Sensorless flow estimation method and system, electronic equipment and storage medium

Abstract

The invention discloses a sensorless flow estimation method and system, electronic equipment and a storage medium, and relates to the technical field of nuclear power and sensorless measurement, and the method comprises the steps: collecting non-flow process variable data in a conventional operation process of an actual reactor cooling system in real time; pre-processed non-flow process variable data serves as input data and is input into the trained feature extraction model and the flow prediction model, feature vectors irrelevant to the field are extracted through the feature extraction model, the feature vectors pass through the flow prediction model, online prediction is conducted, and the flow value of an actual system is output; wherein in the prediction process, whether self-adaptive transfer learning is triggered or not is judged according to monitoring data, the model is updated through the self-adaptive transfer learning, and the updated model is utilized to carry out flow value online prediction. According to the method, high-stability and high-precision flow estimation under any simple or complex working conditions can be realized.

Description

Technical Field

[0001] The present invention relates to the field of nuclear power and sensorless measurement technology, and in particular to a sensorless flow estimation method, system, electronic equipment and storage medium. Background Art

[0002] In recent years, the research and development of various integrated multi-purpose small reactors and the implementation of engineering projects have progressed rapidly, and some reactor types have entered the construction stage. Based on the integrated characteristics of these new reactors, the equipment has been highly integrated and factory-made, which greatly saves on-site assembly and welding work and layout space. However, the built-in features of heat exchangers and fluid flow channels also bring difficulties to the installation of sensors. For example, the primary circuit flow of the integrated pressure vessel is the core indicator of the reactor cooling system. Accurate measurement of this flow can ensure that the core reactivity of the process such as starting boron dilution is controlled, and there will be no unexpected instant criticality and other operating accidents or events. It can timely take away the heat generated by the core to prevent the core from overheating and melting, thereby ensuring the safety of the reactor. However, since the primary circuit flow channel is built into the pressure vessel and is not constructed by a pipeline system, traditional mature flow meters and sensors cannot be installed, and the compact layout inside the shell also makes the environmental conditions inside the shell more harsh, further limiting the feasibility of using sensors directly. In addition, real-time monitoring of key reactor characteristic parameters such as primary circuit flow can help operators better understand the thermal hydraulic state of the reactor, ensure that the flow of each cooling circuit fluctuates within the design range, avoid potential risks caused by insufficient flow, and improve the overall operational stability and reliability of the system; primary circuit flow and other measurements provide key data support for cooling efficiency analysis under steady-state and transient conditions, which can help optimize the design and operating conditions of the cooling system, especially under transient conditions (such as power changes or shutdown and startup processes), where flow changes are directly related to the cooling effect, and timely flow feedback can effectively support the safe regulation of the reactor. Therefore, sensorless measurement technology has the significance of ensuring reactor safety, improving reactor operational reliability, and supporting operating condition analysis and optimized operation for new nuclear power plants.

[0003] Existing sensorless measurement technologies include direct measurement and indirect measurement. However, for new nuclear power plants, existing methods still have certain limitations:

[0004] (1) Direct measurement method: The current direct measurement method is mainly based on the differential pressure method. However, due to the small flow rate of the primary circuit, the differential pressure signal is weak, which affects the measurement accuracy. In addition, the differential pressure method is not sensitive enough to small changes in flow rate. Under low flow conditions, it is more susceptible to interference from system noise and environmental factors, affecting the measurement accuracy.

[0005] (2) Regarding the indirect measurement method: The current indirect measurement method is mainly based on the driving force and resistance balance and the thermal balance equation, and is mainly applicable to steady-state conditions after the system starts heating. This method cannot achieve accurate measurement under transient conditions because the driving force and resistance change dynamically during the transient process, and thermal balance is difficult to maintain, resulting in the model being unable to adapt to these non-steady-state characteristics. This limits the application of the indirect measurement method under normal operating conditions and makes the flow measurement lack dynamic response capabilities. Summary of the invention

[0006] In order to solve the deficiencies of the above-mentioned prior art, the present invention provides a sensorless flow estimation method, system, electronic device and storage medium to make up for the deficiencies of the existing direct and indirect measurement methods. In the absence of direct flow measurement equipment or the accuracy of the equipment is insufficient, through the combination of data-driven model and mechanism model, a single-loop flow estimation method suitable for nuclear power plants under steady-state and transient conditions is proposed, so as to achieve high-stability and high-precision flow estimation under any simple or complex conditions.

[0007] In a first aspect, the present invention provides a sensorless flow estimation method.

