A method for tracing the source of quality anomalies in the sintering process

Abstract

The present invention provides a method for tracing the quality anomaly of a sintering process, and relates to the technical field of sintering process. The method comprises establishing a stationary feature module for extracting the quality of the sintering process, using the collected normal sintering process data as a training set, training the stationary feature module for extracting the quality of the sintering process to obtain the stationary component of the training set and its covariance matrix, and the FeO content prediction submodule obtains the predicted value of the FeO content according to the stationary component of the training set; using the sintering process data monitored online as a sample set, processing the sample set according to the stationary feature extraction submodule based on a twin neural network to obtain the stationary component of the sample set; judging whether the sample set has abnormal data according to the stationary component and covariance matrix of the sample set, and if there is abnormal data, drawing a contribution graph, the present invention takes into account the non-stationarity, nonlinearity and non-Gaussianity of the sintering process data, extracts the stationary features related to the FeO content, and improves the fault detection performance.

Description

Technical Field

[0001] The invention relates to the technical field of sintering technology, and in particular to a method for tracing the source of quality abnormalities in a sintering process. Background Art

[0002] The blast furnace ironmaking process mainly includes four processes: material yard, sintering, pelletizing and blast furnace, which complete the tasks of iron-containing raw material preparation, material pretreatment (producing sintered ore and pelletized ore) and ironmaking in sequence. Among them, sintered ore is the main raw material for blast furnace smelting, accounting for more than 70% of the raw materials entering the blast furnace, providing good air permeability, production efficiency and ironmaking coke ratio for the blast furnace ironmaking process. Therefore, the quality of sintered ore directly affects the blast furnace condition, molten iron quality and production energy consumption, and is crucial to the blast furnace ironmaking process.

[0003] The chemical composition of sintered ore is the basis of its quality, which mainly includes iron grade, ferrous oxide (FeO), alkalinity, silicon dioxide, manganese oxide, aluminum oxide and a small amount of harmful elements such as sulfur and phosphorus; improving the chemical composition of sintered ore based on its quality basis is more conducive to adjusting the overall quality and performance of sintered ore. For example, the FeO content of sintered ore is regarded as a sign of temperature and heat level during the sintering process, which has a direct impact on the physical and metallurgical properties of sintered ore. When the FeO content decreases, the strength and particle size of the sintered ore will decrease, which is not conducive to improving its low-temperature reduction pulverization performance, and will cause changes in the blast furnace block band, making the blast furnace permeability index worse; when the FeO content increases, the reducibility of the sintered ore will decrease, which will lead to an increase in the blast furnace fuel ratio.

[0004] However, the existing technology does not take the FeO content into consideration in the evaluation of sintering quality, so when there is a problem with the sintering quality, it is impossible to fully trace the fault. Summary of the invention

[0005] In order to solve the technical problems existing in the prior art, an embodiment of the present invention provides a method for tracing the source of quality abnormalities in a sintering process. The technical solution is as follows:

[0006] A method for tracing the source of quality abnormalities in a sintering process, comprising:

[0007] S1. Establishing a steady feature module for extracting the quality of the sintering process, wherein the steady feature module for extracting the quality of the sintering process includes a steady feature extraction submodule based on a twin neural network and a FeO content prediction submodule;

[0008] S2, using the collected normal sintering process data as a training set, training the stationary feature module for extracting the sintering process quality to obtain the stationary component of the training set and its covariance matrix, and the FeO content prediction submodule obtains the corresponding FeO content prediction value according to the stationary component of the training set;

[0009] S3, taking the sintering process data monitored online as a sample set, processing the sample set according to the stationary feature extraction submodule based on the twin neural network, and obtaining a stationary component of the sample set;

[0010] S4. Determine whether the sample set has abnormal data according to the stationary component of the sample set and the covariance matrix. If abnormal data exists, draw a contribution graph, and obtain the variable with fault in the sample set according to the contribution graph.

[0011] Optionally, the stationary feature extraction submodule based on the twin neural network in S1 includes: an input layer, a subnetwork and an output layer;

[0012] The number of sub-networks is equal to the number of segments of training data received by the input layer;

[0013] The output layer measures the similarity between data based on Wasserstein distance;

[0014] Among them, when the stationary feature extraction submodule based on the twin neural network extracts stationary features, n segments of training data are input into the twin neural networks of n identical sub-networks through the input layer to obtain hidden features, and the output layer calculates the Wasserstein distance between the hidden features.

