CN117744476A - Real-time dosage monitoring method and system for finished product slurry tank - Google Patents

Abstract

A real-time consumption monitoring method and system for a finished product slurry tank belong to the technical field of lithium battery slurry mixing, and solve the problem of how to accurately monitor the real-time consumption of the finished product slurry tank; according to the invention, an LSTM-KF algorithm slurry consumption prediction model is constructed; collecting the consumption of a slurry tank and the blanking speed data of a sensor in the current slurry mixing process, and processing the collected data; an LSTM model for predicting average blanking speed and predicting instantaneous blanking speed by taking the minimized root mean square error as a loss function; testing the time sequence data classification effect and the model fitting condition by adopting test set data; continuously predicting the real-time consumption of the finished slurry tank by adopting a trained LSTM-KF algorithm; compared with the method relying on a preset single model, repeated modeling aiming at different process conditions can be reduced, and the dynamic change of the sizing process can be responded quickly by establishing a regular LSTM prediction model, so that the real-time performance of tracking is improved; and accurately and reliably evaluating the quality of the combined pulp.

Description

A method and system for real-time monitoring of the amount of finished product slurry tank

Technical Field

The invention belongs to the technical field of lithium battery slurry mixing, and relates to a real-time usage monitoring method and system for a finished product slurry tank.

Background Art

In the lithium battery slurry mixing process, the accuracy of each component of the slurry mixing line directly affects the quality of the final slurry. There are complex dynamic correlations and mutual influences between the actual amount of finished slurry tank in the slurry mixing process, the corresponding theoretical amount prediction, sensor output, and material feeding speed. Traditional tracking mainly relies on the experience of operators and assists in adjustment by manually counting the differences between the two, but the accuracy is difficult to guarantee and the tracking is not timely enough. Although online measurement equipment such as automatic flow meters are used, the tracking effect of dynamic changes is still poor.

In recent years, data-driven prediction methods have been initially applied, such as time series neural networks, which can model regularities. However, this type of single model method has limited generalization and is difficult to adapt to different pulping processes. Some state estimation algorithms such as Kalman filtering can also be used for data fusion, but the dynamic response to the material feeding speed is slow. There are many shortcomings in current tracking technology. The traditional tracking mechanism that relies on empirical rules is difficult to respond to the dynamic changes of the process in real time; the accuracy of existing automatic measurement equipment is limited, and it is difficult to accurately reflect the complex coupling process of multiple variables; the prediction technology that relies only on a single data-driven model has the problem of poor generalization and is difficult to adapt to different process conditions. In addition, the independent state estimation algorithm is slow to respond to the dynamic feeding speed and fails to achieve stable and accurate tracking. What is more serious is that the existing technology generally has a weak ability to identify abnormal situations and cannot provide effective early warning.

In summary, the existing tracking technology is difficult to balance accuracy and stability, and the mutual influence between multiple variables is not considered enough. In addition, the model needs to be repeatedly debugged for different processes, and the reusability is poor. Therefore, it is necessary to develop a new tracking method to achieve accurate monitoring of the real-time usage of the finished slurry tank and ensure the accuracy of the slurry formula.

Summary of the invention

The technical problem to be solved by the present invention is how to accurately monitor the real-time usage of the finished product slurry tank.

The present invention solves the above technical problems through the following technical solutions:

A method for real-time consumption monitoring of a finished product slurry tank comprises the following steps:

Step 1: Construct a slurry dosage prediction model based on LSTM-KF algorithm;

Step 2: Collect the slurry tank usage and sensor feeding speed data during the current slurry mixing process, and process the collected data;

Step 3: Train, verify and test the LSTM-KF algorithm;

Step 4: Use the trained LSTM-KF algorithm to continuously predict the real-time usage of the finished slurry tank.

Furthermore, the method for constructing a slurry dosage prediction model based on the LSTM-KF algorithm described in step 1 is as follows:

(1) Based on the KF algorithm, the transfer process matrix is established, and Gaussian noise is used to model the error in the target measurement information. The Kalman gain is obtained by predicting the state noise covariance matrix and the measurement noise covariance matrix. The state prediction vector obtained after the transfer is combined with the state vector obtained by the sensor using the Kalman gain to obtain the state vector of the filter and the state estimation noise covariance matrix;

(2) Establish an LSTM model and use the forget gate to filter the information that needs to be discarded. The input gate determines the new information in the state of the storage unit, and the output gate determines the output of the current unit. After the forward propagation process is completed, the parameters of the LSTM are updated through backpropagation. This process is repeated until the model converges. The LSTM uses the forget gate, input gate, and output gate mechanism to achieve a trade-off between the input at the previous moment and the input at the new moment, and establish a mapping relationship between the data.

Furthermore, the formula of the transfer process matrix is as follows:

X _k|k-1 =FX _k-1 +w

P _k|k-1 = FP _k-1 F ^T + Q

Wherein, w is the disturbance noise, X _k|k-1 is the state prediction vector, P _k|k-1 is the predicted state noise covariance matrix, X _k-1 is the state vector of the filtered output at time k-1, and P _k-1 is the state estimation covariance matrix of the filtered output at time k-1; F is the transfer matrix, ^FT is the transfer matrix transpose matrix, and Q is the disturbance noise covariance matrix under uniform speed;

Let the disturbance noise obey the Gaussian noise with mean 0 and variance q. The covariance matrix of the disturbance noise at a uniform speed is as follows:

The state transition mode of the target is described by the transfer matrix F, and the formula is as follows:

Wherein, Δt is the measurement interval.

Furthermore, the process of obtaining the Kalman gain by predicting the state noise covariance matrix and the measurement noise covariance matrix, combining the state prediction vector obtained after the transfer with the state vector obtained by the sensor using the Kalman gain, and obtaining the state vector of the filter and the state estimation noise covariance matrix is as follows:

The formula of the measurement noise covariance matrix is:

Among them, σ _c ² is the variance of the usage noise, σ _v ² is the variance of the feeding speed measurement noise;

The Kalman gain is obtained by predicting the state noise covariance matrix P _k|k-1 and the measurement noise covariance matrix R:

K＝P _k|k-1 H ^T (HP _k|k-1 H ^T +R) ^-1

The observation matrix is:

The Kalman gain K _k is used to combine the state prediction vector X _k|k-1 obtained after the transfer with the state vector Z = [c _k ,v _k ] obtained by the sensor to obtain the state vector and state estimation noise covariance matrix of the filter:

X _k =X _k|k-1 +K _k (Z _k -HX _k|k-1 )

P _k =(IK _k H)X _k|k-1

Among them, _Xk and _Pk are used as inputs at the next moment to continuously predict the dose value.

Furthermore, the method of using the forget gate to filter the information to be discarded is as follows:

The output of the current input _xt and the previous output ht _-1 after the forget gate is:

F _t =sigmod(W _f [h _t-1 ,x _t ]+b _f )

Among them, _Wf and _bf are the weight and bias of the forget gate;

The method by which the input gate determines the new information in the storage unit state is as follows:

The sigmoid part determines the update of the value and outputs the vector _it . The tanh part creates a new candidate value ~ _Ct and adds it to the state. The current state is:

i _t =sigmoid(W _i [h _t-1 ,x _t ]+b _i )

Among them, _Wi and _bi are the weight and bias of the input gate, _Wn and _bn are the weight and bias of the tanh part;

The method by which the output gate determines the output of the current unit is as follows:

After the previous moment's output h _t-1 and x _t are processed, the output o _t and the current moment's state C _t are multiplied by the tanh part to obtain the final output:

h _t = o _t tanh(C _t )

o _t =sigmoid(W _o [h _t-1 ,x _t ]+b ₀ )

Where W _o and b _o are the weight and bias of the input gate.