[0008] A sensorless flow estimation method, comprising:

[0009] Real-time acquisition of non-flow process variable data during the normal operation of the actual reactor cooling system and data preprocessing;

[0010] The pre-processed non-flow process variable data is used as input data and input into the trained feature extraction model and flow prediction model. The feature extraction model extracts the feature vector that is not related to the field. The feature vector is then passed through the flow prediction model to predict and output the flow value of the actual system online.

[0011] In the prediction process, the input and output data of the model are monitored in real time. When the monitoring data meets the trigger conditions of adaptive transfer learning, the non-flow process variable data of the actual system collected recently are used to fine-tune the model parameters of the feature extraction model and the flow prediction model based on encoding reconstruction loss, feature alignment loss, pseudo-label loss, and prediction loss. The model is updated and the updated model is used to perform online prediction of the flow value.

[0012] A further technical solution sets multiple adaptive transfer learning trigger conditions, including that the distribution change of input data is greater than a first set value, the flow prediction error is greater than a second set value, and the feature space similarity is less than a third set value;

[0013] Based on the monitoring data, when at least two trigger conditions are met, adaptive transfer learning is triggered to perform fine-tuning training and update of model parameters.

[0014] A further technical solution is to determine whether there is an offset based on the distribution change of the input data monitored in real time, wherein the distribution change of the input data is: the maximum mean difference between the input data in the current time window and the previous time window, or the mean and variance change values ​​of the input data in the current window and the previous window extracted based on the sliding window;

[0015] The model accuracy is determined based on the traffic prediction error monitored in real time, where the traffic prediction error is: the mean square error of the pseudo-label prediction errors at the previous moment and the current moment in steady state; the pseudo-label refers to the predicted traffic value output by the traffic prediction model;

[0016] Whether the current feature vector has changed is determined based on the feature space similarity, wherein the feature space similarity is the cosine similarity of the feature vectors generated in the current window and the previous window extracted based on the sliding window.

[0017] According to a further technical solution, the trigger condition also includes a time trigger condition, and the time trigger condition has the highest priority. The time trigger condition is that the time interval for model update is not less than a set time threshold.

[0018] According to a further technical solution, the data preprocessing includes: deleting erroneous data, filling missing values, transforming timestamps and filtering data.

[0019] In a further technical solution, the non-flow process variables use key characteristic variables for flow prediction, and the key characteristic variables include pressure and temperature; the screening of the key characteristic variables includes:

[0020] Install flow measurement sensors and various types of sensors in the prototype system to collect flow variable data and non-flow variable data at each measuring point under various working condition experiments; the non-flow variables include temperature and pressure;

[0021] Preprocess the collected data of different variables, use the attention mechanism to identify the delay relationship between different variables, and generate delay compensation variable data;

[0022] Based on the delay-compensated data, the Gaussian mixture model is used to fit the joint probability distribution of traffic and each non-traffic variable, so as to calculate the mutual information between traffic and each non-traffic variable;

[0023] Based on mutual information and the preset artificial experience weights, the comprehensive score of each variable is calculated, and the variables most relevant to traffic are selected as key feature variables according to the comprehensive score.

[0024] In a further technical solution, the feature extraction model adopts an encoder-decoder structure, and the traffic prediction model adopts a fully connected neural network architecture; the training process of the feature extraction model and the traffic prediction model includes:

[0025] Collect non-flow variable data of each measuring point of the actual system under various working condition experiments; the non-flow variables include temperature and pressure;

[0026] Preprocess the non-flow variable data collected by the prototype system and the actual system during the test phase, input the preprocessed data into the encoder to extract feature vectors, and then input the feature vectors into the encoder to reconstruct the non-flow variable data;

[0027] Train the encoder-decoder model by constructing a loss function based on the error between the input data and the reconstructed data;

[0028] The trained encoder is used to extract the feature vector of the non-flow variable data of the prototype system, and the feature vector is input into the flow prediction model to output the predicted flow value;

[0029] A loss function is constructed based on the mean square error between the predicted flow value and the actual flow value of the prototype system to train the flow prediction model.

[0030] In a second aspect, the present invention provides a sensorless flow estimation system.