[0015] Optionally, the FeO content prediction submodule in S1 includes:

[0016] The FeO content prediction submodule is based on a long short-term memory recursive neural network.

[0017] Optionally, the normal sintering process data collected in S2 is used as a training set to train the stationary feature module for extracting the sintering process quality to obtain a stationary feature set including:

[0018] Collect normal sintering process data as a training set;

[0019] Performing standardization on the training set to obtain a standardized training set;

[0020] The standardized training set is divided into n time periods, and the standardized training set divided into n time periods is brought into the stationary feature extraction submodule based on the twin neural network to obtain the stationary component.

[0021] Optionally, the standardized training set is divided into n time periods, where n satisfies: n≥((m-d+1) / 2)+1;

[0022] Where m is the data dimension of the training set, d is the number of stationary sources in the training set, and d is determined based on the Johansen cointegration test.

[0023] Optionally, the formula of the covariance matrix is ​​formula (8):

[0024] ; (8)

[0025] in, is the stationary component of the training set, A corresponding covariance matrix is ​​obtained for the stationary component of the training set.

[0026] Optionally, the step of using the online monitored sintering process data as a sample set in S3, and processing the sample set according to the stationary feature extraction submodule based on the twin neural network, to obtain the stationary component of the sample set includes:

[0027] Collecting the sintering process data monitored online as a sample set;

[0028] Performing standardization on the sample set to obtain a standardized sample set;

[0029] The standardized sample set is used as input data for the stationary feature extraction submodule based on the twin neural network to obtain the stationary component of the sample set.

[0030] Optionally, judging whether the sample set contains abnormal data according to the stationary component of the sample set and the covariance matrix in S4 includes:

[0031] Calculate statistics for sample sets;

[0032] Calculate control limits for statistics from sample sets;

[0033] If the statistic of the data is higher than the control limit of the statistic of the sample set, it is determined that abnormal data exists in the sample set.

[0034] Optionally, if there is abnormal data in S4, a contribution graph is drawn, and the cause of the fault in the sample set obtained according to the contribution graph includes:

[0035] An autoencoder is constructed, wherein the stationary feature extraction submodule based on the twin neural network is used as the encoding part of the autoencoder, and the decoder network constructed using LSTM is used as the decoding part of the autoencoder. The loss function of the autoencoder is the reconstruction error, and the formula of the reconstruction error is formula (11):

[0036] ; (11)

[0037] Among them, MSE is the reconstruction error, and N is the number of data in the training set; is the i-th sample after reconstruction;

[0038] is the i-th data of the training set;

[0039] Calculating the contribution of each variable of each sample in the sample set;

[0040] According to the contribution of each variable of each sample in the sample set, a contribution rate of each variable of each sample is obtained;

[0041] Draw a histogram of the contribution rate corresponding to each variable in the abnormal data to obtain a contribution graph.

[0042] Optionally, the calculation formula for the contribution rate of each sample is formula (13):

[0043] ; (13)

[0044] Among them, cr i is the contribution rate of the i-th sample in the sample set, cr ij cr i The jth element in ;

[0045] cp i is the contribution of the i-th sample in the sample set, cp ij cp i The jth element in ;

[0046] m is the number of samples in the sample set.

[0047] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least:

[0048] The present invention fully considers the non-stationary characteristics of the sintering process data, and uses a twin neural network to extract stationary features, which also well considers the nonlinearity of the sintering process data, improves the performance of fault detection, and traces the fault by drawing a contribution graph. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0050] Figure 1 is a network diagram of a method for tracing the source of quality anomalies in a sintering process provided by an embodiment of the present invention;

[0051] Figure 2 is a diagram showing the effect of fault detection provided by an embodiment of the present invention. There are 540 test samples in total. A fault occurs at the 470th sample. 2The statistic exceeds the threshold, but the method proposed in this embodiment extracts and detects faults at 469 samples;

[0052] Figure 3 : is a schematic diagram of the false alarm rate of the validation set provided by an embodiment of the present invention. The validation set has a total of 500 samples, all of which are normal data, and the false alarm rate is 0.8%;

[0053] Figure 4 The contribution graph of the 478th sample in the sample set provided by the embodiment of the present invention shows that the contribution rate of the sixth variable to the fault is the largest. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solution and advantages of the embodiment of the present invention clearer, the technical solution of the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings of the embodiment of the present invention. Obviously, the described embodiment is a part of the embodiment of the present invention, not all of the embodiments. Based on the described embodiment of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] Unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. Similarly, words such as "one", "one" or "the" do not indicate a quantitative limitation, but indicate the existence of at least one. Words such as "include" or "comprise" mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Words such as "connect" or "connected" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect.