Furthermore, the process of collecting the slurry tank usage and sensor feeding speed data in the current slurry mixing process described in step 2 is as follows: the length of the time series of the collected usage is N, let _Li = {c _r,k, k = 1, 2, ..., N} be the time series data set of the i-th usage in the data set, and _vi = {v _r,k , k = 1, 2, ..., N} be the real feeding speed set, where cr _,k and v _r,k are the real usage and real feeding speed of the sensor respectively;

The process of processing the collected data is as follows:

Use a sliding window to extract data of length w _L from the usage time series dataset _Li and obtain:

I＝{c _r,k ,k＝j,j+1,...,W _L+j-1 }

Use the formula to perform differential processing on the amount and convert the input amount into an average feeding speed:

Considering that the target instantaneous material feeding speed and the expected average material feeding speed also have a correlation in time series characteristics, when predicting the average material feeding speed, the instantaneous material feeding speed is also used as input. Then the j-th input vector of the LSTM for predicting the average material feeding speed is:

Use the sliding window to extract the length _wL of the material feeding speed data from the material feeding speed dataset v _i , and get the jth input vector of the LSTM model for predicting the instantaneous material feeding speed:

I _V ={v _r,k ,k=j,...,W _L+j-1 }

The data is normalized to obtain the normalized instantaneous material feeding speed and average material feeding speed time series, namely I′ _L and I′ _v , where the normalization formula is as follows:

The inputs of the constructed LSTM for predicting the average feeding speed of the usage and the LSTM for predicting the instantaneous feeding speed of the usage are respectively, where j ranges from 1 to N:

I′ _V,j ={v′ _r,k ,k＝1,...,W _L }

Among them, I' represents the normalized data set, _Ii represents the time series data set, and _Imax and _Imin represent the minimum values in each feature sequence.

Furthermore, the process of training, verifying and testing the LSTM-KF algorithm described in step 3 is as follows: using the root mean square error as the loss function to minimize, the LSTM model predicting the average feeding speed and the instantaneous feeding speed; using the Adam optimizer to perform adaptive learning rate adjustment on the learning rate, and setting the current batch size; using the training set data to train the LSTM model, and verifying the convergence of the model through loss function visualization; using the verification set data to perform five-fold cross-validation to verify the generalization of the model; using the test set data to test the classification effect of the time series data and the model fitting;

The formula of the loss function is as follows:

Among them, L represents the root mean square error calculation, N represents the length of the input vector, v _r represents the input dosage and feeding speed, and v _p represents the feeding speed predicted by the model.

Furthermore, the steps of using the trained LSTM-KF algorithm to continuously predict the real-time consumption of the finished slurry tank described in step 4 are as follows:

(1) The input vector is constructed through preprocessing and input into the LSTM model. The LSTM model predicts the average material feeding speed and instantaneous material feeding speed of the target respectively. After the prediction is completed, the denormalization is performed to obtain the average material feeding speed prediction. And the instantaneous feeding speed prediction value

(2) The uniform motion model used by the KF algorithm takes the filtered material feeding speed v _f,k-1 at the previous moment as the average material feeding speed between moment k-1 and moment k, and calculates the usage prediction result at moment k; similarly, the KF algorithm directly takes the filtered material feeding speed at moment k-1 as the instantaneous material feeding speed at moment k;

The calculation formula for the usage prediction result at the k moment is:

(3) The LSTM-KF algorithm uses the instantaneous material feeding speed predicted by LSTM And the corrected speed prediction v _f,k-1 , through the KF algorithm, the final feeding speed prediction v _p,k is obtained, and the calculation formula is:

v _p,k =U(v _p,k ,v _f,k-1 )

(4) The KF algorithm is used again to combine the predicted consumption c _p,k and material feeding speed v _p,k at time k with the measured consumption c _m,k and material feeding speed v _m,k at time k to obtain the final consumption estimate and material feeding speed estimate at time k, which are:

c _f,k =U(c _p,k ,c _m,k )

v _f,k =U(v _p,k ,v _m,k )

(5) Repeat the above steps to achieve continuous prediction of the real-time usage of the finished slurry tank.

A finished product slurry tank real-time usage monitoring system, comprising: a model building module, a data acquisition and processing module, a model training module, and a model application module;

The model building module is used to build a slurry dosage prediction model based on the LSTM-KF algorithm;

The data acquisition and processing module is used to collect the slurry tank usage and sensor feeding speed data during the current slurry mixing process, and process the collected data;

The model training module is used to train, verify and test the LSTM-KF algorithm;

The model application module is used to use the trained LSTM-KF algorithm to continuously predict the real-time usage of the finished slurry tank.

Furthermore, the method for constructing a slurry dosage prediction model based on the LSTM-KF algorithm described in the model construction module is as follows:

(1) Based on the KF algorithm, the transfer process matrix is established, and Gaussian noise is used to model the error in the target measurement information. The Kalman gain is obtained by predicting the state noise covariance matrix and the measurement noise covariance matrix. The state prediction vector obtained after the transfer is combined with the state vector obtained by the sensor using the Kalman gain to obtain the state vector of the filter and the state estimation noise covariance matrix;

(2) Establish an LSTM model and use the forget gate to filter the information that needs to be discarded. The input gate determines the new information in the state of the storage unit, and the output gate determines the output of the current unit. After the forward propagation process is completed, the parameters of the LSTM are updated through backpropagation. This process is repeated until the model converges. The LSTM uses the forget gate, input gate, and output gate mechanism to achieve a trade-off between the input at the previous moment and the input at the new moment, and establish a mapping relationship between the data.

Furthermore, the formula of the transfer process matrix is as follows:

X _k|k-1 =FX _k-1 +w

P _k|k-1 = FP _k-1 F ^T + Q

Wherein, w is the disturbance noise, X _k|k-1 is the state prediction vector, P _k|k-1 is the predicted state noise covariance matrix, X _k-1 is the state vector of the filtered output at time k-1, and P _k-1 is the state estimation covariance matrix of the filtered output at time k-1; F is the transfer matrix, ^FT is the transfer matrix transpose matrix, and Q is the disturbance noise covariance matrix under uniform speed;

Let the disturbance noise obey the Gaussian noise with mean 0 and variance q. The covariance matrix of the disturbance noise at a uniform speed is as follows:

The state transition mode of the target is described by the transfer matrix F, and the formula is as follows:

Wherein, Δt is the measurement interval.

