×k; Step c02: Randomly select the weight W using formula (1) L,Layer , the bias b L,Layer ; where λ is a configured fixed parameter, rand(a, b) is a random function that generates a matrix of size a x b with random samples from a uniform distribution on [0, 1), and C in is the number of channels of the input data I; Step c03: Calculate h using equation (2) L,Layer , h L,Layer = g(W L,Layer · I' + b L,Layer ) (2) Wherein, g(·) is an activation function, and · is matrix multiplication; Step c04: Calculate constraint ξ using formula (3) L,Layer If ξ L,Layer < 0, the parameters do not meet the configuration condition, let r = r + τ and return to step c01 to reselect the convolution kernel parameters, otherwise continue step c05; where e T is the transpose of the error e, h L,Layer T is the transpose of h L,Layer , r is a constraint parameter, 0 < r < 1, and τ is a configuration fixed parameter update step. Step c05: constructing weight parameter matrix W of the Layer layer Layer = [W 1,Layer ,W 2,Layer ,...,W L,Layer ], bias parameter matrix b Layer = [b 1,Layer ,...,b L,Layer ], the weight parameter matrix and the bias parameter matrix constitute a convolution parameter matrix; Step d: generating a feature map using the convolutional parameter matrix; the step d comprises: The formula (4) is used to perform convolution operation on the magnesium furnace image training dataset by using the convolution parameter matrix to generate a feature map F L,Layer ; where I i,k is the kth channel matrix of the ith input data, i = 0, 1,..., N, k = 1, 2,..., C in is the cross-correlation operator; Step e: performing a pooling operation on the feature map to generate a down-sampling feature map; Step f: updating the input data using the down-sampling feature map, calculating the error between the predicted output O of the randomly configured convolutional network and the true output T, and if the error is greater than the expected error, the convolutional network parameter is continuously configured; if the error is less than or equal to the expected error, the convolutional network is constructed and used for abnormal working condition detection of a magnesium furnace to be detected.

2. The magnesium furnace abnormal condition recognition method based on the randomly configured convolutional network according to claim 1, characterized in that, The step a comprises: The training dataset input is X = {x1, x2, ..., x} N The output is T = {t1, t2, ..., t}. N }, where N is the number of samples, x i For the i-th input sample, t i Let be the i-th output sample, where i = 1, 2, ..., N.

3. The method of claim 1, wherein the method is a method of magnesium furnace abnormal condition recognition based on a randomly configured convolutional network. The step e comprises: The feature map F is generated by using formula (5) L,Layer The pooling operation is performed to generate a down-sampling feature map F' L,Layer ; wherein F i L,Layer F is a feature map, i = 1,... N is a feature sub-map of the i-th input data, h = 0, 1,..., H - 1, H is a number of rows of the feature map F L,Layer , w = 0, 1,..., W - 1, W is a number of columns of the feature map F L,Layer .

4. The magnesium furnace abnormal condition recognition method based on the randomly configured convolutional network according to claim 3, characterized in that, In the step f, the input data is updated using the down-sampling feature map, comprising: Step f01: update the input data I to I g , g = [F' 1,Layer ,...,F' L,Layer ], and let L = L + 1; Step f02: calculating the output layer parameter β of the convolutional network using formula (6); wherein is the pseudo-inverse operator; Step f03: calculating the predicted output O of the convolutional network using formula (7); O = I g • β (7) Step f04: updating the error e using formula (8); where t i is the i-th component of the output T, o i is the i-th component of the predicted output O.

