Deep learning-based cement pre-decomposition process anomaly detection method
The self-attention and LSTM_AE model constructed using deep learning methods solves the problem of low accuracy in anomaly detection during cement pre-decomposition, achieving real-time anomaly detection and improving the stability and efficiency of the production process.
Patent Information
- Application Number
- CN202410896347.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-05
- Publication Date
- 2026-01-06
AI Technical Summary
The accuracy of anomaly detection in the cement pre-decomposition process is not high in the existing technology, which affects the production process.
We employ deep learning methods to construct an anomaly detection model using a self-attention mechanism and an LSTM_AE module. Combined with data preprocessing steps such as outlier removal, mean filtering, and normalization, we utilize reconstruction error for anomaly detection.
It enables real-time anomaly detection in the cement pre-decomposition process, improves the accuracy of anomaly judgment, and reduces adverse effects on the production process.
Smart Images
Figure CN121278433A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of artificial intelligence technology and relates to time series data processing, specifically a method for detecting time series anomalies in the cement pre-decomposition process. Background Technology
[0002] With the development of the social economy and the rapid growth of the construction industry, the demand for cement is also showing a continuous upward trend. As a basic material in the construction industry, the production efficiency and quality of cement are directly related to global infrastructure construction and economic development. The cement production process involves multiple stages, among which the cement pre-decomposition process is a crucial link in the entire production chain. The complexity of this process and the high-temperature environment make the monitoring and control of the pre-decomposition process very challenging, making monitoring and control even more difficult.
[0003] In recent years, deep learning algorithms have been widely used in time series anomaly detection and prediction, with representative models including self-supervised learning and unsupervised learning. With continuous technological advancements, these models have performed admirably when handling complex time series data, but they also face some challenges. Self-supervised learning methods, through pre-training tasks, enable the model to learn features effectively without labeled data. This is particularly important for time series data, as labeled anomaly data is often scarce. Unsupervised learning methods, on the other hand, utilize the inherent structure of the data for anomaly detection, reducing reliance on manual labeling. For large amounts of unlabeled industrial time series data, unsupervised learning, the relationships between variables, and long-term time dependencies have become problems that researchers need to solve. Summary of the Invention
[0004] In view of the shortcomings of the prior art described above, the purpose of this invention is at least to provide a method for detecting anomalies in the cement pre-decomposition process based on deep learning, which aims to solve the problem that it is difficult to judge anomalies in the cement pre-decomposition process and that related anomalies affect production in the prior art.
[0005] To achieve the above and other related objectives, one embodiment of the present invention provides a method for detecting anomalies in the cement pre-decomposition process, comprising the steps of:
[0006] Optionally, the preliminary steps for detecting anomalies in the cement pre-decomposition process using deep learning methods include the following:
[0007] The outlet temperature of the precalciner refers to the temperature of the gas discharged from the cement precalciner at the outlet during the cement production process. Generally, changes in the outlet temperature affect the degree of decomposition and reaction efficiency of the materials inside the furnace, thus impacting cement quality and yield. Raw meal feed rate refers to the quantity of material fed into the cement production line during the cement production process. Precise control of the raw meal feed rate is crucial to ensure the production line receives the appropriate amount of raw materials. Excessive or insufficient raw meal feed will adversely affect the production process. The flue is located at the bottom of the precalciner process and at the tail of the rotary kiln. Abnormal flue temperature can indicate the degree of crusting in the flue, affecting the calcination of materials inside the kiln and the quality of the cement. The cone pressure of the C5A and C5B cyclone preheaters is an important indicator of ventilation within the preheaters. By monitoring changes in the cone pressure of each stage of the cyclone preheater, the ventilation status, fan damper status, explosion-proof damper status, and system leakage and blockage can be assessed in a timely manner. The pressure in the middle of the precalciner refers to the pressure inside the precalciner during the cement production process. Changes in pressure within the decomposer furnace reflect the gas flow and thermal balance inside, significantly impacting the stability and efficiency of the production process. The air volume supplied to the decomposer furnace via the tertiary air duct is controlled by a flow valve pre-installed on the duct, and the tertiary air pressure is related to the pulverized coal combustion status within the furnace to some extent. Relevant variables are selected as input data for anomaly detection.
