Rotary kiln sensor time sequence abnormity monitoring method, storage medium and electronic equipment

Abstract

The invention provides a time sequence abnormity monitoring method for a rotary kiln sensor, a storage medium and electronic equipment, and the method comprises the steps: carrying out the preliminary selection of input parameters according to a cement clinker sintering technology, and building a time sequence data set from the rotary kiln sensor; secondly, data preprocessing is carried out through abnormal value elimination or marking, mean filtering and homogenization, and a feature set is divided into a training set, a verification set and a test set; establishing a one-dimensional deep learning model for training to obtain an anomaly detection model, and establishing a multi-dimensional time sequence prediction model for training to obtain a prediction model; and carrying out anomaly detection and prediction filling on the input rotary kiln related sensor by using the two trained models. According to the scheme, the mutual relation and long-term time dependence among multiple variables are extracted, and the time sequence prediction result can be more accurate in long-term prediction by utilizing multi-dimensional data.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a rotary kiln sensor timing anomaly monitoring method, storage medium and electronic equipment. Background Art

[0002] Cement is one of the gel materials widely used in civil engineering, energy and electricity, and is an important basic industry of the national economy. Due to the special production conditions of the cement process industry and the tight characteristics of the production process, the failure of some sensors or system components will inevitably cause abnormalities in other modules, thus affecting the entire production process, resulting in low product yield and high defective rate.

[0003] In recent years, deep learning algorithms have been widely used in time series anomaly detection and prediction. Representative models include CNN, RNN, and attention-based neural networks. The attention-based neural network structure was first used for natural translation, and now it is widely used for machine translation, image captioning, and time series prediction. Although attention-based neural networks work well for some data sets, they are not suitable for the prediction of industrial multivariate data and have poor generalization capabilities. For industrial time series data, how to extract the relationship between multiple variables and long-term time dependencies while ensuring the prediction accuracy and generalization ability of the model has become a problem that researchers need to solve. Summary of the invention

[0004] The technical problem to be solved by the present application is that in the prior art, abnormalities of rotary kiln-related sensors cannot be judged, and the related abnormalities affect production. To this end, the present application proposes a rotary kiln sensor timing abnormality monitoring method, storage medium and electronic equipment.

[0005] The technical solution of the present application provides a method for monitoring the abnormal timing of a rotary kiln sensor, comprising: Selecting relevant variable data of the sensor according to the mechanism of cement clinker burning, wherein the relevant variable data is variable data related to predicting the quantity and strength of cement; Preprocessing the relevant variable data to obtain key variable data, dividing the key variable data into a training set and a test set as a data set for model input; Inputting the data set into a trained single-dimensional detection model to obtain anomaly detection results of the sensor; The data set is input into a trained multi-dimensional prediction detection model to obtain the prediction result of the sensor.

[0006] In some embodiments, the rotary kiln sensor timing anomaly monitoring method is described, wherein the sensor related variable data is selected according to the mechanism of cement clinker burning, and the related variable data is variable data related to predicting the quantity and strength of cement, including: The relevant variable data of the sensor include smoke chamber temperature variable, C5A outlet temperature variable, C5A feed pipe temperature variable, C5B outlet temperature variable, C5B feed pipe temperature variable, and kiln head hood temperature variable; wherein, C5A represents a cyclone numbered C5A; C5B represents a cyclone numbered C5B.

[0007] In some embodiments, the rotary kiln sensor timing anomaly monitoring method, the key variable data is obtained after preprocessing the relevant variable data, including: The relevant variable data of the sensor are subjected to outlier elimination or marking, mean filtering, and homogenization processing to obtain the key variable data.

[0008] In some embodiments of the rotary kiln sensor timing anomaly monitoring method, the single-dimensional detection model includes: The time block structure is executed twice to decompose the one-dimensional time series data into the trend part and the short-term fluctuation part according to the periodicity; Use convolutional neural network modeling to extract the two-dimensional time series matrix converted from one-dimensional time series data, and use linear neural network to extract the time trend characteristics in the time series matrix; Use the LSTM-SRU algorithm to extract all sequence splicing and sum to achieve time series prediction results; A reasonable threshold is obtained based on the validation set test, and anomaly detection is performed using the reasonable threshold and the time series prediction result.

