WO2023123779A1

WO2023123779A1 - Dynamic system identification-based product quality prediction method and device for rectification column

Info

Publication number: WO2023123779A1
Application number: PCT/CN2022/090171
Authority: WO
Inventors: 梁新乐; 王峰
Original assignee: 无锡雪浪数制科技有限公司
Priority date: 2021-12-30
Filing date: 2022-04-29
Publication date: 2023-07-06
Also published as: CN114240243B; CN114240243A

Abstract

A dynamic system identification-based product quality prediction method and device for a rectification column, relating to the field of product quality early warning. The method comprises: inputting current state data and planned input data into a dynamic system identification model to obtain a prediction state sequence, the prediction state sequence comprising state data of a rectification column within a predetermined duration in the future, and the dynamic system identification model being obtained by training using historical operation data of the rectification column and on the basis of an SINDy model; and inputting the prediction state sequence into a time series anomaly detection model obtained by pretraining using the historical operation data of the rectification column, so as to obtain a corresponding loss value, then outputting, when the loss value exceeds a loss threshold, early warning information used to warn that the product quality of the rectification column is abnormal. According to the method, accurate prediction of the state of the rectification column can be effectively performed by using the dynamic system identification model, and in combination with an unsupervised time series anomaly detection model, the product quality can be effectively predicted.

Description

A method and device for predicting product quality of a distillation column based on dynamic system identification

technical field

The invention relates to the field of product quality early warning, in particular to a rectification tower product quality prediction method and device based on dynamic system identification.

Background technique

Rectification is the most widely used separation method in chemical production. It uses gas-liquid two-phase mass transfer and heat transfer to achieve the purpose of separation, and rectification tower is currently the mainstream equipment used to realize rectification operation. It is widely used in existing chemical enterprises. Due to the wide range of applications of the rectification tower, it is of great significance to predict the product quality of the rectification tower. By predicting the product quality of the rectification tower, it can be controlled and optimized in time, so that it can operate smoothly and improve product quality. rate, reducing the loss of high-value components in low-value products.

However, there are two main difficulties in the product quality prediction of the rectification tower. On the one hand, the production process of the rectification tower is a dynamic and unsteady state process, and it is difficult to accurately predict the production process. On the other hand, the existing time series abnormal warning The model is mainly a supervised model, but in the production process of chemical enterprises, there are fewer samples with poor product quality. This is because in normal production enterprises, in order to ensure normal production, the product yield must be guaranteed at Only a certain range can be mass-produced, which leads to too little product quality abnormal data accumulated in the normal production process, and the proportion of abnormal samples is far lower than the proportion of normal samples. For supervised time series prediction models, It is difficult to use the current samples to train a more effective prediction model, so there is no better solution for the current product quality prediction of the distillation column.

Contents of the invention

In view of the above-mentioned problems and technical requirements, the present inventor proposes a method and device for predicting product quality of a rectifying tower based on dynamic system identification. The technical scheme of the present invention is as follows:

In the first aspect, a method for predicting product quality of a distillation column based on dynamic system identification is requested for protection, the method includes:

Determine the current state data and planned input data of the rectification tower, the state data of the rectification tower is data used to reflect the operating state of the rectification tower, and the input data of the rectification tower is the data of the input variables of the rectification tower;

The current state data and planned input data are input into the dynamic system identification model to obtain the predicted state sequence, which includes the state data of the rectification tower within a predetermined period of time in the future. The dynamic system identification model uses the historical operation data of the rectification tower based on the SINDy model trained to get;

Input the predicted state sequence into the time series anomaly detection model pre-trained using the historical operation data of the distillation column to obtain the corresponding loss value;

When the obtained loss value exceeds the loss threshold, an early warning message for warning that the product quality of the rectification tower is abnormal is output.

Its further technical solution is that the historical operation data used to train the dynamic system identification model includes the historical state sequence and historical input sequence of the rectification tower, and the historical state sequence includes the state of each working state point of the rectification tower arranged in chronological order Data, the historical input sequence includes the input data of each working status point of the rectification tower arranged in chronological order;

Then the method also includes:

The state data x _t-1 of the t-1th working state point in the historical state sequence of the distillation column and the input data u t _-1 of the t-1th working state point in the historical input sequence are used as the SINDy model Input, take the historical state sequence x _t of the tth working state point in the historical state sequence as the output of the SINDy model, and set the regression model of the SINDy model as LASSO, and use the historical state sequence and historical input sequence of the distillation column to train A dynamic system identification model is obtained.