[0008] Optionally, the data preprocessing steps for the selected parameters include:
[0009] The data is preprocessed by removing or marking outliers, mean filtering, and normalization steps, thus preparing the data for anomaly detection in the cement predecomposition process.
[0010] Optionally, the step of removing and marking outliers includes:
[0011] Based on expert rule base and guidance from on-site technical personnel, we manually remove offsets or drifts in abnormal data. These include: numerical jumps within a period of time, sudden increases or decreases in value, and sudden increases or decreases in the amplitude of fluctuations; data loss or outliers, such as data loss, data exceeding the range, or negative values; and slow response or failure, such as data remaining unchanged for a long time.
[0012] Optionally, the mean filtering formula for processing the selected parameter data is:
[0013]
[0014] In the formula, n is the length of the time window for mean filtering.
[0015] Optionally, the homogenization formula is:
[0016] x'=(x–x min ) / (x max -x min )
[0017] Optionally, the dataset can be divided as follows:
[0018] The collected multivariate data, after outlier removal, is used as the training and validation sets, while the outlier data is used as the test set.
[0019] Optionally, the anomaly detection model is:
[0020] First, a self-attention mechanism block is used to assign different attention weights to different parts of the input so that the model can focus on the information that is most helpful in solving the current task.
[0021] Then, the LSTM_AE module is used to build a data-driven model based on unlabeled data. This model learns by reconstructing the input data, reducing the reconstruction loss between the input and output.
[0022] Finally, a reasonable threshold is obtained based on the validation set test, and the threshold and prediction results are used for anomaly detection.
[0023] Optionally, the self-attention structure is:
[0024] In the self-attention mechanism, given an input sequence X = [x1, x2, ... x...] n ], where x i It is the i-th element of the input sequence. Usually, three new sequences are first obtained through linear transformation: the query sequence Q = [q1, q2, ... q... n The key sequence K = [k1, k2, ... k] n The sum sequence V = [v1, v2, ... v] n Next, by calculating the dot product (or other similarity measure) between the query sequence and the key sequence, the attention score matrix A is obtained, as shown below:
[0025]
[0026] Among them, Q i This indicates that the query is for the i-th element of sequence Q, K j d represents the j-th element of the key sequence K. k This represents the dimension of the query sequence and the key sequence.
[0027] The attention score matrix A is normalized using a softmax operation to obtain the attention weight matrix а, as shown below:
[0028]
[0029] Finally, the value sequence V is multiplied by the attention weight matrix a, and the weighted values at all positions are summed to obtain the final output representation, as shown in the following equation:
[0030]
[0031] The advantage of self-attention mechanisms lies in their ability to consider information from various positions within the input sequence, thus better capturing long-range dependencies. This makes them perform exceptionally well when processing diverse sequence data. Self-attention mechanisms have proven to be a highly effective modeling technique in tasks such as machine translation, language modeling, and text classification.
[0032] Optionally, the LSTM_AE block structure is as follows:
[0033] An autoencoder (AE) is an unsupervised learning method used to build data-driven models based on unlabeled data. These models learn by reconstructing input data, reducing reconstruction loss between input and output. An LSTM network consists of a series of memory units, each containing three gating units: an input gate, a forget gate, and an output gate.