[0009] The rotary kiln sensor time series anomaly monitoring method described in some schemes, wherein the convolutional neural network modeling is used to extract a two-dimensional time series matrix converted from one-dimensional time series data, including: Applying Fast Fourier Transform to mine short-term and long-term trends of sequences in the time dimension: ; ; ; Where: X 1D Represents smoke chamber temperature variable, C5A outlet temperature variable, C5A feed pipe temperature variable, C5B outlet temperature variable, C5B feed pipe temperature variable, kiln head hood temperature variable; A Represents the intensity of the frequency components of the above variables, with the largest intensity k The frequencies correspond to the most significant k The length of the cycle; this k Frequency express; The value of Represents a data sequence; k For each cycle represents, T represents the cycle length; the above formula is simplified to: ; The cycle is obtained by fast Fourier transform. The time series data is periodically decomposed according to different cycles t, and decomposed into trend series and short-term fluctuation series, including: X 1D The most significant k Cycle length X The matrix is ​​copied t-1 times and then concatenated X 1 , X 2 , …, X t is a two-dimensional matrix with a width of period t and a length of sequence length n, and then t Average pooling of windows: ; in X is the input matrix, X t is the trend sequence obtained after pooling. Since the width and window scale are the same, all the output matrices obtained are restored to a one-dimensional sequence. i and j is the absolute position of the input matrix, p and q is the relative position of the input matrix to the pooling area; the periodic sequence is obtained by subtracting the trend sequence from the original sequence X s :

[0010] The above process is repeated q times to obtain the sequence after all period decompositions. , fold the original one-dimensional time series data according to the corresponding period of the fluctuation sequence according to the following formula: ; Indicates that zero padding is performed at the end of the sequence so that the sequence length can be divided evenly; After the above operations, a set of two-dimensional matrices corresponding to the corresponding cycles of the two-dimensional time series matrix is ​​obtained.

[0011] In some schemes, the rotary kiln sensor timing anomaly monitoring method is described, wherein the LSTM-SRU algorithm is used to extract all sequence splicing sums to achieve timing prediction results, and the LSTM-SRU algorithm includes: The LSTM-SRU neural network calculates forgetting, input, and output gating by assigning different weights to the input data. By combining the improved LSTM network and the SRU network, the output of the LSTM network layer becomes the input of the SRU network layer. The dropout layer is introduced to randomly discard some data. Extract the hidden state of the last moment obtained by LSTM neural network training as the initial hidden state of SRU neural network; Two-layer time block output The subsequent calculation process is as follows: ; ; ; The data obtained by the above formula is passed through the fully connected layer and the ELU activation function to obtain the final output: ; Among them, Y represents the LSTM prediction output of the smoke chamber temperature variable, C5A outlet temperature variable, C5A feed pipe temperature variable, C5B outlet temperature variable, C5B feed pipe temperature variable, and kiln head hood temperature variable.

[0012] The rotary kiln sensor time series anomaly monitoring method described in some schemes inputs the data set into a trained single-dimensional detection model to obtain the sensor anomaly detection result, including: according to the unsupervised anomaly detection method, using the error obtained by prediction as the classification judgment basis, using the maximum value of the prediction loss on the verification data set as the anomaly detection threshold; in the test set, using the error generated by the prediction result and the actual result to compare the threshold obtained in the previous step to obtain the prediction classification result, and the prediction is marked as 1, and the prediction is marked as 0; The data set is input into the trained multi-dimensional prediction detection model to obtain the prediction result of the sensor. The multi-dimensional prediction detection model has a three-branch structure, and the three branches include: Historical time series branch, which uses a dual-period Dlinear model to extract the historical time series features of the prediction sequence; Related sequence branch: This branch uses a convolutional neural network to reduce the dimension of related time series data, splices historical time series, uses a scaled reset matrix and performs position encoding at different scales, and then passes through a multi-head attention module and a bidirectional cross attention module to extract inter-series features and internal features of each time series respectively; The relevant real-time branch uses convolutional neural networks to achieve dimensionality reduction and final fully connected output, and then introduces real-time relevant sensor data for real-time correction; The outputs of the above three branches are added and output to complete the prediction of multi-dimensional time series.

