Dry quenching CO online prediction method based on distributed calculation mixing GRU

Abstract

The invention discloses a dry quenching CO online prediction method based on distributed computing hybrid GRU. The method comprises the following steps: S1, data preprocessing; s2, data batching and distribution; s3, carrying out distributed training; according to the method, big data is efficiently managed and organized through Hive, rapid data set construction and query are achieved, data parallel distributed processing is conducted through MapReduce, high parallelization of deep learning model training and efficient resource utilization are achieved, and the training speed and the model performance are remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and big data computing technology, specifically to an online prediction method for CO in dry quenching coke based on a distributed computing hybrid GRU. Background Technology

[0002] In industrial big data applications, single-machine deep learning models suffer from slow data processing speed, long model training cycles, and low utilization of computing resources when dealing with massive amounts of data. Meanwhile, traditional distributed deep learning methods have drawbacks such as difficulty in ensuring data consistency, complex model training processes, and poor ensemble performance. Summary of the Invention

[0003] The purpose of this invention is to provide an online prediction method for CO in dry quenching based on a distributed computing hybrid GRU, so as to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an online prediction method for CO in dry quenching based on a distributed computing hybrid GRU, wherein the prediction method specifically includes: S1, Data Preprocessing; S2, Data batching and distribution; S3, distributed training.

[0005] Preferably, step S1 specifically includes the following steps: a1. Data Acquisition: Real-time data acquisition of PLC point data from the automation program in the dry quenching primary subsystem is achieved using OpenPlant, forming a real-time database; the PLC point data includes: pre-storage chamber pressure ( ), coke loading ( ), coking temperature ( ), coke discharge ( ), coke discharge temperature ( ), CO concentration of the combustible component in the circulating gas ( ), combustible components of circulating gas concentration( ), combustible components of circulating gas concentration( ), combustible components of circulating gas concentration( ), combustible components of circulating gas concentration( ), air intake ( ), amount of gas released ( Bypass flow () ), current furnace number ( ); a2. Noise identification and initial screening; a3. Time series Take a fixed number of values ​​from the real-time denoising backup database according to different processing frequencies and apply the following smoothing or denoising kernels; a4. Align date-granularity data shards with labels: Perform hourly sharding on the real-time smooth backup database at fixed daily times, generating... And at fixed points, cleanup operations are performed on the real-time denoised backup database and the real-time smooth backup database, for each slice. Verify the data feature dimensions, verify the continuity of the time series timestamps, and verify whether the data is null. Stored in a MySQL transactional database.

[0006] Preferably, step a2 specifically includes: a21. Time series data collected from the real-time database are aligned for timestamp consistency and imputed for missing data: Data is collected at a fixed frequency of 500ms and the timestamps of each point are verified to be consistent. Then, the timestamps of each point are verified to be within the threshold of 1000ms and aligned with the latest date. Missing points are filled with the previous value. If the value exceeds the threshold, it is written to the warning log. Then, calculation is started once after collecting data at every two fixed frequencies of the array. The average value at the two frequencies is calculated and written to the queue. a22、 - The 13 real-time data points are stored in a fixed-length queue, and the median of the computation data is taken based on the computation queue using Hampel filtering. Outlier detection, identification, and removal based on median absolute deviation. The outlier points, where t is [1, 2, …, 13], The moving median. Different points take different Value, that is ; a23. Store the 13 calculated values ​​into the real-time noise reduction backup database at a fixed frequency of one second.

[0007] Preferably, step a3 specifically includes: a31, Targeting A simple moving average is used, and the processing method for the simple moving average is as follows: the smoothed value at time t is defined as: in, To smooth the window length, 60; a32, Targeting Using an exponentially weighted average, given a decay factor of 0 < <1, there is in, The weights decay exponentially, with data closer to the current moment having a greater weight. The value is 0.1; a33, targeting A polynomial fitting method is used: within a sliding window of length 2m+1, a polynomial of degree p is performed on the sequence; the solution is obtained under the following conditions: window length 2m+1, polynomial order p. Among them, the center is located in The window determines the coefficient vector. m is 5, the window length is 11 points, and p is 2; a34, the smoothed data and Writes to the real-time smooth backup database at a fixed frequency of one second.