Its further technical solution is that the historical operation data used to train the time series anomaly detection model includes the windowed historical state sequence of the rectification tower, and the windowed historical state sequence includes the chronological order of the rectification tower within the predetermined time window range The state data of each working state point of and expressed as x _{[i,i+1,...i+N]} ;

Then the method also includes:

Input the windowed historical state sequence of the distillation column into the LSTM_VAE model to obtain the corresponding recombination sequence x'_{[i,i+1,...i+N]};

Update the network parameters of the LSTM_VAE model according to the error between the windowed historical state sequence x _{[i,i+1,…i+N]} and its corresponding recombined sequence x′ _{[i,i+1,…i+N]} , Until the time series anomaly detection model is obtained through training.

Its further technical scheme is that the method also includes:

Calculate the recombination loss according to the windowed historical state sequence x _{[i,i+1,...i+N]} and its corresponding recombination sequence x′ _{[i,i+1,...i+N]} , and calculate the KL divergence of VAE , weight the calculated recombination loss and KL divergence according to their corresponding weights to obtain the windowed historical state sequence x _{[i,i+1,…i+N]} and its corresponding recombination sequence x′ _{[i,i+ 1,...i+N]} .

Its further technical scheme is that the method also includes:

Input the K training samples obtained by constructing the historical operation data of the rectification tower into the time series anomaly detection model obtained by training, and obtain the corresponding loss value. The windowed historical state sequence of the state data of each working state point arranged in chronological order;

Among the loss values corresponding to all training samples, the kth loss value from high to low is used as the loss threshold, and the ratio of k/K is the preset abnormal sample ratio.

Its further technical scheme is that the state data of the rectification tower includes at least one of the temperature of the sensitive plate in the rectification section, the pressure at the top of the rectification tower, the cold return flow at the top of the tower, and the pressure at the outlet of the pump, and the input data of the rectification tower includes at least one of the At least one of the amount of material, the current of the regulating valve and the opening degree of the regulating valve of the tower top product.

In the second aspect, a device for predicting the product quality of a distillation column based on dynamic system identification is claimed, which includes:

The data determination module is used to determine the current state data and plan input data of the rectification tower. The state data of the rectification tower is the data used to reflect the operation state of the rectification tower. The input data of the rectification tower is the input of the rectification tower variable data;

The predicted state sequence acquisition module is used to input the current state data and planned input data into the dynamic system identification model to obtain the predicted state sequence. The predicted state sequence includes the state data of the rectification column within a predetermined time in the future. The historical operation data of the tower is obtained based on the SINDy model training;

The loss value acquisition module is used to input the predicted state sequence into the time series anomaly detection model pre-trained using the historical operation data of the rectification tower to obtain the corresponding loss value;

The early warning module is configured to output early warning information for warning that the product quality of the rectification tower is abnormal when the obtained loss value exceeds the loss threshold.

A computer device, including a memory and a processor, the memory stores a computer program, and is characterized in that, when the processor executes the computer program, it implements the steps of a rectification column product quality prediction method based on dynamic system identification claimed in the first aspect .

A computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of a method for predicting product quality of a rectifying column based on dynamic system identification claimed in the first aspect are realized.

A computer program product, including a computer program, when the computer program is executed by a processor, it implements the steps of a method for predicting product quality of a distillation column based on dynamic system identification claimed in the first aspect.

The beneficial technical effect of the present invention is:

This application discloses a rectification column product quality prediction method and device based on dynamic system identification, using the dynamic system identification model trained based on the historical operation data of the rectification column and using the SINDy model to predict the future period based on the current system state and input The state of the time system, and then use the unsupervised time series anomaly detection model to detect the anomaly of the system change time series in the future, which can effectively predict the state of the distillation column accurately and effectively predict the product quality.

Description of drawings

Fig. 1 is a flow chart of a method for predicting product quality of a rectification column disclosed in an embodiment of the present application.

Figure 2 is a schematic diagram of the structure of a rectification tower.

Fig. 3 is a flow chart of a method for predicting product quality of a rectification column disclosed in another embodiment of the present application.

Fig. 4 is a device structure diagram of a rectification tower product quality prediction device disclosed in an embodiment of the present application.

Fig. 5 is a device structure diagram of a rectification column product quality prediction device disclosed in another embodiment of the present application.

Fig. 6 is a device structure diagram of a computer device disclosed in the present application.

Detailed ways

The specific embodiments of the present invention will be further described below in conjunction with the accompanying drawings.