[0013] The technical solution of the present application also provides a computer-readable storage medium, in which program information is stored. After the program information is read and run by a computer, any of the above-mentioned rotary kiln sensor timing anomaly monitoring methods is executed.

[0014] The technical solution of the present application also provides an electronic device, comprising at least one processor and at least one memory, at least one of the memories storing instruction information, and at least one of the processors reading the instruction information and executing any of the above methods for monitoring the timing anomaly of a rotary kiln sensor.

[0015] The above technical solution has the following beneficial effects: The above scheme disclosed in the present application, first, performs preliminary selection of input parameters according to the cement clinker burning process, and then establishes a time series data set from the rotary kiln sensor. Secondly, data preprocessing is performed by removing or marking outliers, mean filtering, and homogenization, and the feature set is divided into a training set, a validation set, and a test set. A single-dimensional deep learning model is established for training to obtain an anomaly detection model, and a multi-dimensional time series prediction model is established for training to obtain a prediction model; the two trained models are used to perform anomaly detection and prediction filling for the input rotary kiln related sensors. This scheme extracts the relationship between multiple variables and long-term time dependence, and uses multi-dimensional data to make the time series prediction results more accurate in long-term predictions. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flow chart of the method for monitoring the abnormal timing of the rotary kiln sensor according to the present invention; Figure 2 It is a structural diagram of a single-dimensional anomaly detection model of the rotary kiln sensor timing anomaly monitoring method of the present invention; Figure 3 It is a time block structure diagram in the single-dimensional anomaly detection model structure of the rotary kiln sensor timing anomaly monitoring method of the present invention; Figure 4 It is a STM-SRU structure diagram in the single-dimensional anomaly detection model structure of the rotary kiln sensor timing anomaly monitoring method described in the present invention; Figure 5 It is a structural diagram of a multi-dimensional time series prediction model of the method for monitoring time series anomaly of a rotary kiln sensor according to the present invention; Figure 6The figure is a schematic diagram of the hardware connection relationship of the electronic device for executing the rotary kiln sensor timing anomaly monitoring method according to the present invention. DETAILED DESCRIPTION

[0017] The specific implementation of the present application is further described below with reference to the accompanying drawings.

[0018] It is easy to understand that according to the technical solution of the present application, without changing the essential spirit of the present application, a variety of structural modes and implementation modes that can be replaced by those skilled in the art can be replaced with each other. Therefore, the following specific implementation modes and drawings are only exemplary descriptions of the technical solution of the present application, and should not be regarded as the entirety of the present application or as a limitation or restriction on the technical solution of the application.

[0019] The directional terms such as up, down, left, right, front, back, front, back, top, bottom, etc. mentioned or may be mentioned in this specification are defined relative to the structures shown in the drawings. They are relative concepts and may change accordingly according to different positions and different usage conditions. Therefore, these or other directional terms should not be interpreted as restrictive terms.

[0020] The present application embodiment provides a method for monitoring abnormal timing of a rotary kiln sensor, such as Figure 1 As shown, the steps include: S10: Select relevant variable data of the sensor according to the mechanism of cement clinker burning, wherein the relevant variable data is variable data related to predicting the quantity and strength of cement.

[0021] First, from the perspective of raw material parameters, clinker index control parameters, clinker mineral composition, etc., the smoke chamber temperature refers to the temperature of the flue gas after passing through the cement clinker rotary kiln, which can represent the temperature at the kiln tail. C5A outlet temperature, C5A discharge pipe temperature, C5B outlet temperature, and C5B discharge pipe temperature are all relevant temperature parameters of the kiln tail part, which can be used to monitor the temperature of the clinker. The kiln head hood temperature can reflect the working status and cooling capacity of the kiln head area. Therefore, the above sensors can realize abnormal detection and prediction.