[0031] A sensorless flow estimation system, comprising:

[0032] Data acquisition and preprocessing module, used to collect non-flow process variable data in real time during the normal operation of the actual reactor cooling system, and perform data preprocessing;

[0033] The flow prediction module is used to input the pre-processed non-flow process variable data into the trained feature extraction model and flow prediction model, extract the feature vector that is not related to the field through the feature extraction model, and then pass the feature vector through the flow prediction model to predict and output the flow value of the actual system online;

[0034] In the prediction process, the input and output data of the model are monitored in real time. When the monitoring data meets the trigger conditions of adaptive transfer learning, the non-flow process variable data of the actual system collected recently are used to fine-tune the model parameters of the feature extraction model and the flow prediction model based on encoding reconstruction loss, feature alignment loss, pseudo-label loss, and prediction loss. The model is updated and the updated model is used to perform online prediction of the flow value.

[0035] In a third aspect, the present invention further provides an electronic device.

[0036] An electronic device, comprising:

[0037] A memory for storing executable instructions;

[0038] The processor is used to implement the sensorless flow estimation method provided by the first aspect when executing the executable instructions stored in the memory.

[0039] In a fourth aspect, the present invention also provides a computer-readable storage medium.

[0040] A computer-readable storage medium stores executable instructions for causing a processor to execute the instructions to complete the sensorless flow estimation method provided in the first aspect.

[0041] One or more of the above technical solutions have the following beneficial effects:

[0042] 1. The present invention provides a sensorless flow estimation method, system, electronic device and storage medium, which are aimed at the primary circuit flow estimation of nuclear power plants under steady-state and transient conditions. By combining data-driven models and mechanism models, and combining dynamic adaptive transfer learning, high-stability, high-adaptability and high-precision flow estimation under simple or complex arbitrary conditions can be achieved without a flow sensor. The shortcomings of the prior art in terms of measurement accuracy, dynamic adaptability and cross-system versatility are solved, and the method has high-precision, strong adaptability and domain-independent flow prediction capabilities, which are suitable for sensorless system flow estimation and model migration between multiple systems, and significantly improve the dynamic response capability and application breadth of flow measurement. The sensorless measurement method can also be migrated to other similar situations for use.

[0043] 2. The present invention realizes high-precision flow estimation without a flow sensor by combining the encoder-decoder structure and the flow prediction model. The encoder extracts stable feature representations from non-flow process variables to achieve accurate prediction of system flow and solve the measurement accuracy problem of the existing differential pressure method under small signals. The present invention can still stably predict flow under complex working conditions through the encoder feature extraction capability and the sensorless prediction technical architecture.

[0044] 3. The present invention uses adaptive transfer learning to monitor the distribution changes of input data in real time, and automatically triggers model updates when system operating conditions change. In this way, the model can adapt to dynamic transient conditions, fill the measurement gap of existing indirect measurement methods under non-steady-state characteristics, and improve the dynamic response capability of the system. The present invention performs adaptive transfer learning through the design of the above-mentioned update trigger conditions and joint optimization objectives, and can maintain high-precision and stable flow prediction under different operating conditions.

[0045] 4. The encoder of the present invention ensures the extraction of domain-independent features through contrast learning or feature alignment, making it suitable for systems of different working conditions or different models, and achieving cross-system versatility. This feature overcomes the limitations of existing methods in cross-system applications, significantly reduces dependence on flow sensors, and provides a stable and reliable feature basis for flow prediction in multiple systems.

[0046] 5. The present invention designs a comprehensive data cleaning, preprocessing and storage process to ensure data quality, time series consistency and traceability, and the data is stored in time series and relational databases to facilitate subsequent query and analysis. This solution ensures the reliability of model input data, effectively improves the accuracy and efficiency of traffic prediction, and provides high-quality data support for subsequent modeling.

[0047] 6. The update trigger mechanism proposed in the present invention reasonably sets the time interval and error monitoring standard to avoid frequent invalid updates, thereby saving computing resources; the error monitoring and update interval control based on the sliding window enable the present invention to have excellent real-time performance and is suitable for industrial application scenarios with long-term stable operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0049] Figure 1 The overall architecture diagram of the sensorless flow estimation method according to an embodiment of the present invention;

[0050] Figure 2 Schematic diagram of the process of the encoder-decoder training stage in an embodiment of the present invention;

[0051] Figure 3 Schematic diagram of the flow chart of the flow prediction model training phase in an embodiment of the present invention;

[0052] Figure 4 It is a flowchart of the actual system operation stage in an embodiment of the present invention;

[0053] Figure 5 The figure is a flow chart of the sensorless flow estimation method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0054] It should be noted that the following detailed descriptions are exemplary only, are intended to describe specific embodiments, are intended to provide further explanation of the present invention, and are not intended to limit exemplary embodiments according to the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those of ordinary skill in the art to which the present invention belongs. In addition, it should also be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0055] Embodiment 1