This application discloses a method for predicting product quality of a distillation column based on dynamic system identification. The method includes the following steps, please refer to Figure 1:

Step S10, determining the current state data and plan input data of the rectification column.

Wherein, the state data of the rectification tower is data used to reflect the operation state of the rectification tower, and the current state data is the data reflecting the real-time operation state of the rectification tower. The input data of the rectification tower is the data of the input variable of the rectification tower, and then the planned input data is the sequence formed by the input data of the rectification tower within a certain period of time in the future.

In terms of specific data types, please refer to the schematic diagram of the rectification tower shown in Figure 2. The state data of the rectification tower includes the temperature T of the sensitive plate in the rectification section, the pressure P _t at the top of the rectification tower, the flow rate R of the cold reflux at the top of the tower, and the pump At least one of outlet pressure P _p . The input data of the rectification column includes at least one of the feed amount F, the current u of the regulating valve and the opening degree V _D of the regulating valve of the overhead product.

Step S20, input the current state data and plan input data into the dynamic system identification model to obtain the predicted state sequence, which contains the state data of the rectification column within a predetermined period of time in the future, that is, the predicted state sequence reflects the state data of the rectification tower in the future. state changes over time.

The dynamic system identification model is trained based on the SINDy model based on the historical operation data of the distillation column. This model can predict the system change sequence in the future according to the current state and the previous sequence input. The basic logic is to preserve the nonlinearity of adjacent state transitions. relation. The historical operation data used to train the dynamic system identification model includes the historical state sequence and historical input sequence of the rectification tower. The input data of each working state point of the distillation column arranged in chronological order.

Optionally, the method also includes a pre-model training process for the dynamic system identification model, please refer to the flow chart shown in Figure 3:

The state data x _t-1 of the t-1th working state point in the historical state sequence of the rectification column and the input data u t _-1 of the t-1th working state point in the historical input sequence are taken as SINDy (not Sparse identification of linear dynamical systems, Sparse Identification of Nonlinear Dynamical systems) model input, the historical state sequence x _t of the tth working state point in the historical state sequence is used as the output of the SINDy model, t is a parameter and t≥2. And set the regression model of the SINDy model as LASSO. The purpose of using LASSO is to generate a sparse regression model. The data corresponding to the working state point under each parameter t is used as the input and output according to the above method, and the model parameters are updated by comparing with the actual output of the model, and then the dynamic system identification model is obtained by using the historical state sequence and historical input sequence training of the distillation column , the specific model training process is relatively conventional, and will not be described in detail in this application.

Step S30, inputting the predicted state sequence into the time series anomaly detection model to obtain the corresponding loss value.

The time series anomaly detection model is pre-trained by using the historical operation data of the distillation tower. Marking the time period in advance can save a lot of manual knowledge, and it is very suitable for scenarios where there are few samples with poor product quality in the distillation column field.

The historical operation data used to train the time series anomaly detection model includes the windowed historical state sequence of the rectification column, and the windowed historical state sequence includes the state of each working state point of the rectification column arranged in chronological order within the predetermined time window range Data and expressed as x _{[i,i+1,…i+N]} , i is a parameter, and N is the time step of the predetermined time window range. And in general, several different windowed historical state sequences are constructed from the historical state sequences of the distillation column as training samples for model training.

Optionally, the method also includes a pre-model training process for the time series anomaly detection model, please refer to Figure 3:

Input the windowed historical state sequence of the distillation column into the LSTM_VAE model to obtain the corresponding recombination sequence x′ _{[i,i+1,…i+N]} . To calculate the error between the windowed historical state sequence x _{[i,i+1,…i+N]} and its corresponding recombined sequence x′ _{[i,i+1,…i+N]} , one way is to Directly calculate the recombination loss of x _{[i,i+1,...i+N]} and x′ _{[i,i+1,...i+N]} as the error. Another alternative is to calculate the reorganization based on the windowed historical state sequence x _{[i,i+1,…i+N]} and its corresponding reorganization sequence x′ _{[i,i+1,…i+N]} loss, and calculate the KL divergence of VAE, and weight the calculated recombined loss and KL divergence according to their corresponding weights to obtain the windowed historical state sequence x _{[i,i+1,...i+N]} and its corresponding The error between recombined sequences x' _{[i,i+1,...i+N]} . Use the calculated error to update the network parameters of the LSTM_VAE model, and then input x _{[i,i+1,…i+N]} into the LSTM_VAE model after updating the network parameters, and re-execute the above process until the error is within the preset error range , train the required time series anomaly detection model.