[0008] Preferably, step S2 specifically includes: b1. Data preprocessing and generation of batch datasets; b2. Data storage and training set generation.

[0009] Preferably, step b1 specifically includes: b11. Assume the sampling step size is fixed. The record sequence is as follows: in Number of sensor channels The value is 14, and the sampling step size T is 2048. Data has been standardized: in, ; b12. Input-output sample construction: Given the backtracking length Predicted offset and target statistics window For each index structure: Where P is 14. The sampling step size is 2048. For the data after standardization in step b11, with dimensions set to (2048, 14), predict the offset value. Predicted offset value for the time interval between dry quenching and coking. The time difference is 512, and the statistics window is... The time interval between the two dry-quenching coke ovens and the coke charging is 2. The time difference is 1024. This is the mean function calculated based on the target value.

[0010] Preferably, step b2 specifically includes: Let the preprocessed batch training sample set be: ; b21. Partition-Bucket Mapping: in, For timestamp integers and For the daily dry quenching coke oven number The hash value modulo 4 As the partition step size, partition by hours, and use 3600. The number of buckets, with a value of 4; After mapping, the physical subset is obtained: ; b22, each according to: PARTITION(p) CLUSTERED BY(b)INTO K BUCKETS Write to a Hive table to form: in, This represents a non-overlapping union.

[0011] Preferably, step S3 specifically includes: c1. Set model parameters: Batch: Timing length: Feature dimension: Window half-width: Sliding step size: 16. Local GRU hidden dimension: Global GRU hidden dimension: Decoder GRU hidden dimension: Internal dimension of attention: , State vector dimension: Number of server data nodes: ( ), number of effective windows ; c2, take the first The first window, the... One effective window, centered on : in, Number of windows ; c3, each server node Each has a local GRU: ; c4. Collect the local output of all data server nodes: ; c5. Using additive attention on the server dimension A weighted fusion is performed to obtain a single local vector for that window. : in, This is the hidden state from the previous step. , , For attention parameters; Each server independently obtains forward pass data. Then, all-gather the computation to the compute node server and perform computation on the main process. and The aggregated local sequence is obtained as follows: ,Right now ; c6. Global encoder input: Uses the same sliding window size and step size as the local encoder input. For each window step... Each sliding window Constructing the global input: ; Global encoder calculation: As the length is Sequence training of global GRU: Update Gate: in, To globally update the gate output, For weights and biases; Reset Door: in, Reset the gate output globally; Candidate hidden state: in, This represents the global candidate hidden state. Hidden status update: in, This is the final globally hidden state; Full output: ,Right now ; c7. Window-level attention and context: Decoding steps Use the hidden state from the previous step of the decoder As a query, the aggregated local sequence Additive attention: in, It is a hidden state. , , For attention parameters; c8. State vector fusion: in, , ; c9. Using GRU decoding, each step's input typically includes a state vector and the previous step's output: The calculation process is as follows: Update Gate: in, To update the gate output, For weights and biases; Reset Door: Candidate hidden state: Hidden status update: in, Final hidden state: in, ; c10. Use MSE as the training loss and RMSE / MAE as the evaluation metric: RMSE = MAE = in, For the true value, These are the model's predicted values.

[0012] Preferably, step c3 specifically includes: one of the server nodes The calculation process is as follows: c31, Update Gate: in, To update the gate output, the value ranges from 0 to 1. To update the gate weight matrix, To update the gate bias term, The Sigmoid activation function is used. Let be the hidden state vector from the previous time step. This is the input vector at the current moment; c32, Reset the door: To reset the gate output, a value of 0 to 1 is used to control the proportion of forgotten historical information. To reset the gate weight matrix, To reset the door offset item; c33, Candidate Hidden State: in, Candidate hidden state vectors For candidate state weight matrix, For candidate state bias terms, For hyperbolic tangent activation function, This is element-wise multiplication; c34, Hidden state update: in, For the first The hidden state vector updated by each server node.