[0056] It should be noted that the terms "up", "down", "left", "right", "front", "back", etc. used in the present invention are only used to indicate relative position relationships. When the absolute position of the object being described changes, the relative position relationship may also change accordingly.

[0057] A method for tracing the source of quality abnormalities in a sintering process comprises the following steps:

[0058] S1. Establish a module for extracting the steady features of the sintering process quality. Here, extracting the steady features of the sintering process quality means extracting the steady features related to the quality in the sintering process, which will affect the quality.

[0059] Among them, the steady feature module for extracting the quality of the sintering process includes a steady feature extraction submodule based on a twin neural network and a FeO content prediction submodule; wherein, the FeO content prediction submodule includes: the FeO content prediction submodule is based on a long short-term memory recursive neural network.

[0060] The stationary feature extraction submodule based on the twin neural network in S1 includes: an input layer, a subnetwork and an output layer; the number of the subnetworks is equal to the number of segments of the training data received by the input layer; the output layer measures the similarity between data based on the Wasserstein distance;

[0061] Among them, when the stationary feature extraction submodule based on the twin neural network extracts stationary features, n segments of training data are input into the twin neural networks of n identical sub-networks through the input layer to obtain hidden features, and the output layer calculates the Wasserstein distance between the hidden features.

[0062] In this step, the module for extracting the stationary features of the sintering process quality is composed of a stationary feature extraction submodule based on a twin neural network and a FeO content prediction submodule based on a long short-term memory network (LSTM).

[0063] In this embodiment, the training set X∈R N×m is divided into n non-overlapping time periods (X 1 ,......,X n ), X 1 ∈R q ×m , q represents the number of samples in each training data segment, m represents the dimension of the training set, in this embodiment, m=15 variables, X 1 [1,:] represents X 1 The first sample of .

[0064] In this embodiment, the stationary feature extraction submodule based on the twin neural network is specifically:

[0065] A siamese neural network (SNN) is usually composed of two identical sub-networks with the same configuration, that is, they share the same parameters and weights. The identical sub-networks are designed to take paired inputs and measure how similar the two inputs are. A siamese neural network consists of three main parts: the input layer, the sub-network, and the output layer. The input layer receives two input samples, and at the output, a distance metric (such as Euclidean distance or Manhattan distance) is used to quantify the similarity between the input pairs. This output can be used directly for comparison tasks, or input into more layers for classification or regression.

[0066] In order to extract stable features, n training data are input into a twin neural network with n identical sub-networks to extract hidden features H(t)∈R nq×d , where d is the dimension of the extracted stationary features, H 1 [1,:] represents X 1 Hidden features of [1,:].

[0067] The differences in the distribution of hidden features of training data at different times are as follows:

[0068] Wasserstein distance is a metric used to measure the difference between two probability distributions. It considers the minimum cost required to transform one distribution into another. Compared with traditional metrics (such as KL divergence), Wasserstein distance does not rely on assumptions about data distribution, so it is more flexible and applicable to various types of data distributions, including non-Gaussian distributions. Therefore, it is suitable for fault detection tasks in actual industrial process data. Assumptions and is from the norm vector space ( , ) of d-dimensional random vectors, where is the basic norm. f and are two probability measures, , .

[0069] and The p-Wasserstein distance (p ≥ 1) between can be defined as follows:

[0070] ; (1)

[0071] yes The set of probability measures π on , whose marginal distributions are f and .

[0072] According to the calculation principle of Wasserstein distance, the first term of the loss function Loss1 is constructed:

[0073] ; (2)

[0074] represents the training data of the i-th period, represent WD represents the Wasserstein distance between the hidden features of two training data segments.