In step S40, when the obtained loss value exceeds the loss threshold, output warning information for warning that the product quality of the rectification tower is abnormal. If it is not exceeded, it means that the product quality of the rectification tower is normal, and relevant results can be output or not.

The loss threshold can be a custom preset value. Alternatively, after the time series anomaly detection model is obtained through training, K training samples obtained by constructing the historical operation data of the rectification tower are respectively input into the trained time series anomaly detection model to obtain corresponding loss values. Each training sample is a windowed historical state sequence including the state data of each working state point of the rectification column arranged in chronological order within the predetermined time window range, and the K training samples here can directly use the above-mentioned training time series Training samples for the anomaly detection model.

Among the loss values corresponding to all training samples, the kth loss value from high to low is used as the loss threshold, and the ratio of k/K is the preset abnormal sample ratio. The preset abnormal sample ratio is a preset ratio value. For example, if the preset abnormal sample ratio is set to 2%, and there are 10,000 training samples in total, then the loss values of the 10,000 training samples are sorted from high to low, and selected from The loss value at the 10000th*2%=200th position from the highest to the lowest is the loss threshold.

Based on the same inventive concept, the embodiment of the present application also provides a rectification tower product quality prediction device based on dynamic system identification for realizing the above-mentioned dynamic system identification-based rectification tower product quality prediction method. The solution to the problem provided by the device is similar to the implementation described in the above method, so the specific limitations in one or more embodiments of the rectification tower product quality prediction device based on dynamic system identification provided below can be found in The above-mentioned limitations on the method for predicting the product quality of a rectification column based on dynamic system identification will not be repeated here.

In one embodiment, as shown in Figure 4, there is provided a rectification column product quality prediction device based on dynamic system identification, including a data determination module, a prediction state sequence acquisition module, a loss value acquisition module and an early warning module, wherein:

In one embodiment, please refer to FIG. 5 , the historical operation data used for training the dynamic system identification model includes the historical state sequence and the historical input sequence of the rectification tower, and the historical state sequence includes each work of the rectification tower in chronological order The state data of the state points, the historical input sequence includes the input data of each working state point of the rectification column in chronological order. Then the device also includes a first model training module, which is used to combine the state data x _t-1 of the t-1th working state point in the historical state sequence of the rectification column and the t-1th working state point in the historical input sequence The input data u _t-1 of the state point is taken as the input of the SINDy model, and the historical state sequence x _t of the tth working state point in the historical state sequence is taken as the output of the SINDy model, and the regression model of the SINDy model is set as LASSO, The dynamic system identification model is obtained by training the historical state sequence and historical input sequence of the distillation column.

In one embodiment, the historical operation data used for training the time series anomaly detection model includes a windowed historical state sequence of the rectification column, and the windowed historical state sequence includes the chronological order of the rectification column within a predetermined time window range The status data of each working status point is expressed as x _{[i,i+1,…i+N]} . Then the device also includes a second model training module, which is used to input the windowed historical state sequence of the rectification column into the LSTM_VAE model to obtain the corresponding recombined sequence x' _{[i, i+1,...i+N]} ; according to the windowed history The error between the state sequence x _{[i,i+1,…i+N]} and its corresponding recombination sequence x′ _{[i,i+1,…i+N]} updates the network parameters of the LSTM_VAE model until the training time Sequential Anomaly Detection Models.

In one embodiment, the second training module is also used for windowing the historical state sequence x _{[i,i+1,...i+N]} and its corresponding reorganization sequence x' _{[i,i+1,...i+ N]} Calculate the reorganization loss, and calculate the KL divergence of VAE, and weight the calculated recombination loss and KL divergence according to their corresponding weights to obtain the windowed historical state sequence x _{[i,i+1,...i+N]} and its corresponding recombined sequence x′ _{[i,i+1,...i+N]} .

In one embodiment, the second training module is also used to input the K training samples obtained by constructing the historical operation data of the rectification tower into the time series anomaly detection model obtained by training respectively to obtain the corresponding loss value, and each training sample It is a windowed historical state sequence including the state data of each working state point of the rectification column arranged in chronological order within the predetermined time window; the kth loss value from high to low among the loss values corresponding to all training samples As the loss threshold, the ratio of k/K is the preset abnormal sample ratio.

Each module in the above-mentioned rectification column product quality prediction device based on dynamic system identification can be fully or partially realized by software, hardware and combinations thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, and can also be stored in the memory of the computer device in the form of software, so that the processor can invoke and execute the corresponding operations of the above-mentioned modules.