[0013] Compared with the prior art, the beneficial effects of the present invention are: This invention discloses an online CO (carbon monoxide) prediction method for dry quenching coke based on a distributed computing hybrid GRU, comprising S1, data preprocessing; S2, data batching and distribution; and S3, distributed training. This invention utilizes Hive for efficient management and organization of large datasets, enabling rapid dataset construction and querying. It employs MapReduce for parallel distributed data processing, achieving highly parallel training of deep learning models and efficient resource utilization, significantly improving training speed and model performance. By training local GRU models in parallel on multiple computing nodes and performing result fusion and parameter optimization at the global level, it achieves efficient modeling and prediction of large-scale time-series data. At the application level, the distributed hybrid GRU model is used for online prediction of CO (carbon monoxide) concentration in dry quenching (CDQ) systems. By sensing CO concentration change trends in advance, it provides decision-making basis for equipment control and operation optimization, thereby reducing energy consumption and burn-off, and improving economic efficiency and system intelligence. Attached Figure Description

[0014] Figure 1 The present invention is a denoised version after Hampel filtering and Savitzky-Golay filtering. Comparison images before and after processing; Figure 2 This is a model layer diagram of the present invention; Figure 3 This is a comparison chart of the CO prediction performance data of the present invention; Figure 4 This is a comparison chart of the CO prediction performance data of the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0017] In the description of this patent, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "setting" should be interpreted broadly. For example, they can refer to a fixed connection or setting, a detachable connection or setting, or an integrated connection or setting. Those skilled in the art can understand the specific meaning of the above terms in this patent according to the specific circumstances.

[0018] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a number" means two or more, unless otherwise explicitly specified.