Furthermore, the process of obtaining the Kalman gain by predicting the state noise covariance matrix and the measurement noise covariance matrix, combining the state prediction vector obtained after the transfer with the state vector obtained by the sensor using the Kalman gain, and obtaining the state vector of the filter and the state estimation noise covariance matrix is as follows:

The formula of the measurement noise covariance matrix is:

Among them, σ _c ² is the variance of the usage noise, σ _v ² is the variance of the feeding speed measurement noise;

The Kalman gain is obtained by predicting the state noise covariance matrix P _k|k-1 and the measurement noise covariance matrix R:

K＝P _k|k-1 H ^T (HP _k|k-1 H ^T +R) ^-1

The observation matrix is:

The Kalman gain K _k is used to combine the state prediction vector X _k|k-1 obtained after the transfer with the state vector Z = [c _k ,v _k ] obtained by the sensor to obtain the state vector and state estimation noise covariance matrix of the filter:

X _k =X _k|k-1 +K _k (Z _k -HX _k|k-1 )

P _k =(IK _k H)X _k|k-1

Among them, _Xk and _Pk are used as inputs at the next moment to continuously predict the dose value.

Furthermore, the method of using the forget gate to filter the information to be discarded is as follows:

The output of the current input _xt and the previous output ht _-1 after the forget gate is:

F _t =sigmod(W _f [h _t-1 ,x _t ]+b _f )

Among them, _Wf and _bf are the weight and bias of the forget gate;

The method by which the input gate determines the new information in the storage unit state is as follows:

The sigmoid part determines the update of the value and outputs the vector _it , while the tanh part creates a new candidate value Added to the state, the current state is:

i _t =sigmoid(W _i [h _t-1 ,x _t ]+b _i )

Among them, _Wi and _bi are the weight and bias of the input gate, _Wn and _bn are the weight and bias of the tanh part;

The method by which the output gate determines the output of the current unit is as follows:

After the previous moment's output h _t-1 and x _t are processed, the output o _t and the current moment's state C _t are multiplied by the tanh part to obtain the final output:

h _t = o _t tanh(C _t )

o _t =sigmoid(W _o [h _t-1 ,x _t ]+b ₀ )

Where W _o and b _o are the weight and bias of the input gate.

Furthermore, the process of collecting the slurry tank usage and sensor feeding speed data in the current slurry mixing process described in the data collection and processing module is as follows: the time series length of the collected usage is N, let _Li = {c _r,k , k = 1, 2, ..., N} be the time series data set of the i-th usage in the data set, and _vi = {v _r,k , k = 1, 2, ..., N} be the real feeding speed set, where cr _,k and v _r,k are the real usage and real feeding speed of the sensor respectively;

The process of processing the collected data is as follows:

Use a sliding window to extract data of length w _L from the usage time series dataset _Li and obtain:

I＝{c _r,k ,k＝j,j+1,...,W _L+j-1 }

Use the formula to perform differential processing on the amount and convert the input amount into an average feeding speed:

Considering that the target instantaneous feeding speed and the expected average feeding speed also have a correlation in time series characteristics, when predicting the average feeding speed, the instantaneous feeding speed is also used as input. Then the j-th input vector of the LSTM for predicting the average feeding speed is:

Use the sliding window to extract the length _wL of the material feeding speed data from the material feeding speed dataset v _i , and get the jth input vector of the LSTM model for predicting the instantaneous material feeding speed:

I _V ={v _r,k ,k=j,...,W _L+j-1 }

The data is normalized to obtain the normalized instantaneous material feeding speed and average material feeding speed time series, namely I′ _L and I′ _v , where the normalization formula is as follows:

The inputs of the constructed LSTM for predicting the average feeding speed of the usage and the LSTM for predicting the instantaneous feeding speed of the usage are respectively, where j ranges from 1 to N:

I′ _V,j ={v′ _r,k ,k＝1,...,W _L }

Among them, I' represents the normalized data set, _Ii represents the time series data set, and _Imax and _Imin represent the minimum values in each feature sequence.

Furthermore, the process of training, verifying and testing the LSTM-KF algorithm described in the model training module is as follows: using the root mean square error as the loss function to minimize, the LSTM model predicting the average feeding speed and the instantaneous feeding speed; using the Adam optimizer to perform adaptive learning rate adjustment on the learning rate, and setting the current batch size; using the training set data to train the LSTM model, and verifying the convergence of the model through loss function visualization; using the verification set data to perform five-fold cross-validation to verify the generalization of the model; using the test set data to test the classification effect of the time series data and the model fitting;

The formula of the loss function is as follows:

Among them, L represents the root mean square error calculation, N represents the length of the input vector, v _r represents the input dosage and feeding speed, and v _p represents the feeding speed predicted by the model.

Furthermore, the steps of using the trained LSTM-KF algorithm to continuously predict the real-time consumption of the finished slurry tank described in the model application module are as follows:

(1) The input vector is constructed through preprocessing and input into the LSTM model. The LSTM model predicts the average material feeding speed and instantaneous material feeding speed of the target respectively. After the prediction is completed, the denormalization is performed to obtain the average material feeding speed prediction. And the instantaneous feeding speed prediction value

(2) The uniform motion model used by the KF algorithm takes the filtered material feeding speed v _f,k-1 at the previous moment as the average material feeding speed between moment k-1 and moment k, and calculates the usage prediction result at moment k; similarly, the KF algorithm directly takes the filtered material feeding speed at moment k-1 as the instantaneous material feeding speed at moment k;

The calculation formula for the usage prediction result at the k moment is:

(3) The LSTM-KF algorithm uses the instantaneous material feeding speed predicted by LSTM And the corrected speed prediction v _f,k-1 , through the KF algorithm, the final feeding speed prediction v _p,k is obtained, and the calculation formula is:

v _p,k =U(v _p,k ,v _f,k-1 )

(4) The KF algorithm is used again to combine the predicted consumption c _p,k and material feeding speed v _p,k at time k with the measured consumption c _m,k and material feeding speed v _m,k at time k to obtain the final consumption estimate and material feeding speed estimate at time k, which are:

c _f,k =U(c _p,k ,c _m,k )

v _f,k =U(v _p,k ,v _m,k )

(5) Repeat the above steps to achieve continuous prediction of the real-time usage of the finished slurry tank.

A storage medium stores a computer program, which executes the steps of the method for real-time consumption monitoring of a finished product slurry tank when the computer program is run by a processor.

The advantages of the present invention are:

The technical solution of the present invention can quickly respond to changes by training and applying the LSTM network to learn patterns from history; at the same time, the LSTM prediction output is integrated with the Kalman filter recursive estimation, which can make the state evaluation more stable and accurate; compared with relying on a preset single model, the present invention has stronger adaptability, can reduce repeated modeling for different process conditions, and by establishing a regular LSTM prediction model, can quickly respond to the dynamic changes of the slurry mixing process and improve the real-time tracking. Compared with the prior art, the technical solution of the present invention adopts the fusion of LSTM and Kalman filtering, which can significantly improve the accuracy and stability of state estimation; fully utilizes the LSTM modeling ability for complex time series patterns and the recursive estimation algorithm of Kalman filtering to enhance the adaptability of the solution to dynamic changes. The solution provides real-time, accurate and reliable state evaluation, which greatly improves the level of slurry quality monitoring; the solution has moderate computational complexity and can realize real-time online monitoring and control of industrial processes through parallelization. By quickly and accurately identifying anomalies, the solution enhances the stability of the slurry mixing process and reduces quality risks. In summary, the solution significantly improves the controllability of slurry mixing, which is of great significance to ensuring battery quality.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a method for real-time consumption monitoring of a finished product slurry tank according to an embodiment of the present invention;

FIG2 is a structural diagram of a slurry consumption prediction model of an LSTM-KF algorithm of a method for real-time consumption monitoring of a finished product slurry tank according to an embodiment of the present invention.