In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure may be as shown in FIG. 6 . The computer device includes a processor, memory and a network interface connected by a system bus. Wherein, the processor of the computer device is used to provide calculation and control capabilities. The memory of the computer device includes non-volatile storage media and memory. The non-volatile storage medium stores an operating system, computer programs and databases. The memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the computer device is at least used for storing the dynamic system identification model and the time series anomaly detection model obtained through pre-training. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by a processor, a method for predicting product quality of a distillation column based on dynamic system identification is realized.

Those skilled in the art can understand that the structure shown in FIG. 6 is only a block diagram of a part of the structure related to the solution of this application, and does not constitute a limitation on the computer equipment to which the solution of this application is applied. The specific computer equipment can be More or fewer components than shown in the figures may be included, or some components may be combined, or have a different arrangement of components.

In one embodiment, a computer device is provided, including a memory and a processor, a computer program is stored in the memory, and the processor implements the following steps when executing the computer program:

Inputting the current state data and plan input data into a dynamic system identification model to obtain a predicted state sequence, the predicted state sequence includes the state data of the rectification tower within a predetermined time in the future, and the dynamic system identification model uses the The historical operation data of the distillation column is obtained based on the SINDy model training;

Inputting the predicted state sequence into a time series anomaly detection model obtained by pre-training the historical operation data of the rectification tower to obtain a corresponding loss value;

In one embodiment, the historical operation data used for training the dynamic system identification model includes the historical state sequence and historical input sequence of the rectification column, and the historical state sequence includes the state data of each working state point of the rectification column in chronological order , the historical input sequence includes the input data of each working state point of the rectification tower in chronological order; when the processor executes the computer program, it also realizes the following steps: the t-1th working state in the historical state sequence of the rectification tower The state data x _t-1 of the state point and the input data u _t-1 of the t-1th working state point in the historical input sequence are used as the input of the SINDy model, and the history of the t-th working state point in the historical state sequence The state sequence x _t is taken as the output of the SINDy model, and the regression model of the SINDy model is set as LASSO, and the dynamic system identification model is obtained by training the historical state sequence and historical input sequence of the distillation column.

In one embodiment, the historical operation data used for training the time series anomaly detection model includes a windowed historical state sequence of the rectification column, and the windowed historical state sequence includes the chronological order of the rectification column within a predetermined time window range The state data of each working state point is expressed as x _{[i,i+1,...i+N]} ; when the processor executes the computer program, the following steps are also implemented: input the windowed historical state sequence of the rectification column into the LSTM_VAE model to obtain The corresponding recombination sequence x′ _{[i,i+1,…i+N]} ; according to the windowed historical state sequence x _{[i,i+1,…i+N]} and its corresponding recombination sequence x′ _{[i,i +1,...i+N]} to update the network parameters of the LSTM_VAE model until the time series anomaly detection model is trained.

In one embodiment, the processor also implements the following steps when executing the computer program: according to the windowed history state sequence x _{[i, i+1,...i+N]} and its corresponding recombination sequence x′ _{[i, i+ 1,...i+N]} Calculate the reorganization loss, and calculate the KL divergence of VAE, and weight the calculated recombination loss and KL divergence according to their corresponding weights to obtain the windowed historical state sequence x _{[i,i+1, ...i+N]} and its corresponding recombination sequence x′ _{[i,i+1,...i+N]} .

In one embodiment, the processor also implements the following steps when executing the computer program: input K training samples obtained by constructing the historical operation data of the rectification tower into the time series anomaly detection model obtained by training respectively, and obtain corresponding loss values , each training sample is a windowed historical state sequence including the state data of each working state point of the rectification column arranged in chronological order within the predetermined time window range; for all training samples corresponding to the loss value from high to low The kth loss value is used as the loss threshold, and the ratio of k/K is the preset abnormal sample ratio.

In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the steps in the foregoing method embodiments are implemented.

In one embodiment, a computer program product is also provided, including a computer program, when the processor executes the computer program, the steps in the foregoing method embodiments are implemented.

What is described above is only a preferred embodiment of the present application, and the present invention is not limited to the above examples. It can be understood that other improvements and changes directly derived or conceived by those skilled in the art without departing from the spirit and concept of the present invention should be considered to be included in the protection scope of the present invention.