[0075] Eliminate the correlation between stationary features. The specific contents are as follows:

[0076] In order to remove redundant information in the features and retain the most representative features, a term is added here to make the covariance matrix of the extracted features the unit matrix, so that the decorrelated stationary features can be obtained and the second loss function can be constructed:

[0077] ; (3)

[0078] Where cov(.) represents the calculation of the covariance matrix, represents the square of the F norm, and I is the identity matrix.

[0079] Establish a FeO content prediction submodule based on LSTM, the specific contents are as follows:

[0080] The hidden feature H(t) extracted above is input, and the output is the predicted value of FeO content The actual value of FeO content is . The loss function is:

[0081] ; (4)

[0082] The above three formulas (2) to (4) constitute the final loss function, which can be expressed as formula (5):

[0083] ; (5)

[0084] Among them, α> 0 and β> 0 are weight hyperparameters. After multiple adjustments, this embodiment finally sets α=0.5 and β=e -2 .

[0085] S2, using the collected normal sintering process data as a training set, training the module for extracting the stationary features of the sintering process quality to obtain the stationary component of the training set, and the FeO content prediction submodule obtains the corresponding covariance matrix according to the stationary component of the training set;

[0086] The normal sintering process data collected in S2 is used as a training set to train the stationary feature module for extracting the sintering process quality to obtain a stationary feature set including:

[0087] Collect normal sintering process data as a training set;

[0088] Performing standardization on the training set to obtain a standardized training set;

[0089] The standardized training set is divided into n time periods, and the standardized training set divided into n time periods is brought into the stationary feature extraction submodule based on the twin neural network to obtain the stationary component.

[0090] The S2 step is mainly offline training. The specific steps are as follows:

[0091] Normal sintering process data is collected as a training set, and the training set is standardized in the following way:

[0092] ; (5)

[0093] in, represents the mean of the j-th dimension feature, Represents the standard deviation of the j-th dimension feature.

[0094] Divide the standardized training set into n periods;

[0095] If the value of n is too small, it may lead to false stationarity. Although the entire time series appears to be stationary, it actually contains non-stationary components. This will make it impossible to accurately identify the true stationary source. Therefore, n should meet certain conditions:

[0096] ; (7)

[0097] The number of stationary sources d is determined using the Johansen cointegration test;

[0098] The formula of the covariance matrix is ​​formula (8):

[0099] ; (8)

[0100] in, is the stationary component of the training set, A corresponding covariance matrix is ​​obtained for the stationary component of the training set.

[0101] S3, taking the sintering process data monitored online as a sample set, processing the sample set according to the stationary feature extraction submodule based on the twin neural network, and obtaining a stationary component of the sample set;

[0102] Collecting the sintering process data monitored online as a sample set;

[0103] Performing standardization on the sample set to obtain a standardized sample set;

[0104] The standardized sample set is used as input data for the stationary feature extraction submodule based on the twin neural network to obtain the stationary component of the sample set.

[0105] In step S3, the main task is online detection, and the specific steps include:

[0106] For online monitoring samples, standardization is performed;

[0107] Input the standardized sample set into the model trained in step 1 to obtain the stationary component ;

[0108] Calculate T 2 Statistics, calculated as follows:

[0109] ; (9)

[0110] Calculate T 2 Control limits of statistics T a , the calculation formula is as follows:

[0111] ; (10)

[0112] Among them, a represents the significance level, and its value is set to 0.05. It obeys the F distribution with the first degree of freedom h and the second degree of freedom Nh, where N is the number of samples in the training data set.

[0113] If the T of the monitoring sample 2 The statistic is above the control limit T a , then it means a fault has occurred.

[0114] S4. Determine whether the sample set has abnormal data according to the stationary component of the sample set and the covariance matrix. If abnormal data exists, draw a contribution graph, and obtain the variation of the fault in the sample set according to the contribution graph.

[0115] Calculate statistics for sample sets;

[0116] Calculate control limits for statistics from sample sets;

[0117] If the statistic of the data is higher than the control limit of the statistic of the sample set, it is determined that abnormal data exists in the sample set.