[0022] S20: Preprocessing the relevant variable data to obtain key variable data, dividing the key variable data into a training set and a test set as a data set for model input; Among them, data preprocessing is performed by removing and marking outliers, mean filtering, and homogenization, and the determined key variables are divided into training sets and test sets as model inputs; The method for removing or marking abnormal values ​​is: according to the expert rule library and the guidance of on-site technicians, the offset or drift existing in the sensor data is manually manifested as the sensor measurement value and the actual temperature are continuously higher or lower for a period of time; the value jump is manifested as a trend of stable change under normal circumstances, and the value suddenly becomes larger or suddenly becomes smaller; the periodic fluctuation anomaly is manifested as normal periodic fluctuations, the frequency of temperature fluctuations suddenly accelerates or slows down, and the amplitude of fluctuations suddenly increases or decreases; data loss or abnormal values ​​are manifested as data loss or data out of range, negative values, etc.; slow response or failure is manifested as data remaining unchanged for a long time, or the response speed decreases, etc., and then they are removed and marked.

[0023] The mean filtering formula is: ; Where n is the time window length of the mean filter, Represents the i-th related variable data; The homogenization formula is: .

[0024] Dividing the key variable data into a training set and a test set includes: removing a total of twelve sensor parameter data and eighty thousand abnormal data as a training set and a verification set, and marking ten thousand abnormal data as a test set.

[0025] S30: Input the data set into a trained single-dimensional detection model to obtain anomaly detection results of the sensor.

[0026] In this step, twelve types of sensor parameter data and 80,000 abnormal data are removed as training sets and validation sets, and 10,000 abnormal data are marked as test sets.

[0027] S40: Inputting the data set into a trained multi-dimensional prediction detection model to obtain a prediction result of the sensor.

[0028] The solution of this embodiment aims at the problem of abnormal detection and prediction of cement rotary kiln related sensors. A single-dimensional time series anomaly detection model and a multi-dimensional time series prediction model are designed based on a deep learning method, thereby realizing real-time abnormal detection and prediction of rotary kiln related sensors, and solving the problem in the prior art that the abnormalities of rotary kiln related sensors cannot be judged and the related abnormalities affect production.

[0029] In the above scheme, step S30 specifically includes: Step 31: Establish a single-dimensional anomaly detection model; Step 311: Establish a time block structure; Step 312: Establishing LSTM-SRU structure; Step 313: Anomaly detection classification method; Step 32: Establish a multi-dimensional time series prediction model; Step 321: Establish a historical time series branch; Step 322: Establishing related sequence branches; Step 323: Establish relevant real-time branches; Step 33: Obtain a one-dimensional anomaly detection model; like Figure 2 As shown, the anomaly detection model is: First, the time block structure is executed twice, with the aim of decomposing the one-dimensional time series data into two parts according to periodicity - the trend part and the short-term fluctuation part. Then, the convolutional neural network (CNN) is used to model and extract the two-dimensional time series matrix converted from the one-dimensional time series. The linear neural network (Linear) is used to extract the time trend features in order to improve the extraction of long-term dependencies.

[0030] Then, the LSTM-SRU module is used to extract all the sequences and concatenate and sum them to achieve single-dimensional time series prediction.

[0031] Finally, a reasonable threshold is obtained based on the validation set test, and the threshold and prediction results are used for anomaly detection.

[0032] Step 311: Establish a time block structure; like Figure 3 As shown, the time block structure is: Apply Fast Fourier Transform (FFT) in the time dimension to mine short-term and long-term trends of the sequence:

[0033]

[0034]

[0035] in: A Represents the intensity of each frequency component, with the largest intensity k The frequencies correspond to the most significant k The above process is simplified as follows:

[0036] According to the cycle obtained by fast Fourier transform, the time series data is periodically decomposed according to different periods t, and decomposed into trend series and short-term fluctuation series.