Claims

A method for predicting product quality of a distillation tower based on dynamic system identification, characterized in that the method comprises:

Determine the current state data and planned input data of the rectification tower, the state data of the rectification tower is data used to reflect the operating state of the rectification tower, and the input data of the rectification tower is the data of the input variables of the rectification tower;

Inputting the current state data and plan input data into a dynamic system identification model to obtain a predicted state sequence, the predicted state sequence includes the state data of the rectification tower within a predetermined time in the future, and the dynamic system identification model uses the The historical operation data of the distillation column is obtained based on the SINDy model training;

Inputting the predicted state sequence into a time series anomaly detection model obtained by pre-training the historical operation data of the rectification tower to obtain a corresponding loss value;

When the obtained loss value exceeds the loss threshold, output warning information for warning that the product quality of the rectification tower is abnormal.
The method according to claim 1, wherein the historical operation data used for training the dynamic system identification model includes the historical state sequence and historical input sequence of the rectification column, and the historical state sequence includes the rectification column. The state data of each working state point of the distillation column arranged in chronological order, the historical input sequence includes the input data of each working state point of the distillation column arranged in chronological order;

Then described method also comprises:

The state data x t- 1 of the t-1th working state point in the historical state sequence of the rectification column and the input data u t - 1 of the t-1th working state point in the historical input sequence As the input of the SINDy model, the historical state sequence x t of the tth working state point in the historical state sequence is used as the output of the SINDy model, and the regression model of the SINDy model is set as LASSO, utilizing the rectifying tower The historical state sequence and historical input sequence are trained to obtain the dynamic system identification model.
The method according to claim 1, wherein the historical operation data used for training the time series anomaly detection model includes a windowed historical state sequence of the rectification column, and the windowed historical state sequence includes a predetermined time The state data of each working state point arranged in chronological order of the rectifying tower within the window range is expressed as x [i, i+1,...i+N] ;

Then described method also comprises:

Inputting the windowed historical state sequence of the rectification column into the LSTM_VAE model to obtain the corresponding recombination sequence x' [i, i+1,...i+N] ;

Update the LSTM_VAE model according to the error between the windowed historical state sequence x [i,i+1,...i+N] and its corresponding recombined sequence x' [i,i+1,...i+N] The network parameters until the time series anomaly detection model is obtained through training.
The method according to claim 3, characterized in that the method further comprises:

Calculate the recombination loss according to the windowed historical state sequence x [i,i+1,...i+N] and its corresponding recombination sequence x' [i,i+1,...i+N] , and calculate the KL of VAE Divergence, weighting the calculated reorganization loss and KL divergence according to their corresponding weights to obtain the windowed historical state sequence x [i,i+1,...i+N] and its corresponding reorganization sequence x′ [ i,i+1,...i+N] .
The method according to claim 1, further comprising:

Input the K training samples obtained by constructing the historical operation data of the rectification tower into the time series anomaly detection model obtained by training respectively, and obtain the corresponding loss value, and each training sample is included in the precise time window within the predetermined time window. The windowed historical state sequence of the state data of each working state point of the distillation column arranged in chronological order;

The k-th loss value from high to low among the loss values corresponding to all training samples is used as the loss threshold, and the ratio of k/K is the preset abnormal sample ratio.
The method according to any one of claims 1-5, characterized in that,

The state data of the rectification tower includes at least one of the temperature of the sensitive plate in the rectification section, the pressure at the top of the rectification tower, the cold reflux flow at the top of the tower, and the outlet pressure of the pump. At least one of the opening degrees of the overhead product regulating valve.
A rectification column product quality prediction device based on dynamic system identification, characterized in that the device comprises:

The data determination module is used to determine the current state data and plan input data of the rectification tower. The state data of the rectification tower is the data used to reflect the operation state of the rectification tower. The input data of the rectification tower is the input of the rectification tower variable data;

A predicted state sequence acquisition module, configured to input the current state data and planned input data into the dynamic system identification model to obtain a predicted state sequence, the predicted state sequence includes the state data of the rectification tower within a predetermined time period in the future, so The dynamic system identification model is obtained based on the SINDy model training using the historical operation data of the rectification tower;

The loss value acquisition module is used to input the predicted state sequence into the time series anomaly detection model obtained by pre-training the historical operation data of the rectification tower to obtain a corresponding loss value;

The early warning module is configured to output early warning information for warning that the product quality of the rectification tower is abnormal when the obtained loss value exceeds a loss threshold.
A computer device, comprising a memory and a processor, the memory stores a computer program, wherein the processor implements the steps of the method according to any one of claims 1 to 6 when executing the computer program.
A computer-readable storage medium, on which a computer program is stored, wherein, when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are realized.
A computer program product, comprising a computer program, characterized in that, when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.