[0019] Example Please see Figure 1-4 As shown, this invention provides a technical solution for online prediction of CO in dry quenching based on a distributed computing hybrid GRU: the prediction method specifically includes: S1, Data Preprocessing; a1. Data Acquisition: Real-time data acquisition of PLC point data from the automation program in the dry quenching primary subsystem is achieved using OpenPlant, forming a real-time database; the PLC point data includes: pre-storage chamber pressure ( ), coke loading ( ), coking temperature ( ), coke discharge ( ), coke discharge temperature ( ), CO concentration of the combustible component in the circulating gas ( ), combustible components of circulating gas concentration( ), combustible components of circulating gas concentration( ), combustible components of circulating gas concentration( ), combustible components of circulating gas concentration( ), air intake ( ), amount of gas released ( Bypass flow () ), current furnace number ( ); a2. Noise identification and initial screening; a21. Time series data collected from the real-time database are aligned for timestamp consistency and imputed for missing data: Data is collected at a fixed frequency of 500ms and the timestamps of each point are verified to be consistent. Then, the timestamps of each point are verified to be within the threshold of 1000ms and aligned with the latest date. Missing points are filled with the previous value. If the value exceeds the threshold, it is written to the warning log. Then, calculation is started once after collecting data at every two fixed frequencies of the array. The average value at the two frequencies is calculated and written to the queue. a22、 - The 13 real-time data points are stored in a fixed-length queue, and the median of the computation data is taken based on the computation queue using Hampel filtering. Outlier detection, identification, and removal based on median absolute deviation. The outlier points, where t is [1, 2, …, 13], The moving median. Different points take different Value, that is ; a23. Store the 13 calculated values ​​into the real-time noise reduction backup database at a fixed frequency of one second; a3. Time series Take a fixed number of values ​​from the real-time denoising backup database according to different processing frequencies and apply the following smoothing or denoising kernels; a31, Targeting A simple moving average is used, and the processing method for the simple moving average is as follows: the smoothed value at time t is defined as: in, To smooth the window length, 60; a32, Targeting Using an exponentially weighted average, given a decay factor of 0 < <1, there is in, The weights decay exponentially, with data closer to the current moment having a greater weight. The value is 0.1; a33, targeting A polynomial fitting method is used: within a sliding window of length 2m+1, a polynomial of degree p is performed on the sequence; the solution is obtained under the following conditions: window length 2m+1, polynomial order p. Among them, the center is located in The window determines the coefficient vector. m is 5, the window length is 11 points, and p is 2; a34, the smoothed data and Writes to the real-time smooth backup database at a fixed frequency of one second; a4. Align date-granularity data shards with labels: Perform hourly sharding on the real-time smooth backup database at fixed daily times, generating... And at fixed points, cleanup operations are performed on the real-time denoised backup database and the real-time smooth backup database, for each slice. Verify the data feature dimensions, verify the continuity of the time series timestamps, and verify whether the data is null. Stored in a MySQL transactional database; S2, Data batching and distribution; b1. Data preprocessing and generation of batch datasets; b11. Assume the sampling step size is fixed. The record sequence is as follows: in Number of sensor channels The value is 14, and the sampling step size T is 2048. Data has been standardized: in, ; b12. Input-output sample construction: Given the backtracking length Predicted offset and target statistics window For each index structure: Where P is 14. The sampling step size is 2048. For the data after standardization in step b11, with dimensions set to (2048, 14), predict the offset value. Predicted offset value for the time interval between dry quenching and coking. The time difference is 512, and the statistics window is... The time interval between the two dry-quenching coke ovens and the coke charging is 2. The time difference is 1024. This is the mean function calculated based on the target value; b2. Data storage and training set generation; Let the preprocessed batch training sample set be: ; b21. Partition-Bucket Mapping: in, For timestamp integers and For the daily dry quenching coke oven number The hash value modulo 4 As the partition step size, partition by hours, and use 3600. The number of buckets, with a value of 4; After mapping, the physical subset is obtained: ; b22, each according to: PARTITION(p) CLUSTERED BY(b)INTO K BUCKETS Write to a Hive table to form: in, This represents a non-overlapping union; S3, distributed training; c1. Set model parameters: Batch: Timing length: Feature dimension: Window half-width: Sliding step size: 16. Local GRU hidden dimension: Global GRU hidden dimension: Decoder GRU hidden dimension: Internal dimension of attention: , State vector dimension: Number of server data nodes: ( ), number of effective windows ; c2, take the first The first window, the... One effective window, centered on : in, Number of windows ; c3, each server node Each has a local GRU: ; One of the server nodes The calculation process is as follows: c31, Update Gate: in, To update the gate output, the value ranges from 0 to 1. To update the gate weight matrix, To update the gate bias term, The Sigmoid activation function is used. Let be the hidden state vector from the previous time step. This is the input vector at the current moment; c32, Reset the door: To reset the gate output, a value of 0 to 1 is used to control the proportion of forgotten historical information. To reset the gate weight matrix, To reset the door offset item; c33, Candidate Hidden State: in, Candidate hidden state vectors For candidate state weight matrix, For candidate state bias terms, For hyperbolic tangent activation function, This is element-wise multiplication; c34, Hidden state update: in, For the first The hidden state vector updated by each server node c4. Collect the local output of all data server nodes: ; c5. Using additive attention on the server dimension A weighted fusion is performed to obtain a single local vector for that window. : in, This is the hidden state from the previous step. , , For attention parameters; Each server independently obtains forward pass data. Then, all-gather the computation to the compute node server and perform computation on the main process. and The aggregated local sequence is obtained as follows: ,Right now ; c6. Global encoder input: Uses the same sliding window size and step size as the local encoder input. For each window step... Each sliding window Constructing the global input: ; Global encoder calculation: As the length is Sequence training of global GRU: Update Gate: in, To globally update the gate output, For weights and biases; Reset Door: in, Reset the gate output globally; Candidate hidden state: in, This represents the global candidate hidden state. Hidden status update: in, This is the final globally hidden state; Full output: ,Right now ; c7. Window-level attention and context: Decoding steps Use the hidden state from the previous step of the decoder As a query, the aggregated local sequence Additive attention: in, It is a hidden state. , , For attention parameters; c8. State vector fusion: in, , ; c9. Using GRU decoding, each step's input typically includes a state vector and the previous step's output: The calculation process is as follows: Update Gate: in, To update the gate output, For weights and biases; Reset Door: Candidate hidden state: Hidden status update: in, Final hidden state: in, ; c10. Use MSE as the training loss and RMSE / MAE as the evaluation metric: RMSE = MAE = in, For the true value, These are the model's predicted values.