[0037] First of all X The sequences were replicated t-1 times and then spliced ​​X 1 , X 2 , …, X tis a two-dimensional matrix with a width of period t and a length of sequence length n, and then t Average pooling of windows:

[0038] in X is the input matrix, X t is the trend sequence obtained after pooling. Since the width and window scale are the same, all the output matrices obtained are restored to a one-dimensional sequence. i and j is the absolute position of the input matrix, p and q is the relative position of the input matrix to the pooling area. Then the periodic sequence is obtained by subtracting the trend sequence from the original sequence X s .

[0039]

[0040] This process will be repeated k times to obtain the sequence after all period decompositions For the fluctuation sequence, fold the original one-dimensional time series data according to the corresponding period according to the following formula:

[0041] In order to adapt the sequence length to FFT processing, It indicates that zero padding is performed at the end of the sequence so that the sequence length can be divided evenly. Finally, after the above operation, a set of two-dimensional matrices corresponding to the corresponding cycle-dominated two-dimensional time series changes are obtained.

[0042] Then, a convolutional neural network is used to extract the timing features from the two-dimensional timing matrix.

[0043] The network contains a 2D convolution layer, a Layer Norm layer, a 2D convolution layer, an Elu activation function, and then the residual summation after Dropout. Calculate according to .

[0044] ; ; ; ; ; in is the input and output vector, increasing from front to back.* represents the convolution operation, represents the convolution kernel, where Represents a 7*7 convolution kernel with a step size of 3*3. Represents a 1*1 convolution kernel with a step size of 1*1. is the bias term. γ is a learnable scaling parameter, b 1 is a learnable bias parameter, μ is The mean of The activation function between layers uses the linear unsaturated activation function (ELU) as the activation function, where exp Express Calculate the corresponding power of the natural constant e, thereby effectively solving the neuron death problem of the ReLU function.

[0045] Transform two dimensions to one dimension: For the extracted time features, transform them into one-dimensional space for subsequent fusion: ; Trunc It means to remove the 0 in the previous padding expansion and restore the sequence to its original state.

[0046] For the trend sequence generated previously X t , with the i Trend Series As an example, apply a single-layer linear neural network: ; Finally, the trend part and the periodic part of the time block output are summed with the input data: ; The above operations are run twice in the model, and the output is input into the lSTM-SRU module.

[0047] Step 312: Establishing LSTM-SRU structure; like Figure 4 As shown, the LSTM-SRU block structure is: The LSTM-SRU neural network calculates forgetting, input, and output gating by assigning different weights to the input data. By combining the improved LSTM network and the SRU network, the output of the LSTM network layer becomes the input of the SRU network layer. This structure helps the SRU layer fit the data faster. To prevent overfitting, the dropout layer is introduced to randomly discard some data to ensure the generalization ability of the network model. In order to maximize the continuity of the network, the hidden state of the last moment obtained by the LSTM neural network training is extracted as the initial hidden state of the SRU neural network. Effectively improve the learning efficiency of the network and achieve higher accuracy.

[0048] The two-layer time block output in this section The subsequent calculation process is as follows: ; ; ; The final data is passed through the fully connected layer and the ELU activation function to obtain the final output.

[0049] ; Step 313: Anomaly detection classification method; The anomaly detection methods are: According to the current unsupervised anomaly detection method, the error obtained by prediction is used as the basis for classification judgment, and the maximum value of the predicted loss on the validation data set is used as the anomaly detection threshold. In the test set, the error generated by the predicted result and the actual result is compared with the threshold obtained in the previous step to obtain the predicted classification result. Similarly, the prediction of abnormality is marked as 1, and the prediction of normality is marked as 0.

[0050] Step 32: Establish a multi-dimensional time series prediction model; like Figure 5 As shown, the multidimensional prediction model is: The model has a three-branch structure: Historical time series branch: This branch uses a dual-period Dlinear model. The model operation is simple and can well extract the historical time series features of the prediction sequence. Related sequence branch: This branch uses convolutional neural network to reduce the dimension of related time series data, splices historical time series, uses scaled reset matrix and scaled position encoding, and then passes through multi-head attention module and bidirectional cross attention module respectively to extract inter-time series features and internal features of each time series; The relevant real-time branch uses convolutional neural networks to achieve dimensionality reduction and final fully connected output, and then introduces real-time relevant sensor data for real-time correction; the outputs of the above three branches are added and output, and finally the prediction of multi-dimensional time series is completed.