[0034] The forget gate in LSTM is a key component for controlling the flow of information. Its role is to determine whether to retain information from past memories and how much of those memories to retain for use at the current time step. Specifically, the forget gate uses a sigmoid activation function to output a value between 0 and 1, representing the amount of past memory to retain. If the gate's output is close to 0, it means past memories are forgotten; if it's close to 1, it means past memories are retained. The calculation method for the forget gate is as follows:
[0035] f t =σ(W f ·[h t-1 ,X t ]+b f )#
[0036] Among them, f t This refers to the output of the forget gate, W. f and b f These are the weight matrix and bias vector of the forget gate, h t-1 It is the hidden state of the previous time step, x t σ is the input at the current time step, and σ is the sigmoid activation function.
[0037] An input gate is a key component used to control the influence of the input information at the current time step on the current state unit. The input gate determines which parts of the input information at the current time step will be added to the memory unit. The input gate outputs a value between 0 and 1 using a sigmoid activation function, representing the importance of each input dimension. Then, a tanh activation function produces a value between -1 and 1, representing candidate values for the input information. Finally, the output of the input gate is multiplied by the candidate values to obtain the information to be added to the memory unit. The calculation method for the input gate is as follows:
[0038] i t =σ(W i .[h t-1 ,X t ]+b i )#
[0039]
[0040] Among them, i t It is the output of the input gate. Candidate values for the input information at the current time step, W i W c b i b c It is the weight matrix and bias vector of the input gate and candidate values, h t-1 It is the hidden state of the previous time step, X t This is the input at the current time step. Finally, the contents of the memory cell are updated by multiplying the output of the input gate by the candidate value and adding it to the memory cell.
[0041]
[0042] The output gate determines how the information in the memory cell affects the hidden state at the current time step and outputs it as the network's output. Specifically, the output gate uses a sigmoid activation function to output a value between 0 and 1, representing the degree of influence of the information in the memory cell on the hidden state. Then, the contents of the memory cell are passed through a tanh activation function to generate a value between -1 and 1, representing the current state of the memory cell. Finally, the output of the output gate is multiplied by the current state of the memory cell to obtain the hidden state at the current time step. The calculation method for the output gate is as follows:
[0043] O t =o(W o .[h t-1 ,X t ]+b o )#
[0044] h t=O t .tanh(C t )#
[0045] Among them, O t It is the output of the output gate, W o and b o These are the weight matrix and bias vector of the output gate. The product of the output gate's output and the current state of the memory cell controls how the information in the memory cell affects the hidden state at the current time step.
[0046] Optionally, the anomaly detection method is:
[0047] Based on current unsupervised anomaly detection methods, this approach uses the error between the reconstructed data and the actual data as the classification criterion, and the maximum error on the validation dataset as the anomaly detection threshold. In the test set, the anomaly detection result is obtained by comparing the error between the reconstructed result and the actual result with the threshold obtained in the previous step.
[0048] The method of this invention can solve the problem in the prior art where the accuracy of abnormal judgment in the cement pre-decomposition process is not high, and related abnormalities affect production.
[0049] This invention addresses the problem of anomaly detection in the cement pre-decomposition process. Based on deep learning, an anomaly detection model for the cement pre-decomposition process is designed, enabling real-time anomaly detection in the cement pre-decomposition process. This solves the problem in existing technologies where the accuracy of anomaly detection in the cement pre-decomposition process is low, and related anomalies affect production. Attached Figure Description
[0050] Figure 1 This is a flowchart of an anomaly detection method for cement pre-decomposition process based on deep learning, as described in this invention.
[0051] Figure 2 This is a structural diagram of a deep learning-based abnormal detection model for cement pre-decomposition process as described in this invention.
[0052] Figure 3 This is an LSTM structure diagram of a deep learning-based method for detecting anomalies in the cement pre-decomposition process, as described in this invention. Detailed Implementation
[0053] The present invention will be further described below with reference to the accompanying drawings and embodiments:
[0054] To clearly illustrate the technical features of the present invention, the present invention will be described in detail below through specific embodiments and in conjunction with the accompanying drawings.