DETAILED DESCRIPTION

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

The technical solution of the present invention is further described below in conjunction with the accompanying drawings and specific embodiments:

Embodiment 1

As shown in FIG1 , the method for real-time consumption monitoring of a finished product slurry tank according to an embodiment of the present invention comprises the following steps:

1. Construct a slurry dosage prediction model based on the LSTM-KF algorithm, including: KF algorithm and LSTM model, as follows:

1.1 KF algorithm

Let c _k-1 be the finished product consumption at time k-1, v _k-1 be the estimated material feeding speed at time k-1, then the state vector of the filter output at time k-1 is X _k-1 = [c _k-1 , v _k-1 ], then P _k-1 is the state estimation covariance matrix of the filter output at time k-1. _{X k|k-1} = [c _k , v _k ] is the state prediction vector, P _k|k-1 is the state prediction noise covariance matrix, then the transfer process can be described as:

X _k|k-1 =FX _k-1 +w

P _k|k-1 = FP _k-1 F ^T + Q

Among them, w is the disturbance noise, which represents the interference in the transfer process. Let the transfer noise obey the Gaussian noise with a mean of 0 and a variance of q, and transfer the noise covariance matrix at a uniform speed, the formula is as follows:

The state transition mode of the target is described by the transfer matrix F, and the formula is as follows:

Where Δt is the measurement interval.

Gaussian noise is used to model the error in the target's measurement information, and the measurement noise covariance matrix is:

Where: σ _c ² is the variance of the usage noise, σ _v ² is the variance of the feeding speed measurement noise.

The Kalman gain is obtained by predicting the state noise covariance matrix P _k|k-1 and the measurement noise covariance matrix R:

K＝P _k|k-1 H ^T (HP _k|k-1 H ^T +R) ^-1

The observation matrix is:

The KF algorithm uses the Kalman gain _Kk to combine the state prediction vector Xk _|k-1 obtained after the transfer with the state vector Z=[c _k ,v _k ] obtained by the sensor to obtain the state vector and state estimation noise covariance matrix of the filter:

X _k =X _k|k-1 +K _k (Z _k -HX _k|k-1 )

P _k =(IK _k H)X _k|k-1

Among them, _Xk and _Pk are used as inputs at the next moment to continuously predict the dose value.

1.2 LSTM Model

As shown in Figure 2, first, the forget gate filters the information that needs to be discarded. The output of the current moment input _xt and the previous moment output _ht-1 after the forget gate is:

F _t =sigmod(W _f [h _t-1 ,x _t ]+b _f )

Where _Wf and _bf are the weight and bias of the forget gate.

Next, the input gate determines the new information in the state of the storage unit. The sigmoid part determines the update of the value and outputs the vector _it . The tanh part creates a new candidate value Added to the state, the current state is:

i _t =sigmoid(W _i [h _t-1 ,x _t ]+b _i )

Among them, _Wi and _bi are the weight and bias of the input gate, and _Wn and _bn are the weight and bias of the tanh part.

Finally, the output gate determines the output of the current unit. After the outputs h _t-1 and x _t of the previous moment are processed, the output o _t and the current state C _t are multiplied by the tanh part to obtain the final output:

h _t = o _t tanh(C _t )

o _t =sigmoid(W _o [h _t-1 ,x _t ]+b ₀ )

Where W _o and b _o are the weight and bias of the input gate;

After the forward propagation process is completed, the parameters of the LSTM are updated through back propagation, and the process is repeated until the model converges. LSTM uses the forget gate, input gate, and output gate mechanism to achieve a trade-off between the input at the previous moment and the input at the new moment, and establish a mapping relationship between the data.

2. Collect the slurry tank usage and sensor feeding speed data during the current slurry mixing process, and process the collected data

Currently, the usage is directly used as input and the expected usage is used as a label to achieve the prediction of usage. Since the usage used in the training set is a finite set, while the usage in the actual scenario is an infinite set, the model cannot train all usages, so it is difficult to achieve a high prediction accuracy after changing the usage scenario or coordinate system. The LSTM_KF algorithm uses LSTM to predict the average unloading speed of the target between time k-1 and time k respectively and instantaneous material feeding speed at time k The current usage is further obtained through calculation.

Let the length of the time series of usage be N, let _Li = {c _r,k , k = 1, 2, ..., N} be the time series data set of the i-th usage in the data set, and _vi = {v _r,k , k = 1, 2, ..., N} be the real material discharge speed set, where cr _,k and v _r,k are the real usage and real material discharge speed of the sensor respectively.

The specific data processing process is as follows:

First, use a sliding window to extract data of length w _L from the usage time series dataset _Li , and obtain:

I＝{c _r,k ,k＝j,j+1,...,W _L+j-1 }

The formula is further used to perform differential processing on the amount, converting the input amount into the average feeding speed:

Considering that the target instantaneous material feeding speed and the expected average material feeding speed also have a temporal correlation, when predicting the average material feeding speed, taking the instantaneous material feeding speed as input can add one-dimensional information and improve the prediction accuracy. Then the j-th input vector of the LSTM for predicting the average material feeding speed is:

Using a sliding window to extract the length _wL of the material feeding speed data from the material feeding speed dataset v _i , the j-th input vector of the LSTM model for predicting the instantaneous material feeding speed can be obtained as:

I _V ={v _r,k ,k=j,...,W _L+j-1 }

According to verification, the size of the sliding window is set to 9 here.

Furthermore, the data is normalized here to obtain the normalized instantaneous material feeding speed and average material feeding speed time series, namely I′ _L and I′ _v , where the normalization formula is as follows:

Among them, I' represents the normalized data set, _Ii represents the time series data set, and _Imax and _Imin represent the minimum values in each feature sequence.

The inputs of the constructed LSTM for predicting the average feeding speed of the usage and the LSTM for predicting the instantaneous feeding speed of the usage are respectively, where j ranges from 1 to N:

I′ _V,j ={v′ _r,k ,k＝1,...,W _L }

3. Train, verify and test the LSTM-KF algorithm

3.1 LSTM-KF algorithm flow

1) Perform feature analysis on the data and propose a method based on average feeding speed and instantaneous feeding speed to achieve prediction, which solves the problem of poor generalization of existing non-parametric model solutions. Subsequently, the usage data and feeding speed data are processed and converted into the features required by LSTM.

2) Select a reasonable structure and loss function, and use the back-propagation algorithm to train the LSTM for predicting the average feeding speed and the LSTM for predicting the instantaneous feeding speed.

3) Preprocess the coordinate and feeding speed input data and input them into LSTM for prediction. Output average feeding speed prediction And filter feeding speed The average material feeding speed prediction is used to calculate the consumption c _p,k ; the instantaneous material feeding speed prediction is output Combined with the filtered material feeding speed v _f,k-1 , the material feeding speed prediction v _p,k is obtained. Finally, the usage c _p,k and the material feeding speed prediction v _p,k are combined with the current measurements c _m,k and v _m,k to obtain the final estimate c _f,k and the material feeding speed estimate v _f,k .