[0118] If there is abnormal data in S4, a contribution graph is drawn, and the causes of the failure in the sample set obtained according to the contribution graph include:

[0119] An autoencoder is constructed, wherein the stationary feature extraction submodule based on the twin neural network is used as the encoding part of the autoencoder, and the decoder network constructed using LSTM is used as the decoding part of the autoencoder. The loss function of the autoencoder is the reconstruction error, and the formula of the reconstruction error is formula (11):

[0120] ; (11)

[0121] Among them, MSE is the reconstruction error, and N is the number of data in the training set; is the i-th sample after reconstruction;

[0122] is the i-th data of the training set;

[0123] Calculating the contribution of each variable of each sample in the sample set;

[0124] According to the contribution of each variable of each sample in the sample set, a contribution rate of each variable of each sample is obtained;

[0125] Draw a histogram of the contribution rate corresponding to each variable in the abnormal data to obtain a contribution graph.

[0126] Optionally, the calculation formula for the contribution rate of each sample is formula (13):

[0127] ; (13)

[0128] Among them, cr i is the contribution rate of the i-th sample in the sample set, cr ij cr i The jth element in ;

[0129] cp i is the contribution of the i-th sample in the sample set, cp ij cp i The jth element in ;

[0130] m is the number of samples in the sample set.

[0131] In this step S4, the main purpose is to draw a contribution graph and trace the fault, which mainly includes:

[0132] After a fault is detected, it is necessary to trace the fault to its source, that is, to find out the variables related to the fault. This embodiment uses a contribution graph-based method to trace the fault. This method constructs a contribution index for each variable after a fault occurs to show the possibility of a fault occurring due to the variable. Here, a method based on an autoencoder is used to construct a contribution graph.

[0133] When a fault occurs, the fault information will be propagated throughout the autoencoder, so all output variables will contain the fault information. The encoding part of the autoencoder is based on the stationary feature extraction submodule of the twin neural network.

[0134] The decoding part uses LSTM to construct a decoder network, and the input of the network is the stationary component of the training set. , the output is the reconstructed data of the training set, so the loss function of the autoencoder is the reconstruction error:

[0135] ; (11)

[0136] Among them, MSE is the reconstruction error, and N is the number of data in the training set; is the i-th sample after reconstruction; is the i-th data of the training set.

[0137] Since normal training data is used when training the network, if the feature is not faulty, the encoder can reconstruct it well and the reconstruction error is naturally small. However, if the feature is faulty, the reconstruction error should increase. Represents the contribution of the i-th sample in the data set, which is calculated as follows:

[0138] ; (12)

[0139] in , and Represents vectors , and In order to reduce the impact of the contribution index on the feature mining process, the contribution is further converted into a contribution rate by normalization. , the calculation method is as follows:

[0140] ; (13)

[0141] Among them, cr i is the contribution rate of the i-th sample in the sample set, cr ij cr i The jth element in ;

[0142] cp i is the contribution of the i-th sample in the sample set, cp ij cp i The jth element in ;

[0143] m is the number of samples in the sample set.

[0144] like Figure 4 As shown, variables 1-15 are the horizontal axes, and the contribution rate of each variable is the vertical axis. In the figure, variable 6 has the largest contribution rate. We can analyze and trace the source of variable 6 to find the source of the fault. The contribution of variable 15 is almost 0, and its impact on this fault is basically 0.

[0145] The present invention fully considers the non-stationary characteristics of the sintering process data, and uses a twin neural network to extract stationary features, which also well considers the nonlinearity of the sintering process data, improves the performance of fault detection, and traces the fault by drawing a contribution graph.

[0146] The following points need to be explained:

[0147] (1) The drawings of the embodiments of the present invention only involve structures related to the embodiments of the present invention. Other structures may refer to conventional designs.

[0148] (2) For the sake of clarity, in the drawings used to describe the embodiments of the present invention, the thickness of layers or regions is exaggerated or reduced, that is, these drawings are not drawn according to the actual scale. It is understood that when an element such as a layer, film, region or substrate is referred to as being "on" or "under" another element, the element may be "directly" "on" or "under" the other element or there may be intermediate elements.

[0149] (3) In the absence of conflict, the embodiments of the present invention and the features therein may be combined with each other to obtain new embodiments.