[0055] like Figure 1 As shown, the present invention provides a method for detecting anomalies in the cement pre-decomposition process, comprising the following steps:
[0056] Step 1: Preliminary selection of relevant parameters based on the mechanism of cement clinker calcination;
[0057] Step 2: Data preprocessing is performed by removing outliers, mean filtering, and normalization, and finally the training set, validation set, and test set are divided.
[0058] Step 3: Input the dataset into the cement pre-decomposition process anomaly detection model, train the model, and then perform anomaly detection in the cement pre-decomposition process.
[0059] Step 4: Anomaly detection in the cement pre-decomposition process was achieved, solving the problem that the accuracy of anomaly detection in the cement pre-decomposition process was not high in the existing technology, and that related anomalies affected production.
[0060] Step 1: Preliminary selection of relevant parameters based on the mechanism of cement clinker calcination;
[0061] First, from a process perspective, we analyze the parameters of raw materials, clinker control parameters, and clinker mineral composition. The decomposition furnace outlet temperature refers to the temperature of the gas discharged from the cement pre-decomposition furnace at the outlet during cement production. Generally, changes in the decomposition furnace outlet temperature affect the degree of decomposition and reaction efficiency of the materials inside the furnace, thus affecting the quality and yield of cement. The raw material feed rate refers to the amount of material fed into the cement production line during the cement production process. Precise control of the raw material feed rate is necessary to ensure the production line receives an appropriate amount of raw materials. Too much or too little raw material feed will adversely affect the production process. The smoke chamber is located at the bottom of the entire pre-decomposition process and at the tail of the rotary kiln. Abnormal smoke chamber temperature can indicate the degree of crusting in the smoke chamber, affecting the calcination of materials inside the kiln and the quality of cement. The cone pressure of the C5A and C5B cyclone preheaters is an important indicator reflecting the ventilation within the preheater. By monitoring changes in the cone pressure of each stage of the cyclone preheater, the ventilation status, fan damper status, explosion-proof damper status, and system leakage and blockage can be determined in a timely manner. The pressure in the middle of the decomposition furnace refers to the pressure inside the decomposition furnace during cement production. Changes in the pressure in the middle of the decomposition furnace reflect the gas flow and thermal balance inside, significantly impacting the stability and efficiency of the production process. The air volume supplied to the decomposition furnace through the tertiary air duct is controlled by a flow valve pre-installed on the tertiary air duct, and the tertiary air pressure is related to the pulverized coal combustion status within the decomposition furnace to some extent. Relevant variables are selected as input data for anomaly detection.
[0062] Optionally, the data preprocessing steps for the selected parameters include:
[0063] Step 21: Process the cement pre-decomposition process dataset by removing outliers from the training and validation sets;
[0064] Step 22: Perform mean filtering on the cement pre-decomposition process dataset;
[0065] Step 23: Homogenize the cement pre-decomposition process dataset;
[0066] Step 24: Divide the cement pre-decomposition process dataset into a training set, a validation set, and a test set;
[0067] The following is a specific implementation plan for removing and marking outliers:
[0068] Step 21: Mark and remove outliers from the cement pre-decomposition process dataset;
[0069] Based on expert rule base and guidance from on-site technical personnel, we manually remove offsets or drifts in abnormal data. These include: numerical jumps within a period of time, sudden increases or decreases in value, and sudden increases or decreases in the amplitude of fluctuations; data loss or outliers, such as data loss, data exceeding the range, or negative values; and slow response or failure, such as data remaining unchanged for a long time.
[0070] Step 22: Perform mean filtering on the cement pre-decomposition process dataset;
[0071] The formula for mean filtering the parameter data after removing or marking outliers is:
[0072]
[0073] In the formula, m is the length of the time window for mean filtering.
[0074] Step 23: Homogenize the cement pre-decomposition process dataset;
[0075] The formula for normalizing the data after mean filtering is:
[0076] x'=(xx min ) / (x max -x min )
[0077] Step 24: Divide the homogenized dataset into a training set, a validation set, and a test set;
[0078] The collected multivariate data, after outlier removal, is used as the training and validation sets, while the outlier data is used as the test set.