3.2 Model Training

The model uses the minimum root mean square error (RMSE) as the loss function. The loss function of the LSTM model for predicting the average feeding speed and predicting the instantaneous feeding speed is set as follows:

Among them, L represents the root mean square error calculation, N represents the length of the input vector, v _r represents the input dosage and feeding speed, and v _p represents the feeding speed predicted by the model.

For the learning rate, the Adam optimizer is used here for adaptive learning rate adjustment. At the same time, the current batch size is set.

Here, the training set, validation set, and test set are divided according to the ratio of 7:1:2. The LSTM model of 70% of the data is trained to verify the convergence effect of the model through loss function visualization. The generalization of the model is verified by performing a five-fold cross-validation on 10% of the data. Finally, the optimal model parameters are applied to the remaining 20% test set to test the classification effect of time series data and the model fitting.

4. Use the trained LSTM-KF algorithm to continuously predict the real-time consumption of the finished slurry tank

First, after the existing data length meets the LSTM requirements, the input vector is constructed through preprocessing and input into the LSTM. The LSTM model predicts the average material feeding speed and instantaneous material feeding speed of the target respectively. After the prediction is completed, the two models are denormalized to obtain the average material feeding speed prediction And the instantaneous feeding speed prediction value

Next, the uniform motion model used by the KF algorithm takes the filtered material feeding speed v _f,k-1 of the previous moment as the average material feeding speed between moment k-1 and moment k, and calculates the usage prediction result at moment k. The LSTM-KF algorithm uses the average material feeding speed predicted by LSTM and the previous material feeding speed estimate v _f,k-1 , the KF algorithm is used to calculate the usage prediction cp _,k at time k, which can make the usage material feeding speed closer to the actual material feeding speed value, thereby improving the prediction accuracy. The usage prediction at time k is:

Similarly, the KF algorithm uses the filtered material feeding speed at time k-1 as the instantaneous material feeding speed at time k. The LSTM-KF algorithm uses the instantaneous material feeding speed predicted by LSTM. And the modified speed prediction v _f,k-1 , through the KF algorithm, the final feeding speed prediction v _p,k is obtained, which can improve the prediction accuracy. The instantaneous feeding speed of the amount at time k is:

v _p,k =U(v _p,k ,v _f,k-1 )

Finally, the KF algorithm is used again to combine the predicted usage c _p,k and material feeding speed v _p,k at time k with the measured usage c _m,k and material feeding speed v _m,k at time k to obtain the final usage estimate and material feeding speed estimate at time k, which are:

c _f,k =U(c _p,k ,c _m,k )

v _f,k =U(v _p,k ,v _m,k )

Repeat the above steps to achieve continuous prediction of usage.

Embodiment 2

A finished product slurry tank real-time usage monitoring system, comprising: a model building module, a data acquisition and processing module, a model training module, and a model application module;

The model building module is used to build a slurry dosage prediction model based on the LSTM-KF algorithm, and the method is as follows:

(1) Let c _k-1 be the finished product quantity at time k-1, v _k-1 be the estimated material feeding speed at time k-1, then the state vector of the filter output at time k-1 is X _k-1 = [c _k-1 , v _k-1 ], then P _k-1 is the state estimation covariance matrix of the filter output at time k-1; X _k|k-1 = [c _k , v _k ] is the state prediction vector, P _k|k-1 is the state prediction noise covariance matrix, then the transfer process can be described as:

X _k|k-1 =FX _k-1 +w

P _k|k-1 = FP _k-1 F ^T + Q

Among them, w is the disturbance noise, which represents the interference in the transfer process. Let the transfer noise obey the Gaussian noise with a mean of 0 and a variance of q, and transfer the noise covariance matrix at a uniform speed, the formula is as follows:

The state transition mode of the target is described by the transfer matrix F, and the formula is as follows:

Where Δt is the measurement interval;

Gaussian noise is used to model the error in the target's measurement information, and the measurement noise covariance matrix is:

Where: σ _c ² is the variance of the usage noise, σ _v ² is the variance of the feeding speed measurement noise;

The Kalman gain is obtained by predicting the state noise covariance matrix P _k|k-1 and the measurement noise covariance matrix R:

K＝P _k|k-1 H ^T (HP _k|k-1 H ^T +R) ^-1

The observation matrix is:

The KF algorithm uses the Kalman gain _Kk to combine the state prediction vector Xk _|k-1 obtained after the transfer with the state vector Z=[c _k ,v _k ] obtained by the sensor to obtain the state vector and state estimation noise covariance matrix of the filter:

X _k =X _k|k-1 +K _k (Z _k -HX _k|k-1 )

P _k =(IK _k H)X _k|k-1

Among them, X _k and P _k are used as inputs at the next moment to continuously predict the dose value;

(2) The output of the current input _xt and the previous output ht _-1 after the forget gate is:

F _t =sigmod(W _f [h _t-1 ,x _t ]+b _f )

Where _Wf and _bf are the weight and bias of the forget gate;

The input gate determines the new information in the state of the storage unit, the sigmoid part determines the update of the value, the output vector _it , and the tanh part creates a new candidate value. Added to the state, the current state is:

i _t =sigmoid(W _i [h _t-1 ,x _t ]+b _i )

Among them, _Wi and _bi are the weight and bias of the input gate, _Wn and _bn are the weight and bias of the tanh part;

The output gate determines the output of the current unit. After the outputs h _t-1 and x _t of the previous moment are processed, the output o _t and the state C _t of the current moment are multiplied by the tanh part to obtain the final output:

h _t = o _t tanh(C _t )

o _t =sigmoid(W _o [h _t-1 ,x _t ]+b ₀ )

Where W _o and b _o are the weight and bias of the input gate;

After the forward propagation process is completed, the parameters of LSTM are updated through back propagation, and the process is repeated until the model converges; LSTM uses the forget gate, input gate and output gate mechanism to achieve a trade-off between the input at the previous moment and the input at the new moment, and establish a mapping relationship between the data.

The data acquisition and processing module is used to collect the slurry tank usage and sensor feeding speed data in the current slurry mixing process. The time series length of the collected usage is N. Let _Li = {c _r,k , k = 1, 2, ..., N} be the time series data set of the i-th usage in the data set, and _vi = {v _r,k , k = 1, 2, ..., N} be the real feeding speed set, where _cr,k and v _r,k are the real usage and real feeding speed of the sensor respectively, and the collected data is processed;

The process of processing the collected data is as follows:

Use a sliding window to extract data of length w _L from the usage time series dataset _Li and obtain:

I＝{c _r,k ,k＝j,j+1,...,W _L+j-1 }

Use the formula to perform differential processing on the amount and convert the input amount into an average feeding speed:

Considering that the target instantaneous feeding speed and the expected average feeding speed also have a correlation in time series characteristics, when predicting the average feeding speed, the instantaneous feeding speed is also used as input. Then the j-th input vector of the LSTM for predicting the average feeding speed is:

Use the sliding window to extract the length _wL of the material feeding speed data from the material feeding speed dataset v _i , and get the jth input vector of the LSTM model for predicting the instantaneous material feeding speed:

I _V ={v _r,k ,k=j,...,W _L+j-1 }

The data is normalized to obtain the normalized instantaneous material feeding speed and average material feeding speed time series, namely I′ _L and I′ _v , where the normalization formula is as follows:

The inputs of the constructed LSTM for predicting the average feeding speed of the usage and the LSTM for predicting the instantaneous feeding speed of the usage are respectively, where j ranges from 1 to N:

I′ _V,j ={v′ _r,k ,k＝1,...,W _L }

Among them, I' represents the normalized data set, _Ii represents the time series data set, and _Imax and _Imin represent the minimum values in each feature sequence.