[0150] The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. The protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for tracing the source of quality abnormalities in a sintering process, characterized in that: include: S1. Establishing a steady feature module for extracting the quality of the sintering process, wherein the steady feature module for extracting the quality of the sintering process includes a steady feature extraction submodule based on a twin neural network and a FeO content prediction submodule; S2, using the collected normal sintering process data as a training set, training the stationary feature module for extracting the sintering process quality to obtain the stationary component of the training set and its covariance matrix, and the FeO content prediction submodule obtains the predicted value of the FeO content according to the stationary component of the training set; S3, taking the sintering process data monitored online as a sample set, processing the sample set according to the stationary feature extraction submodule based on the twin neural network, and obtaining a stationary component of the sample set; S4. judging whether there is abnormal data in the sample set according to the stationary component of the sample set and the covariance matrix, and if there is abnormal data, drawing a contribution graph, and obtaining the variable with fault in the sample set according to the contribution graph; The stationary feature extraction submodule based on the twin neural network in S1 includes: an input layer, a subnetwork and an output layer; The number of sub-networks is equal to the number of segments of training data received by the input layer; The output layer measures the similarity between data based on Wasserstein distance; When the stationary feature extraction submodule based on the twin neural network extracts stationary features, n segments of training data are input into the twin neural networks of n identical subnetworks through the input layer to obtain hidden features, and the output layer calculates the Wasserstein distance between the hidden features; The FeO content prediction submodule in S1 includes: The FeO content prediction submodule is based on a long short-term memory recurrent neural network; The normal sintering process data collected in S2 is used as a training set to train the stationary feature module for extracting the sintering process quality to obtain a stationary feature set including: Collect normal sintering process data as a training set; Performing standardization on the training set to obtain a standardized training set; The standardized training set is divided into n time periods, and the standardized training set divided into n time periods is brought into the stationary feature extraction submodule based on the twin neural network to obtain the stationary component.

2. The method for tracing the source of quality abnormality in the sintering process according to claim 1, characterized in that: The standardized training set is divided into n time periods where n satisfies: n≥((m-d+1) / 2)+1; Where m is the data dimension of the training set, d is the number of stationary sources in the training set, and d is determined based on the Johansen cointegration test.

3. The method for tracing the source of quality abnormality in the sintering process according to claim 2, characterized in that: The formula for obtaining the corresponding covariance matrix according to the stationary component of the training set is formula (8): in, is the stationary component of the training set, and S is the stationary component of the training set to obtain the corresponding covariance matrix.

4. The method for tracing the source of quality abnormality in the sintering process according to claim 3 is characterized in that: In S3, the sintering process data monitored online is used as a sample set, and the sample set is processed according to the stationary feature extraction submodule based on the twin neural network to obtain the stationary component of the sample set including: Collecting the sintering process data monitored online as a sample set; Performing standardization on the sample set to obtain a standardized sample set; The standardized sample set is used as input data for the stationary feature extraction submodule based on the twin neural network to obtain the stationary component of the sample set.

5. According to the sintering process quality anomaly tracing method of claim 4, the step of determining whether the sample set has abnormal data according to the stationary component of the sample set and the covariance matrix in S4 comprises: Calculate statistics for sample sets; Calculate control limits for statistics from sample sets; If the statistic of the data is higher than the control limit of the statistic of the sample set, it is determined that abnormal data exists in the sample set.

6. The method for tracing the source of quality abnormality in the sintering process according to claim 5, characterized in that: If there is abnormal data in S4, a contribution graph is drawn, and the reasons for the failure in the sample set obtained according to the contribution graph include: An autoencoder is constructed, wherein the stationary feature extraction submodule based on the twin neural network is used as the encoding part of the autoencoder, and the decoder network constructed using LSTM is used as the decoding part of the autoencoder. The loss function of the autoencoder is the reconstruction error, and the formula of the reconstruction error is formula (11): Among them, MSE is the reconstruction error, and N is the number of data in the training set; is the i-th sample after reconstruction; x i is the i-th data of the training set; Calculating the contribution of each variable of each sample in the sample set; According to the contribution of each variable of each sample in the sample set, a contribution rate of each variable of each sample is obtained; Draw a histogram of the contribution rate corresponding to each variable in the abnormal data to obtain a contribution graph.

7. The method for tracing the source of quality abnormality in the sintering process according to claim 6, characterized in that: The calculation formula for the contribution rate of each variable in the abnormal data is formula (13): Among them, cr i is the contribution rate of the i-th sample in the sample set, cr ij cr i The jth element in ; cp i is the contribution of the i-th sample in the sample set, cp ij cp i The jth element in ; m is the number of samples in the sample set.

Images