[0079] Step 31: Establish an anomaly detection model;
[0080] Step 311: Establish the structure of the self-attention mechanism;
[0081] Step 312: Establish the LSTM_AE structure;
[0082] Step 313: Anomaly detection classification method;
[0083] The anomaly detection model is:
[0084] First, a self-attention mechanism block is used to assign different attention weights to different parts of the input so that the model can focus on the information that is most helpful in solving the current task.
[0085] Then, the LSTM_AE module is used to build a data-driven model based on unlabeled data. This model learns by reconstructing the input data, reducing the reconstruction loss between the input and output.
[0086] Finally, a reasonable threshold is obtained based on the validation set test, and the threshold and prediction results are used for anomaly detection.
[0087] Step 311: Establish the structure of the self-attention mechanism;
[0088] In the self-attention mechanism, given an input sequence X = [x1, x2, ... x...] n ], where x i It is the i-th element of the input sequence. Usually, three new sequences are obtained first through linear transformation: the query sequence Q = [q1, q2, ... q... n The key sequence K = [k1, k2, ... k] n The sum sequence V = [v1, v2, ... v] n Next, by calculating the dot product (or other similarity measure) between the query sequence and the key sequence, the attention score matrix A is obtained, as shown below:
[0089]
[0090] Among them, Q i This indicates that the query is for the i-th element of sequence Q, K. j d represents the j-th element of the key sequence K. k This represents the dimension of the query sequence and the key sequence.
[0091] The attention score matrix A is normalized using a softmax operation to obtain the attention weight matrix а, as shown below:
[0092]
[0093] Finally, the value sequence V is multiplied by the attention weight matrix a, and the weighted values at all positions are summed to obtain the final output representation, as shown in the following equation:
[0094]
[0095] The advantage of self-attention mechanisms lies in their ability to consider information from various positions within the input sequence, thus better capturing long-range dependencies. This makes them perform exceptionally well when processing diverse sequence data. In tasks such as machine translation, language modeling, and text classification, self-attention mechanisms have proven to be a highly effective modeling technique.
[0096] Step 312: Establish the LSTM_AE structure;
[0097] The LSTM_AE block structure is as follows:
[0098] An autoencoder (AE) is an unsupervised learning method used to build data-driven models based on unlabeled data. These models learn by reconstructing input data, reducing reconstruction loss between input and output. An LSTM network consists of a series of memory units, each containing three gating units: an input gate, a forget gate, and an output gate.
[0099] The forget gate in LSTM is a key component for controlling the flow of information. Its role is to determine whether to retain information from past memories and how much of those memories to retain for use at the current time step. Specifically, the forget gate uses a sigmoid activation function to output a value between 0 and 1, representing the amount of past memory to retain. If the gate's output is close to 0, it means past memories are forgotten; if it's close to 1, it means past memories are retained. The calculation method for the forget gate is as follows:
[0100] f t =σ(W f .[h t-1 ,X t ]+b f )#
[0101] Among them, f t This refers to the output of the forget gate, W. f and b f These are the weight matrix and bias vector of the forget gate, h t-1 It is the hidden state of the previous time step, x t σ is the input at the current time step, and σ is the sigmoid activation function.
[0102] An input gate is a key component used to control the influence of the input information at the current time step on the current state unit. The input gate determines which parts of the input information at the current time step will be added to the memory unit. The input gate outputs a value between 0 and 1 using a sigmoid activation function, representing the importance of each input dimension. Then, a tanh activation function produces a value between -1 and 1, representing candidate values for the input information. Finally, the output of the input gate is multiplied by the candidate values to obtain the information to be added to the memory unit. The calculation method for the input gate is as follows:
[0103] i t =σ(W i .[h t-1 ,X t ]+b i )#
[0104]
[0105] Among them, i t It is the output of the input gate. Candidate values for the input information at the current time step, W i W c b i b c It is the weight matrix and bias vector of the input gate and candidate values, h t-1 It is the hidden state of the previous time step, X t This is the input at the current time step. Finally, the contents of the memory cell are updated by multiplying the output of the input gate by the candidate value and adding it to the memory cell.