The model training module is used to minimize the root mean square error as the loss function, and predict the LSTM model of the average feeding speed and the instantaneous feeding speed; the learning rate is adaptively adjusted using the Adam optimizer, and the current batch size is set at the same time; the LSTM model is trained using the training set data, and the convergence effect of the model is verified by visualizing the loss function; the validation set data is used for five-fold cross validation to verify the generalization of the model; the test set data is used to test the classification effect of the time series data and the model fitting;

The formula for minimizing the root mean square error loss function is as follows:

Among them, L represents the root mean square error calculation, N represents the length of the input vector, v _r represents the input dosage and feeding speed, and v _p represents the feeding speed predicted by the model.

The model application module is used to continuously predict the real-time consumption of the finished slurry tank using the trained LSTM-KF algorithm, and the steps are as follows:

(1) The input vector is constructed through preprocessing and input into the LSTM model. The LSTM model predicts the average material feeding speed and instantaneous material feeding speed of the target respectively. After the prediction is completed, the denormalization is performed to obtain the average material feeding speed prediction. And the instantaneous feeding speed prediction value

(2) The uniform motion model used by the KF algorithm takes the filtered material feeding speed v _f,k-1 at the previous moment as the average material feeding speed between moment k-1 and moment k, and calculates the usage prediction result at moment k; similarly, the KF algorithm directly takes the filtered material feeding speed at moment k-1 as the instantaneous material feeding speed at moment k;

The calculation formula for the usage prediction result at the k moment is:

(3) The LSTM-KF algorithm uses the instantaneous material feeding speed predicted by LSTM And the corrected speed prediction v _f,k-1 , through the KF algorithm, the final feeding speed prediction v _p,k is obtained, and the calculation formula is:

v _p,k =U(v _p,k ,v _f,k-1 )

(4) The KF algorithm is used again to combine the predicted consumption c _p,k and material feeding speed v _p,k at time k with the measured consumption c _m,k and material feeding speed v _m,k at time k to obtain the final consumption estimate and material feeding speed estimate at time k, which are:

c _f,k =U(c _p,k ,c _m,k )

v _f,k =U(v _p,k ,v _m,k )

(5) Repeat the above steps to achieve continuous prediction of the real-time usage of the finished slurry tank.

Embodiment 3

A storage medium stores a computer program, which, when executed by a processor, executes the steps of the method for real-time consumption monitoring of a finished product slurry tank in Example 1.

The technical solution of the present invention faces the problems of insensitive dynamic response, unstable evaluation, poor generalization applicability, etc. in tracking. The technical solution aims to build a unified tracking framework that integrates learning and estimation algorithms to achieve real-time and accurate capture of complex dynamic processes. By training and applying the LSTM network to learn patterns from history, changes can be responded to quickly; at the same time, the LSTM prediction output is integrated with the Kalman filter recursive estimation to make the state evaluation more stable and accurate. Compared with relying on a preset single model, this framework has stronger adaptability, can reduce repeated modeling for different process conditions, and better serve a wide range of intelligent manufacturing applications. In general, this solution aims to achieve an intelligent, accurate, and responsive tracking technology to meet the needs of industrial process quality control and ensure stable output.

Compared with the current technology, the core innovation of the technology of the present invention is that it fully integrates the advantages of the two types of algorithms, LSTM prediction and Kalman filter estimation algorithm, to achieve a more powerful tracking effect. Specifically, this technology builds a unified data-driven framework and uses the LSTM network to learn historical patterns to achieve online modeling and prediction of dynamic rules, greatly improving the response and feeding speed to changes. At the same time, this technology fuses the prediction output of LSTM with the state estimation of Kalman filter to make the final state assessment more accurate and stable. This effective combination of prediction and filtering not only enhances the ability to identify and warn of anomalies, but also makes the technical framework highly reusable, simplifies the model construction and adjustment process for different application scenarios, and realizes real-time online monitoring. Overall, through the collaborative innovation of algorithms, this technology has achieved accuracy and high responsiveness that are difficult to balance with current tracking technologies.

By establishing a regular LSTM prediction model, it is possible to respond quickly to the dynamic changes of the slurry mixing process and improve the real-time tracking. Compared with the existing technology, this solution adopts the fusion of LSTM and Kalman filtering, which can significantly improve the accuracy and stability of state estimation. By making full use of LSTM's modeling ability for complex time series patterns and Kalman filtering's recursive estimation algorithm, the adaptability of the solution to dynamic changes is enhanced. This solution provides real-time, accurate and reliable state assessment, greatly improving the level of slurry mixing quality monitoring. The solution has a moderate amount of computation and can achieve real-time online monitoring and control of industrial processes through parallelization. By quickly and accurately identifying anomalies, the solution enhances the stability of the slurry mixing process and reduces quality risks. In summary, this solution significantly improves the controllability of slurry mixing, which is of great significance to ensuring battery quality.

The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (10)