5. The method of claim 4, wherein the method is characterized by, In the step f, the error between the predicted output O of the randomly configured convolutional network and the true output T is calculated, and if the error is greater than the expected error, the convolutional network parameter is continuously configured; if the error is less than or equal to the expected error, the convolutional network is constructed and used for abnormal working condition detection of a magnesium furnace to be detected, comprising: Step f05: judging whether the error e is greater than the expected error ε; if e>ε, step f06 is continued; otherwise, the convolutional network is constructed and step f08 is jumped; Step f06: judge whether the convolution times L is less than or equal to the maximum convolution times L max , if L <= L max , then jump to step c01 to continue configuring the convolution parameters, otherwise, let Layer = Layer + 1, L = 1; Step f07: judging whether the convolution layer number Layer is less than or equal to the maximum convolution layer number Layer max , if Layer <= Layer max , then jumping to step c01 to continue configuring the convolution parameter, otherwise, the convolution network is completed, and jumping to step f08; Step f08: the constructed convolutional network is used for abnormal working condition detection of a magnesium furnace to be detected.

6. The magnesium furnace abnormal condition recognition device based on a randomly configured convolutional network, characterized in that, The device comprises: A data set acquisition module for acquiring a magnesium furnace image training data set; A network initialization module for setting parameters of a convolutional network and initializing; the network initialization module is further used for: Setting the parameters of the convolutional network, the maximum number of convolutional layers Layer max , the maximum number of convolutions per layer L max , the desired error ε, the size of the convolution kernel k; Initialize the convolutional network, the number of convolutions L = 0, the number of convolutional layers Layer = 0, the input I = X, the error A parameter configuration module for configuring the Lth convolutional parameter of the Layerth convolutional layer of the convolutional network, and judging whether the parameter meets the configuration condition; if yes, a convolutional parameter matrix is constructed; if not, the parameter is reconfigured until the configuration condition is met; the parameter configuration module is further used for: Step c01: cropping the input data I into input data I' with a size of k×k; Step c02: Randomly select the weight W using equation (1) L,Layer , the bias b L,Layer ; where λ is a configured fixed parameter, rand(a, b) is a random function that generates a matrix of size a x b with random samples from a uniform distribution on [0, 1), and C in is the number of channels of the input data I; Step c03: Calculate h using equation (2) L,Layer , h L,Layer = g(W L,Layer · I' + b L,Layer ) (2) Wherein, g(·) is an activation function, and · is matrix multiplication; Step c04: Calculate the constraint ξ using formula (3) L,Layer If ξ < 0, the parameter does not meet the configuration condition, let r = r + τ and return to step c01 to reselect the convolution kernel parameter, otherwise continue step c05. L,Layer ​ wherein e T is the transpose of the error e, h L,Layer T is the transpose of h L,Layer , r is a constraint parameter, 0 < r < 1, and τ is a configuration fixed parameter update step. Step c05: constructing weight parameter matrix W of the Layer layer Layer = [W 1,Layer ,...,W 2,Layer ,...,W L,Layer ], bias parameter matrix b Layer = [b 1,Layer ,...,b L,Layer ], the weight parameter matrix and the bias parameter matrix constitute a convolution parameter matrix; A feature map generation module for generating a feature map using the convolutional parameter matrix; the feature map generation module is further used for: The formula (4) is used to perform convolution operation on the magnesium furnace image training data set by using the convolution parameter matrix to generate a feature map F L,Layer ; where I i,k is the kth channel matrix of the ith input data, i = 0, 1,..., N, k = 1, 2,..., C in is the cross-correlation operator; A pooling operation module for performing a pooling operation on the feature map to generate a down-sampling feature map; The error judgment module is configured to update the input data by using the down-sampling feature map, calculate the error between the prediction output O of the random configuration convolutional network and the true output T, continue to configure the parameters of the convolutional network if the error is greater than the expected error, and complete the construction of the convolutional network if the error is less than or equal to the expected error, so as to detect the abnormal working condition of the magnesium furnace to be detected.

7. The magnesium furnace abnormal condition recognition device based on the random configuration convolutional network according to claim 6, characterized in that, The data set acquisition module is further configured to: The training dataset input is X = {x1, x2, ..., x...} N The output is T = {t1, t2, ..., t}. N }, where N is the number of samples, x i For the i-th input sample, t i Let be the i-th output sample, where i = 1, 2, ..., N.

Images