[0106]
[0107] The output gate determines how the information in the memory cell affects the hidden state at the current time step and outputs it as the network's output. Specifically, the output gate uses a sigmoid activation function to output a value between 0 and 1, representing the degree of influence of the information in the memory cell on the hidden state. Then, the contents of the memory cell are passed through a tanh activation function to generate a value between -1 and 1, representing the current state of the memory cell. Finally, the output of the output gate is multiplied by the current state of the memory cell to obtain the hidden state at the current time step. The calculation method for the output gate is as follows:
[0108] O t =σ(W o .[h t-1 ,X t ]+b o )#
[0109] h t=O t .tanh(C t )#
[0110] Among them, O t It is the output of the output gate, W o and b o These are the weight matrix and bias vector of the output gate. The product of the output gate's output and the current state of the memory cell controls how the information in the memory cell affects the hidden state at the current time step.
[0111] Step 313: Anomaly detection classification method;
[0112] The anomaly detection method is as follows:
[0113] Based on current unsupervised anomaly detection methods, this approach uses the error between the reconstructed data and the actual data as the classification criterion, and the maximum error on the validation dataset as the anomaly detection threshold. In the test set, the anomaly detection result is obtained by comparing the error between the reconstructed result and the actual result with the threshold obtained in the previous step.
[0114] Step 4: Input the dataset into different models, train the models, and then perform anomaly detection and prediction in the cement pre-decomposition process; the training set is used to train the models separately, the model hyperparameters are adjusted according to the validation set, and the test set is used to verify the anomaly detection effect in the cement pre-decomposition process.
[0115] Step 5: After obtaining the final trained model and prediction results, the detection effect and prediction effect are obtained;
[0116] The method of this invention can solve the problem in the prior art where the accuracy of abnormal judgment in the cement pre-decomposition process is not high, and related abnormalities affect production.
[0117] This invention addresses the problem of anomaly detection in the cement pre-decomposition process. Based on deep learning, an anomaly detection model for the cement pre-decomposition process is designed, enabling real-time anomaly detection in the cement pre-decomposition process. This solves the problem in existing technologies where the accuracy of anomaly detection in the cement pre-decomposition process is low, and related anomalies affect production.
Claims
1. A deep learning based pre-decomposition process anomaly detection method, characterized in that, Comprise: Step 1: selection of relevant parameters according to the mechanism of the cement pre-decomposition process; Step 2: data preprocessing by homogenization and other operations, dividing the determined key variables into training set, validation set and test set as model input; Step 3: input the data set into the anomaly detection model, train the model and then perform anomaly detection on the pre-decomposition process; Step 4: obtain the detection results of the cement pre-decomposition process through the model, solve the problem of difficult pre-decomposition process anomaly judgment and related abnormality affecting production in the prior art.