[0051] Step 321: Establish a historical time series branch; The historical time series branches are: According to the prediction sequence X 1 According to different cycles t Perform periodic decomposition into trend series and short-term fluctuation series.

[0052] First of all X 1 Sequences were replicatedt-1 After the splicing is X 1 , X 2 , …, X t Width is cycle t The length is the sequence length N A two-dimensional matrix of length t The windows of are averaged and pooled according to formula 3.5 to obtain , then the periodic sequence is obtained by subtracting the trend sequence from the original sequence X s .

[0053] ; This process will be repeated to obtain the sequence after all period decomposition .

[0054] For all the sequences generated above, apply a single-layer linear neural network and sum them up: ; ; ； ； ； in represents the weight of the lth network input of the first branch, is the value by which the output is offset. The same linear layer is applied to the two periodic sequences and the trend sequence to improve the fitting effect of the historical sequence. The final summation is used to obtain the predicted sequence branch output.

[0055] Step 322: establishing multiple correlation sequence branches; The relevant sequence branches are: First, a convolutional neural network is used to classify the related sequences X 2 Dimensionality reduction can be performed to extract the correlation features between different time series and reduce noise and improve the robustness of the model. X 2 , pass it into the subsequent convolutional layer, assuming that Indicates l-1 The feature mapping result of the layer, then l The output is as follows: ; in * represents the convolution operation, represents the convolution kernel, represents the output of the lth layer of the convolutional network of the second branch, is the bias term. Since there are four convolutional layers, the output result is Represents four convolutional layers, and the specific parameters are shown in the following table.

[0056] Table 1 Convolutional layer parameters Convolutional Layer Number of convolution kernels Convolution stride Convolution kernel size Conv2d_1 8 3×3 7×7 Conv2d_2 128 3×3 7×7 Conv2d_3 128 3×3 7×7 Conv2d_4 8 3×3 7×7 Then for the convolutional layer output After performing the maximum pooling, the results are as follows: ; in is the input feature map, size is the size of the pooling window, and stride is the step size of the pooling window. Perform the maximum value operation on each size×size window in the input feature map and slide it with the step size stride, and finally get . Then it is compared with the predicted sequence in the feature dimension X 1 Stitching: .

[0057] Wherein, step S40 includes: The training set is used to train the model respectively, the model hyperparameters are adjusted according to the validation set, and the test set is used to test the sensor anomaly detection effect and prediction effect. After obtaining the final training model and prediction results, the anomaly detection effect shown in Table 2 and the prediction effect after anomaly detection shown in Table 3 are obtained; Table 2 Anomaly detection effect

[0058] Optional, prediction effect after anomaly detection by anomaly algorithm: Table 3 Prediction results after anomaly detection

[0059] Through the method of the present invention, it can be clearly seen that the abnormality detection and prediction effect is good and the accuracy is high, which solves the problem in the prior art that the abnormality of the rotary kiln related sensor cannot be judged and the related abnormalities affect production.

[0060] The example of the present invention uses 90,000 groups of data from a cement production enterprise, of which 70,000 groups of data with abnormal data removed are used to train the model, 10,000 groups of data with abnormal data removed are used to verify and determine hyperparameters, and the last 10,000 groups of labeled abnormal data are used as test sets for prediction verification. The algorithm of the present invention shows good prediction performance in all six types of data, with good prediction effect and high prediction accuracy, solving the problem in the prior art that the abnormality of the rotary kiln related sensor cannot be judged and the related abnormality affects production.

[0061] The present application also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps in any of the above-mentioned embodiments of the method for monitoring timing anomaly of a rotary kiln sensor.