1. A method for real-time consumption monitoring of a finished product slurry tank, characterized in that it comprises the following steps: Step 1: Construct a slurry dosage prediction model based on the LSTM-KF algorithm as follows: (1) Based on the KF algorithm, the transfer process matrix is established, and Gaussian noise is used to model the error in the target measurement information. The Kalman gain is obtained by predicting the state noise covariance matrix and the measurement noise covariance matrix. The state prediction vector obtained after the transfer is combined with the state vector obtained by the sensor using the Kalman gain to obtain the state vector of the filter and the state estimation noise covariance matrix; (2) Establish an LSTM model, use the forget gate to filter the information that needs to be discarded, the input gate determines the new information in the state of the storage unit, and the output gate determines the output of the current unit; after the forward propagation process is completed, the parameters of the LSTM are updated through backpropagation, and the process is repeated until the model converges. The LSTM uses the forget gate, input gate, and output gate mechanism to achieve a trade-off between the input at the previous moment and the input at the new moment, and establish a mapping relationship between the data; Step 2: Collect the slurry tank usage and sensor feeding speed data during the current slurry mixing process, and process the collected data; Step 3: Train, verify and test the LSTM-KF algorithm; Step 4: Use the trained LSTM-KF algorithm to continuously predict the real-time usage of the finished slurry tank. 2. The method for real-time consumption monitoring of a finished product slurry tank according to claim 1, characterized in that the formula of the transfer process matrix is as follows: X _k|k-1 =FX _k-1 +w P _k|k-1 = FP _k-1 F ^T + Q Wherein, w is the disturbance noise, X _k|k-1 is the state prediction vector, P _k|k-1 is the predicted state noise covariance matrix, X _k-1 is the state vector of the filtered output at time k-1, and P _k-1 is the state estimation covariance matrix of the filtered output at time k-1; F is the transfer matrix, ^FT is the transfer matrix transpose matrix, and Q is the disturbance noise covariance matrix under uniform speed; Let the disturbance noise obey the Gaussian noise with mean 0 and variance q. The covariance matrix of the disturbance noise at a uniform speed is as follows: The state transition mode of the target is described by the transfer matrix F, and the formula is as follows: Where, Δt is the measurement interval; The formula of the measurement noise covariance matrix is: Among them, σ _c ² is the variance of the usage noise, σ _v ² is the variance of the feeding speed measurement noise; The Kalman gain is obtained by predicting the state noise covariance matrix P _k|k-1 and the measurement noise covariance matrix R: K＝P _k|k-1 H ^T (HP _k|k-1 H ^T +R) ^-1 The observation matrix is: The Kalman gain K _k is used to combine the state prediction vector X _k|k-1 obtained after the transfer with the state vector Z = [c _k ,v _k ] obtained by the sensor to obtain the state vector and state estimation noise covariance matrix of the filter: X _k =X _k|k-1 +K _k (Z _k -HX _k|k-1 ) P _k =(IK _k H)X _k|k-1 Among them, X _k and P _k are used as inputs at the next moment to continuously predict the dose value; The method of using the forget gate to filter the information to be discarded is as follows: The output of the current input _xt and the previous output ht _-1 after the forget gate is: F _t =sigmod(W _f [h _t-1 ,x _t ]+b _f ) Among them, _Wf and _bf are the weight and bias of the forget gate; The method by which the input gate determines the new information in the storage unit state is as follows: The sigmoid part determines the update of the value and outputs the vector _it . The tanh part creates a new candidate value ~ _Ct and adds it to the state. The current state is: i _t =sigmoid(W _i [h _t-1 ,x _t ]+b _i ) Among them, _Wi and _bi are the weight and bias of the input gate, _Wn and _bn are the weight and bias of the tanh part; The method by which the output gate determines the output of the current unit is as follows: After the previous moment's output h _t-1 and x _t are processed, the output o _t and the current moment's state C _t are multiplied by the tanh part to obtain the final output: h _t = o _t tanh(C _t ) o _t =sigmoid(W _o [h _t-1 ,x _t ]+b ₀ ) Where W _o and b _o are the weight and bias of the input gate. 3. The method for real-time consumption monitoring of a finished product slurry tank according to claim 1 is characterized in that the process of collecting the consumption of the slurry tank and the sensor feeding speed data in the current slurry mixing process described in step 2 is as follows: the time series length of the collected consumption is N, let _Li = {c _r,k , k = 1, 2, ..., N} be the time series data set of the i-th consumption in the data set, _vi = {v _r,k , k = 1, 2, ..., N} be the real feeding speed set, where cr _,k and v _r,k are the real consumption and real feeding speed of the sensor respectively; The process of processing the collected data is as follows: Use a sliding window to extract data of length w _L from the usage time series dataset _Li and obtain: I＝{c _r,k ,k＝j,j+1,...,W _L+j-1 } Use the formula to perform differential processing on the amount and convert the input amount into an average feeding speed: Considering that the target instantaneous feeding speed and the expected average feeding speed also have a correlation in time series characteristics, when predicting the average feeding speed, the instantaneous feeding speed is also used as input. Then the j-th input vector of the LSTM for predicting the average feeding speed is: Use the sliding window to extract the length _wL of the material feeding speed data from the material feeding speed dataset v _i , and get the jth input vector of the LSTM model for predicting the instantaneous material feeding speed: I _V ={v _r,k ,k=j,...,W _L+j-1 } The data is normalized to obtain the normalized instantaneous material feeding speed and average material feeding speed time series, namely I′ _L and I′ _v , where the normalization formula is as follows: The inputs of the constructed LSTM for predicting the average feeding speed of the usage and the LSTM for predicting the instantaneous feeding speed of the usage are respectively, where j ranges from 1 to N: I′ _V,j ={v′ _r,k ,k＝1,...,W _L } Among them, I' represents the normalized data set, _Ii represents the time series data set, and _Imax and _Imin represent the minimum values in each feature sequence. 4. The method for real-time consumption monitoring of a finished slurry tank according to claim 1 is characterized in that the process of training, verifying and testing the LSTM-KF algorithm described in step 3 is as follows: using the root mean square error as the loss function to minimize the average feeding speed and the LSTM model that predicts the instantaneous feeding speed; using the Adam optimizer to perform adaptive learning rate adjustment on the learning rate, and setting the current batch size at the same time; using the training set data to train the LSTM model, and visually verifying the convergence of the model through the loss function; using the verification set data to perform five-fold cross-validation to verify the generalization of the model; using the test set data to test the classification effect of the time series data and the model fitting; The formula of the loss function is as follows: Among them, L represents the root mean square error calculation, N represents the length of the input vector, v _r represents the input dosage and feeding speed, and v _p represents the feeding speed predicted by the model. 5. The method for monitoring the real-time usage of a finished product slurry tank according to claim 1 is characterized in that the step of using the trained LSTM-KF algorithm to continuously predict the real-time usage of the finished product slurry tank in step 4 is as follows: (1) The input vector is constructed through preprocessing and input into the LSTM model. The LSTM model predicts the average material feeding speed and instantaneous material feeding speed of the target respectively. After the prediction is completed, the denormalization is performed to obtain the average material feeding speed prediction. And the instantaneous feeding speed prediction value (2) The uniform motion model used by the KF algorithm takes the filtered material feeding speed v _f,k-1 at the previous moment as the average material feeding speed between moment k-1 and moment k, and calculates the usage prediction result at moment k; similarly, the KF algorithm directly takes the filtered material feeding speed at moment k-1 as the instantaneous material feeding speed at moment k; The calculation formula for the usage prediction result at the k moment is: (3) The LSTM-KF algorithm uses the instantaneous material feeding speed ~v _r,k predicted by LSTM and the modified speed prediction v _f,k-1 through the KF algorithm to obtain the final material feeding speed prediction v _p,k . The calculation formula is: v _p,k =U(v _p,k ,v _f,k-1 ) (4) The KF algorithm is used again to combine the predicted consumption c _p,k and material feeding speed v _p,k at time k with the measured consumption c _m,k and material feeding speed v _m,k at time k to obtain the final consumption estimate and material feeding speed estimate at time k, which are: c _f,k =U(c _p,k , _cm,k ) v _f,k =U(v _p,k ,v _m,k ) (5) Repeat the above steps to achieve continuous prediction of the real-time usage of the finished slurry tank. 6. A finished product slurry tank real-time usage monitoring system, characterized by comprising: a model building module, a data