2. The method of claim 1, wherein the method is based on deep learning. The preliminary step of obtaining anomaly detection related parameters through the cement pre-decomposition process comprises: Firstly, from the aspects of raw material parameters, clinker index control parameters, clinker mineral composition, etc., analyze from the process point of view, the decomposition furnace outlet temperature refers to the temperature of the gas discharged from the cement pre-decomposition furnace at the outlet in the cement production process. Under normal circumstances, the change of the decomposition furnace outlet temperature will affect the decomposition degree and reaction efficiency of the materials in the furnace, and then affect the quality and yield of the cement. The raw material feeding amount refers to the amount of raw material fed into the cement production line during the cement production process. The raw material feeding amount needs to be accurately controlled in the cement production process to ensure that the production line can obtain an appropriate amount of raw materials for production. If the raw material feeding amount is too much or too little, it will adversely affect the production process. The smoke chamber is located at the bottom of the entire pre-decomposition process and the tail of the rotary kiln, and the abnormality of the smoke chamber temperature can indicate the skinning condition of the smoke chamber, which affects the calcination of the materials in the kiln and the quality of the cement. The C5A and C5B cyclone preheater cone pressure is an important indicator reflecting the ventilation condition in the preheater. By monitoring the change of the cone pressure of each stage of the cyclone preheater, the ventilation condition inside the system, the opening and closing state of the fan gate, the opening and closing state of the explosion-proof damper, and the system air leakage and blockage condition can be judged in time. The decomposition furnace middle pressure refers to the pressure condition inside the decomposition furnace during the cement production process. The change of the decomposition furnace middle pressure can reflect the gas flow and heat balance inside the decomposition furnace, which has important influence on the stability and efficiency of the production process. The amount of air sent into the decomposition furnace by the tertiary air pipe is controlled by the flow valve pre-installed on the tertiary air pipe, and the tertiary air pressure is related to the coal powder combustion condition in the decomposition furnace to a certain extent. The above variables have a greater relationship with the abnormal working condition of the pre-decomposition process, so they are selected as input data for multi-dimensional anomaly detection. 3.The deep learning based pre-decomposition process anomaly detection method of claim 1, wherein, The data preprocessing step of each parameter selected in claim 2 comprises: Homogenization step for predicting the time series data of the rotary kiln related sensors, preprocessing the data to prepare for the pre-decomposition process anomaly detection.
4. The deep learning based pre-decomposition process anomaly detection and root cause analysis method of claim 3, wherein, The methods for removing or marking abnormal values are: According to the expert rule base and the guidance of the field technical personnel, the offset or drift existing in the abnormal data is removed manually, which is manifested as a sudden jump in the value within a period of time, a sudden increase or decrease in the value, a sudden increase or decrease in the fluctuation amplitude; data loss or abnormal value is manifested as data loss or data exceeding the range, negative value, etc.; response delay or failure is manifested as the data remaining unchanged for a long time, etc.
5. The method of claim 3, wherein the method is based on deep learning. The mean filtering formula for processing the selected parameter data is: where n is the time window length of the mean filtering.
6. The method of claim 3, wherein the method is based on deep learning. The normalization formula is: x' = (x - x min ) / (x max -x min ) 7. The method of claim 3, wherein the method is based on deep learning. The data set is divided into: The collected multivariate data is used as the training set and the validation set after removing the abnormal data, and the abnormal data is used as the test set. 8.The deep learning based pre-decomposition process anomaly detection method of claim 1, wherein, The abnormal detection model is: First, use the self-attention mechanism block to assign different attention weights to different parts of the input, so that the model can focus on the information that is most helpful for solving the current task. Then use the LSTM_AE module to build a data-driven model based on unlabeled data, which learns by reconstructing the input data and reducing the reconstruction loss between the input and output. Finally, according to the test of the validation set, a reasonable threshold is obtained, and finally the threshold and the prediction result are used for abnormal detection.
9. The deep learning-based pre-decomposition process anomaly detection method of claim 8, wherein, The self-attention mechanism block is: In the self-attention mechanism, given an input sequence X = [x1, x2, ... x...] n ], where x i It is the i-th element of the input sequence. Usually, three new sequences are obtained first through linear transformation: the query sequence Q = [q1, q2, ... q... n The key sequence K = [k1, k2, ... k] n The sum sequence V = [v1, v2, ... v] n Next, by calculating the dot product (or other similarity measure) between the query sequence and the key sequence, the attention score matrix A is obtained, as shown below: where Q i represents the i-th element of the query sequence Q, K j represents the j-th element of the key sequence K, d k represents the dimension of the query sequence and the key sequence. The attention score matrix A is normalized by the softmax operation to obtain the attention weight matrix a, as follows: Finally, multiply the value sequence V by the attention weight matrix a, and sum the weighted values of all positions to obtain the final output representation, as follows: The advantage of self-attention mechanism is that it can consider the information of each position in the input sequence, so it can better capture the long-distance dependencies in the sequence. This makes it perform well in handling various sequence data. In machine translation, language modeling, text classification and other tasks, self-attention mechanism has been proven to be a very effective modeling technique.