[0062] The present application also provides an electronic device, such as Figure 6 As shown, the electronic device includes at least one processor 31 and at least one memory 32, at least one of the memories 32 stores program information, and at least one of the processors 31 reads the program information and executes the rotary kiln sensor timing abnormality monitoring method described in any one of the above method embodiments. The device may also include: an input device 33 and an output device 34. The processor 31, the memory 32, the input device 33 and the output device 34 can be connected in communication. The memory 32, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer executable programs and modules. The processor 31 executes various functional applications and data processing by running the non-volatile software programs, instructions and modules stored in the memory 32, that is, the rotary kiln sensor timing abnormality monitoring method provided by any of the above schemes is realized. The memory 32 may include a program storage area and a data storage area, wherein the program storage area may store an operating system and an application required by at least one function; the data storage area may store data created according to the use of the rotary kiln sensor timing abnormality monitoring method, etc. In addition, the memory 32 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 32 may optionally include a memory remotely arranged relative to the processor 31, and these remote memories may be connected to a device for executing the rotary kiln sensor timing anomaly monitoring method through a network. Examples of the above-mentioned network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network and a combination thereof. The input device 33 may receive input user clicks, and generate signal inputs related to user settings and function controls of the rotary kiln sensor timing anomaly monitoring method. The output device 34 may include a display device such as a display screen. When the one or more modules are stored in the memory 32, when they are run by the one or more processors 31, the rotary kiln sensor timing anomaly monitoring method in any of the above-mentioned method embodiments is executed.

[0063] As needed, the above technical solutions can be combined to achieve the best technical effect.

[0064] The above are only the principles and preferred embodiments of the present application. It should be noted that, for ordinary technicians in this field, on the basis of the principles of the present application, several other modifications can be made, which should also be considered as the protection scope of the present application.

Claims

1. A method for monitoring abnormal timing of a rotary kiln sensor, characterized in that: include: Selecting relevant variable data of the sensor according to the mechanism of cement clinker burning, wherein the relevant variable data is variable data related to predicting the quantity and strength of cement; Preprocessing the relevant variable data to obtain key variable data, dividing the key variable data into a training set and a test set as a data set for model input; Inputting the data set into a trained single-dimensional detection model to obtain anomaly detection results of the sensor; The data set is input into a trained multi-dimensional prediction detection model to obtain the prediction result of the sensor.

2. The rotary kiln sensor timing abnormality monitoring method according to claim 1 is characterized in that: The relevant variable data of the sensor is selected according to the mechanism of cement clinker burning, and the relevant variable data is variable data related to predicting the quantity and strength of cement, including: The relevant variable data of the sensor include smoke chamber temperature variable, C5A outlet temperature variable, C5A feed pipe temperature variable, C5B outlet temperature variable, C5B feed pipe temperature variable, and kiln head hood temperature variable; wherein, C5A represents a cyclone numbered C5A; C5B represents a cyclone numbered C5B.

3. The rotary kiln sensor timing abnormality monitoring method according to claim 1 is characterized in that: The key variable data is obtained after preprocessing the relevant variable data, including: The relevant variable data of the sensor are subjected to outlier elimination or marking, mean filtering, and homogenization processing to obtain the key variable data.

4. The rotary kiln sensor timing abnormality monitoring method according to claim 3 is characterized in that: The one-dimensional detection model includes: The time block structure is executed twice to decompose the one-dimensional time series data into the trend part and the short-term fluctuation part according to the periodicity; Use convolutional neural network modeling to extract the two-dimensional time series matrix converted from one-dimensional time series data, and use linear neural network to extract the time trend characteristics in the time series matrix; Use the LSTM-SRU algorithm to extract all sequence splicing and sum to achieve time series prediction results; A reasonable threshold is obtained based on the validation set test, and anomaly detection is performed using the reasonable threshold and the time series prediction result.