[0020] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An online prediction method for CO in dry quenching based on a distributed computing hybrid GRU, characterized in that: The prediction method specifically includes: S1, Data Preprocessing; S2, Data batching and distribution; S3, distributed training.

2. The online prediction method for CO in dry quenching based on a distributed computing hybrid GRU according to claim 1, characterized in that: Step S1 specifically includes the following steps: a1. Data Acquisition: Real-time data acquisition of PLC point data from the automation program in the dry quenching primary subsystem is achieved using OpenPlant, forming a real-time database; the PLC point data includes: pre-storage chamber pressure ( ), coke loading ( ), coking temperature ( ), coke discharge ( ), coke discharge temperature ( ), CO concentration of the combustible component in the circulating gas ( ), combustible components of circulating gas concentration( ), combustible components of circulating gas concentration( ), combustible components of circulating gas concentration( ), combustible components of circulating gas concentration( ), air intake ( ), amount of gas released ( Bypass flow () ), current furnace number ( ); a2. Noise identification and initial screening; a3. Time series Take a fixed number of values ​​from the real-time denoising backup database according to different processing frequencies and apply the following smoothing or denoising kernels; a4. Align date-granularity data shards with labels: Perform hourly sharding on the real-time smooth backup database at fixed daily times, generating... And at fixed points, cleanup operations are performed on the real-time denoised backup database and the real-time smooth backup database, for each slice. Verify the data feature dimensions, verify the continuity of the time series timestamps, and verify whether the data is null. Stored in a MySQL transactional database.

3. The online prediction method for CO in dry quenching based on a distributed computing hybrid GRU according to claim 2, characterized in that: Step a2 specifically includes: a21. Time series data collected from the real-time database are aligned for timestamp consistency and imputed for missing data: Data is collected at a fixed frequency of 500ms and the timestamps of each point are verified to be consistent. Then, the timestamps of each point are verified to be within the threshold of 1000ms and aligned with the latest date. Missing points are filled with the previous value. If the value exceeds the threshold, it is written to the warning log. Then, calculation is started once after collecting data at every two fixed frequencies of the array. The average value at the two frequencies is calculated and written to the queue. a22、 - The 13 real-time data points are stored in a fixed-length queue, and the median of the computation data is taken based on the computation queue using Hampel filtering. Outlier detection, identification, and removal based on median absolute deviation. The outlier points, where t is [1, 2, …, 13], The moving median. Different points take different Value, that is ; a23. Store the 13 calculated values ​​into the real-time noise reduction backup database at a fixed frequency of one second.

4. The online prediction method for CO in dry quenching based on distributed computing hybrid GRU according to claim 2, characterized in that: Step a3 specifically includes: a31, Targeting A simple moving average is used, and the processing method for the simple moving average is as follows: the smoothed value at time t is defined as: in, To smooth the window length, 60; a32, Targeting Using an exponentially weighted average, given a decay factor of 0 < <1, there is in, The weights decay exponentially, with data closer to the current moment having a greater weight. The value is 0.1; a33, targeting A polynomial fitting method is used: within a sliding window of length 2m+1, a polynomial of degree p is performed on the sequence; the solution is obtained under the following conditions: window length 2m+1, polynomial order p. Among them, the center is located in The window determines the coefficient vector. m is 5, the window length is 11 points, and p is 2; a34, the smoothed data and Writes to the real-time smooth backup database at a fixed frequency of one second.

5. The online prediction method for CO in dry quenching based on a distributed computing hybrid GRU according to claim 1, characterized in that: Step S2 specifically includes: b1. Data preprocessing and generation of batch datasets; b2. Data storage and training set generation.