10. The deep learning based pre-decomposition process anomaly detection method of claim 8, wherein, The LSTM_AE module is: An autoencoder (AE) is an unsupervised learning method used to build a data-driven model based on unlabeled data. This model learns by reconstructing the input data and reducing the reconstruction loss between the input and output. The LSTM network consists of a series of memory cells, each containing three gating units: input gate, forget gate and output gate. The forget gate in LSTM is a key component for controlling information flow. The role of the forget gate is to decide whether to retain information in the past memory and decide how much past memory to retain for use at the current time step. Specifically, the forget gate outputs a value between 0 and 1 through a sigmoid activation function, indicating the amount of past memory to be retained. If the output of the gate is close to 0, it means forgetting the past memory; if it is close to 1, it means retaining the past memory. The calculation of the forget gate is as follows: f t = σ(W f· [h t-1 , X t ]+b f ) where f t is the output of the forget gate, W f and b f are the weight matrix and bias vector of the forget gate, h t-1 is the hidden state of the previous time step, x t is the input of the current time step, and σ is the sigmoid activation function. The input gate is a key component for controlling the influence of the input information at the current time step on the current state cell. The role of the input gate is to decide which parts of the input information at the current time step will be added to the memory cell. The input gate outputs a value between 0 and 1 through a sigmoid activation function, indicating the importance of each input dimension. Then, a tanh activation function is used to produce a value between -1 and 1, representing the candidate value of the input information. Finally, multiply the output of the input gate with the candidate value to get the information to be added to the memory cell. The calculation of the input gate is as follows: i t = σ(W i· [h t-1 , X t ]+b i ) where i t is the output of the input gate, the candidate value of the input information at the current time step, W i , W c , b i , b c is the weight matrix and bias vector of the input gate and the candidate value, h t-1 is the hidden state at the previous time step, X t is the input at the current time step, and finally, the content of the memory cell is updated by multiplying the output of the input gate with the candidate value and adding it to the memory cell: The output gate determines how the information in the memory cell affects the hidden state of the current time step and outputs it as the output of the network. Specifically, the output gate outputs a value between 0 and 1 through a sigmoid activation function, representing the degree of influence of the information in the memory cell on the hidden state. Then, the content of the memory cell is passed through a tanh activation function to generate a value between -1 and 1, representing the current state of the memory cell. Finally, the output of the output gate is multiplied by the current state of the memory cell to obtain the hidden state of the current time step. The calculation method of the output gate is as follows: O t = σ(W o· [h t-1 , X t ] + b o ) h t =O t· tanh(C t ) where O t is the output of the output gate, W o and b o are the weight matrix and bias vector of the output gate. The product of the output of the output gate and the current state of the memory cell controls how the information in the memory cell influences the hidden state at the current time step.
11. The deep learning based pre-decomposition process anomaly detection method of claim 1, wherein, The abnormality detection method is: According to the current unsupervised anomaly detection method, the error obtained by comparing the reconstructed data with the actual data is used as the classification judgment basis, and the maximum error on the validation dataset is used as the anomaly detection threshold. In the test set, the error between the reconstruction result and the actual result is compared with the threshold obtained in the last step to obtain the abnormality detection result.
12. The deep learning based pre-decomposition process anomaly detection method of claim 1, wherein, Train the model and obtain the threshold for anomaly detection using the validation set: The training set is used to train the model respectively, and the model hyperparameters are adjusted according to the validation set to obtain the related threshold.
13. The method of claim 1, wherein, Effect verification: The test set is used to verify the sensor anomaly detection effect and prediction effect.