5. The rotary kiln sensor timing abnormality monitoring method according to claim 4 is characterized in that: The method of using convolutional neural network modeling to extract one-dimensional time series data into a two-dimensional time series matrix includes: Applying Fast Fourier Transform to mine short-term and long-term trends of sequences in the time dimension: ； ； ； Where: X 1D Represents the smoke chamber temperature variable, C5A outlet temperature variable, C5A feed pipe temperature variable, C5B outlet temperature variable, C5B feed pipe temperature variable, and kiln head hood temperature variable; A Represents the intensity of the frequency components of the above variables, with the largest intensity k The frequencies correspond to the most significant k The length of the cycle; this k Frequency express; The value of Represents a data sequence; k For each cycle represents, T represents the cycle length; the above formula is simplified to: ； The cycle is obtained by fast Fourier transform. The time series data is periodically decomposed according to different cycles, and decomposed into trend series and short-term fluctuation series, including: X 1D The most significant k The length of the cycle X The matrices are copied t-1 times and concatenated into X1, X2, ..., X t is a two-dimensional matrix with a width of period t and a length of sequence length n, and then t Average pooling of windows: ； in X is the input matrix, X t is the trend sequence obtained after pooling. Since the width and window scale are the same, all the output matrices obtained are restored to a one-dimensional sequence. i and j is the absolute position of the input matrix, p and q is the relative position of the input matrix to the pooling area; the periodic sequence is obtained by subtracting the trend sequence from the original sequence X s : ； The above process is repeated q times to obtain the sequence after all period decompositions. , fold the original one-dimensional time series data according to the corresponding period of the fluctuation sequence according to the following formula: ； Indicates that zero padding is performed at the end of the sequence so that the sequence length can be divided evenly; The value of is 1, 2, ... k ; After the above operations, a set of two-dimensional matrices corresponding to the corresponding cycles of the two-dimensional time series matrix is ​​obtained.

6. The rotary kiln sensor timing abnormality monitoring method according to claim 5 is characterized in that: In the method of using the LSTM-SRU algorithm to extract the concatenation and sum of all sequences to achieve the time series prediction result, the LSTM-SRU algorithm includes: The LSTM-SRU neural network calculates forgetting, input, and output gating by assigning different weights to the input data. By combining the improved LSTM network and the SRU network, the output of the LSTM network layer becomes the input of the SRU network layer. The dropout layer is introduced to randomly discard some data. The hidden state at the last moment obtained by LSTM neural network training is extracted as the initial hidden state of SRU neural network.

7. The rotary kiln sensor timing abnormality monitoring method according to claim 6 is characterized in that: The data set is input into the trained single-dimensional detection model to obtain the abnormality detection result of the sensor, including: according to the unsupervised abnormality detection method, the error obtained by prediction is used as the classification judgment basis, and the maximum value of the predicted loss on the verification data set is used as the abnormality detection threshold; in the test set, the error generated by the predicted result and the actual result is compared with the threshold obtained in the previous step to obtain the predicted classification result, and the predicted abnormal result is marked as 1, and the predicted normal result is marked as 0; The data set is input into the trained multi-dimensional prediction detection model to obtain the prediction result of the sensor. The multi-dimensional prediction detection model has a three-branch structure, and the three branches include: Historical time series branch, which uses a dual-period Dlinear model to extract the historical time series features of the prediction sequence; Related sequence branch: This branch uses a convolutional neural network to reduce the dimension of related time series data, splices historical time series, uses a scaled reset matrix and performs position encoding at different scales, and then passes through a multi-head attention module and a bidirectional cross attention module to extract inter-series features and internal features of each time series respectively; The relevant real-time branch uses convolutional neural networks to achieve dimensionality reduction and final fully connected output, and then introduces real-time relevant sensor data for real-time correction; The outputs of the above three branches are added and output to complete the prediction of multi-dimensional time series.

8. A computer-readable storage medium, characterized in that: The storage medium stores program information, and the program information is read and run by a computer to execute the rotary kiln sensor timing abnormality monitoring method according to any one of claims 1 to 7.

9. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, the steps of the rotary kiln sensor timing abnormality monitoring method described in any one of claims 1-7 are implemented.

10. An electronic device, characterized in that: The method comprises at least one processor and at least one memory, wherein at least one of the memories stores instruction information, and at least one of the processors reads the instruction information and executes the rotary kiln sensor timing anomaly monitoring method according to any one of claims 1 to 7.

Images