6. The online prediction method for CO in dry quenching based on a distributed computing hybrid GRU according to claim 5, characterized in that: Step b1 specifically includes: b11. Assume the sampling step size is fixed. The record sequence is as follows: in Number of sensor channels The value is 14, and the sampling step size T is 2048. Data has been standardized: in, ; b12. Input-output sample construction: Given the backtracking length Predicted offset and target statistics window For each index structure: Where P is 14. The sampling step size is 2048. For the data after standardization in step b11, with dimensions set to (2048, 14), predict the offset value. Predicted offset value for the time interval between dry quenching and coking. The time difference is 512, and the statistics window is... The time interval between the two dry-quenching coke ovens and the coke charging is 2. The time difference is 1024. This is the mean function calculated based on the target value.

7. The online prediction method for CO in dry quenching based on a distributed computing hybrid GRU according to claim 5, characterized in that: The specific steps of step b2 include: Let the preprocessed batch training sample set be: ； b21. Partition-Bucket Mapping: in, For timestamp integers and For the daily dry quenching coke oven number The hash value modulo 4 As the partition step size, partition by hours, and use 3600. The number of buckets, with a value of 4; After mapping, the physical subset is obtained: ； b22, each according to: PARTITION(p) CLUSTERED BY(b)INTO K BUCKETS Write to a Hive table to form: in, This represents a non-overlapping union.

8. The online prediction method for CO in dry quenching based on a distributed computing hybrid GRU according to claim 1, characterized in that: Step S3 specifically includes: c1. Set model parameters: Batch: Timing length: Feature dimension: Window half-width: Sliding step size:

16. Local GRU hidden dimension: Global GRU hidden dimension: Decoder GRU hidden dimension: Internal dimension of attention: , State vector dimension: Number of server data nodes: ( ), number of effective windows ; c2, take the first The first window, the... One effective window, centered on : in, Number of windows ; c3, each server node Each has a local GRU: ； c4. Collect the local output of all data server nodes: ； c5. Using additive attention on the server dimension A weighted fusion is performed to obtain a single local vector for that window. : in, This is the hidden state from the previous step. , , For attention parameters; Each server independently obtains forward pass data. Then, all-gather the computation to the compute node server and compute on the main process. and The aggregated local sequence is obtained as follows: ,Right now ; c6. Global encoder input: Uses the same sliding window size and step size as the local encoder input. For each window step... Each sliding window Constructing the global input: ； Global encoder calculation: As the length is Sequence training of global GRU: Update Gate: in, To globally update the gate output, For weights and biases; Reset Door: in, Reset the gate output globally; Candidate hidden state: in, This represents the global candidate hidden state. Hidden status update: in, This is the final globally hidden state; Full output: ,Right now ; c7. Window-level attention and context: Decoding steps Use the hidden state from the previous step of the decoder As a query, the aggregated local sequence Additive attention: in, It is a hidden state. , , For attention parameters; c8. State vector fusion: in, , ; c9. Using GRU decoding, each step's input typically includes a state vector and the previous step's output: The calculation process is as follows: Update Gate: in, To update the gate output, For weights and biases; Reset Door: Candidate hidden state: Hidden status update: in, Final hidden state: in, ; c10. Use MSE as the training loss and RMSE / MAE as the evaluation metric: RMSE = IS = in, For the true value, These are the model's predicted values.

9. The online prediction method for CO in dry quenching based on a distributed computing hybrid GRU according to claim 8, characterized in that: Step c3 specifically includes: one of the server nodes The calculation process is as follows: c31, Update Gate: in, To update the gate output, the value ranges from 0 to 1. To update the gate weight matrix, To update the gate bias term, The Sigmoid activation function is used. Let be the hidden state vector from the previous time step. This is the input vector at the current moment; c32, Reset the door: To reset the gate output, a value of 0 to 1 is used to control the proportion of forgotten historical information. To reset the gate weight matrix, To reset the door offset item; c33, Candidate Hidden State: in, Candidate hidden state vectors For candidate state weight matrix, For candidate state bias terms, For hyperbolic tangent activation function, This is element-wise multiplication; c34, Hidden state update: in, For the first The hidden state vector updated by each server node.

Images