acquisition and processing module, a model training module, and a model application module; The model building module is used to build a slurry dosage prediction model based on the LSTM-KF algorithm, and the method is as follows: (1) Based on the KF algorithm, the transfer process matrix is established, and Gaussian noise is used to model the error in the target measurement information. The Kalman gain is obtained by predicting the state noise covariance matrix and the measurement noise covariance matrix. The state prediction vector obtained after the transfer is combined with the state vector obtained by the sensor using the Kalman gain to obtain the state vector of the filter and the state estimation noise covariance matrix; (2) Establish an LSTM model, use the forget gate to filter the information that needs to be discarded, the input gate determines the new information in the state of the storage unit, and the output gate determines the output of the current unit; after the forward propagation process is completed, the parameters of the LSTM are updated through backpropagation, and the process is repeated until the model converges. The LSTM uses the forget gate, input gate, and output gate mechanism to achieve a trade-off between the input at the previous moment and the input at the new moment, and establish a mapping relationship between the data; The data acquisition and processing module is used to collect the slurry tank usage and sensor feeding speed data during the current slurry mixing process, and process the collected data; The model training module is used to train, verify and test the LSTM-KF algorithm; The model application module is used to use the trained LSTM-KF algorithm to continuously predict the real-time usage of the finished slurry tank. 7. The real-time consumption monitoring system for finished slurry tanks according to claim 6, characterized in that the formula of the transfer process matrix is as follows: X _k|k-1 =FX _k-1 +w P _k|k-1 = FP _k-1 F ^T + Q Wherein, w is the disturbance noise, X _k|k-1 is the state prediction vector, P _k|k-1 is the predicted state noise covariance matrix, X _k-1 is the state vector of the filtered output at time k-1, and P _k-1 is the state estimation covariance matrix of the filtered output at time k-1; F is the transfer matrix, ^FT is the transfer matrix transpose matrix, and Q is the disturbance noise covariance matrix under uniform speed; Let the disturbance noise obey the Gaussian noise with mean 0 and variance q. The covariance matrix of the disturbance noise at a uniform speed is as follows: The state transition mode of the target is described by the transfer matrix F, and the formula is as follows: Where, Δt is the measurement interval; The formula of the measurement noise covariance matrix is: Among them, σ _c ² is the variance of the usage noise, σ _v ² is the variance of the feeding speed measurement noise; The Kalman gain is obtained by predicting the state noise covariance matrix P _k|k-1 and the measurement noise covariance matrix R: K＝P _k|k-1 H ^T (HP _k|k-1 H ^T +R) ^-1 The observation matrix is: The Kalman gain K _k is used to combine the state prediction vector X _k|k-1 obtained after the transfer with the state vector Z = [c _k ,v _k ] obtained by the sensor to obtain the state vector and state estimation noise covariance matrix of the filter: X _k =X _k|k-1 +K _k (Z _k -HX _k|k-1 ) P _k =(IK _k H)X _k|k-1 Among them, X _k and P _k are used as inputs at the next moment to continuously predict the dose value; The method of using the forget gate to filter the information to be discarded is as follows: The output of the current input _xt and the previous output ht _-1 after the forget gate is: F _t =sigmod(W _f [h _t-1 ,x _t ]+b _f ) Among them, _Wf and _bf are the weight and bias of the forget gate; The method by which the input gate determines the new information in the storage unit state is as follows: The sigmoid part determines the update of the value and outputs the vector _it . The tanh part creates a new candidate value ~ _Ct and adds it to the state. The current state is: i _t =sigmoid(W _i [h _t-1 ,x _t ]+b _i ) Among them, _Wi and _bi are the weight and bias of the input gate, _Wn and _bn are the weight and bias of the tanh part; The method by which the output gate determines the output of the current unit is as follows: After the previous moment's output h _t-1 and x _t are processed, the output o _t and the current moment's state C _t are multiplied by the tanh part to obtain the final output: h _t = o _t tanh(C _t ) o _t =sigmoid(W _o [h _t-1 ,x _t ]+b ₀ ) Where W _o and b _o are the weight and bias of the input gate. 8. The real-time consumption monitoring system for the finished slurry tank according to claim 6 is characterized in that the process of collecting the consumption of the slurry tank and the sensor feeding speed data in the current slurry mixing process in the data collection and processing module is as follows: the time series length of the collected consumption is N, let _Li = {c _r,k , k = 1, 2, ..., N} be the time series data set of the i-th consumption in the data set, _vi = {v _r,k , k = 1, 2, ..., N} be the real feeding speed set, where cr _,k and v _r,k are the real consumption and real feeding speed of the sensor respectively; The process of processing the collected data is as follows: Use a sliding window to extract data of length w _L from the usage time series dataset _Li and obtain: I＝{c _r,k ,k＝j,j+1,...,W _L+j-1 } Use the formula to perform differential processing on the amount and convert the input amount into an average feeding speed: Considering that the target instantaneous feeding speed and the expected average feeding speed also have a correlation in time series characteristics, when predicting the average feeding speed, the instantaneous feeding speed is also used as input. Then the j-th input vector of the LSTM for predicting the average feeding speed is: Use the sliding window to extract the length _wL of the material feeding speed data from the material feeding speed dataset v _i , and get the jth input vector of the LSTM model for predicting the instantaneous material feeding speed: I _V ={v _r,k ,k=j,...,W _L+j-1 } The data is normalized to obtain the normalized instantaneous material feeding speed and average material feeding speed time series, namely I′ _L and I′ _v , where the normalization formula is as follows: The inputs of the constructed LSTM for predicting the average feeding speed of the usage and the LSTM for predicting the instantaneous feeding speed of the usage are respectively, where j ranges from 1 to N: I′ _V,j ={v′ _r,k ,k＝1,...,W _L } Among them, I' represents the normalized data set, _Ii represents the time series data set, and _Imax and _Imin represent the minimum values in each feature sequence. 9. The real-time consumption monitoring system for finished slurry tanks according to claim 6 is characterized in that the process of training, verifying and testing the LSTM-KF algorithm described in the model training module is as follows: using the root mean square error as the loss function to minimize the average feeding speed and the LSTM model that predicts the instantaneous feeding speed; the learning rate uses the Adam optimizer to perform adaptive learning rate adjustment, and sets the current batch size at the same time; the LSTM model is trained using the training set data, and the convergence of the model is verified by visualizing the loss function; the validation set data is used for five-fold cross validation to verify the generalization of the model; the test set data is used to test the classification effect of the time series data and the model fitting; The formula of the loss function is as follows: Among them, L represents the root mean square error calculation, N represents the length of the input vector, v _r represents the input dosage and feeding speed, and v _p represents the feeding speed predicted by the model. 10. The real-time consumption monitoring system for the finished product slurry tank according to claim 6 is characterized in that the steps of using the trained LSTM-KF algorithm to continuously predict the real-time consumption of the finished product slurry tank in the model application module are as follows: (1) The input vector is constructed through preprocessing and input into the LSTM model. The LSTM model predicts the average material feeding speed and instantaneous material feeding speed of the target respectively. After the prediction is completed, the denormalization is performed to obtain the average material feeding speed prediction. And the instantaneous feeding speed prediction value (2) The uniform motion model used by the KF algorithm takes the filtered material feeding speed v _f,k-1 at the previous moment as the average material feeding speed between moment k-1 and moment k, and calculates the usage prediction result at moment k; similarly, the KF algorithm directly takes the filtered material feeding speed at moment k-1 as the instantaneous material feeding speed at moment k; The calculation formula for the usage prediction result at the k moment is: (3) The LSTM-KF algorithm uses the instantaneous material feeding speed predicted by LSTM And the corrected speed prediction v _f,k-1 , through the KF algorithm, the final feeding speed prediction v _p,k is obtained, and the calculation formula is: v _p,k =U(v _p,k ,v _f,k-1 ) (4) The KF algorithm is used again to combine the predicted consumption c _p,k and material feeding speed v _p,k at time k with the measured consumption c _m,k and material feeding speed v _m,k at time k to obtain the final consumption estimate and material feeding speed estimate at time k, which are: c _f,k =U(c _p,k ,c _m,k ) v _f,k =U(v _p,k ,v _m,k ) (5) Repeat the above steps to achieve continuous prediction of the real-time usage of the finished slurry tank.