[0056] In view of the problems of low measurement accuracy and lack of dynamic response capability in the existing sensorless measurement technology, this embodiment provides a sensorless flow estimation method, which can effectively improve the accuracy of flow estimation. The method specifically includes the following steps:

[0057] Real-time acquisition of non-flow process variable data during the normal operation of the actual reactor cooling system and data preprocessing;

[0058] The pre-processed non-flow process variable data is used as input data and input into the trained feature extraction model and flow prediction model. The feature extraction model extracts the feature vector that is not related to the field. The feature vector is then passed through the flow prediction model to predict and output the flow value of the actual system online.

[0059] In the prediction process, the input and output data of the model are monitored in real time. When the monitoring data meets the trigger conditions of adaptive transfer learning, the non-flow process variable data of the actual system collected recently are used to fine-tune the model parameters of the feature extraction model and the flow prediction model based on encoding reconstruction loss, feature alignment loss, pseudo-label loss, and prediction loss. The model is updated and the updated model is used to perform online prediction of the flow value.

[0060] The sensorless flow estimation method proposed in this embodiment is introduced in more detail through the following content.

[0061] like Figure 5 As shown, the sensorless flow estimation method proposed in this embodiment can be specifically divided into a prototype system test phase, an actual system test and data-driven model building phase, and an actual system operation phase. In the prototype system test phase, non-flow variables (or features) related to the flow variable are screened out; in the actual system test and data-driven model building phase, a flow prediction model is constructed and trained; in the actual system operation phase, the flow prediction model is updated according to the actual working conditions, and the flow prediction model is used to accurately predict the flow online according to the actual system operation data.

[0062] Specifically:

[0063] (1) Prototype system testing phase

[0064] (1.1) Sensor installation and flow measurement

[0065] like Figure 1 As shown in the figure, a flow measurement sensor is installed on the test prototype system, and the sensor is used to collect the test data of the prototype system. Among them, a variety of redundant measurement methods can be used, including resistance measurement, driving force fitting measurement, descending section throttling scheme and water flow holes, etc., to ensure the high accuracy and reliability of flow data.

[0066] (1.2) Experimental data collection

[0067] For the reactor cooling prototype system, various operating condition experiments including startup, steady-state and transient conditions are carried out, and flow variable data and non-flow variable data of each measuring point under each experimental condition are collected, where the non-flow variables include temperature, pressure, etc.

[0068] (1.3) Data preprocessing and storage

[0069] After data collection, comprehensive data preprocessing is carried out to ensure data dimension and time series consistency; after data preprocessing and cleaning, the data is stored in the database, which is divided into time series database and relational database to ensure efficient query and traceability of the test prototype system data.

[0070] The above data preprocessing includes: error data deletion, missing value filling, timestamp transformation and data filtering. The specific steps are as follows:

[0071] (1.3.1) Data cleaning

[0072] Through data cleaning, noise, errors and abnormal data are identified and removed to ensure data quality. The cleaning process includes the following steps:

[0073] (1.3.1.1) Error data deletion: Use statistical methods or reasonable range thresholds to identify and remove abnormal data points. For example, suppose the reasonable range of a certain measurement variable data x is [x min ,x max ], then the data validity is defined as: x∈[x min ,x max ], and data points outside this range are marked as abnormal and deleted.

[0074] (1.3.1.2) Missing value filling: When there are missing values ​​in the data, linear interpolation, mean value filling or time series interpolation (such as Lagrange interpolation) are used to fill them. Taking linear interpolation as an example, if the missing value is x i , then the filling formula is:

[0075] (1.3.2) Timestamp conversion

[0076] Unify the timestamps of each measurement point to ensure the consistency of the data in time sequence. That is, record the timestamp of each measurement point as t_i and convert the timestamps into a unified reference time axis.

[0077] Assume that the time interval between the measurement points is non-uniform Δt i , then linear interpolation can be used to map the data to a uniform time step Δt, and the new timestamp can be defined as:

[0078] t j =t0+j·Δt;

[0079] Among them, j is the number of time steps.

[0080] For any data point x(t i ), mapping it to a new timestamp t j The value x(t j ) is obtained by linear interpolation and can be expressed as:

[0081]

[0082] (1.3.3) Data filtering

[0083] Data filtering is used to remove high-frequency noise in the signal and retain valid information. Common filtering methods include moving average filtering and low-pass filtering, etc. Specifically:

[0084] (1.3.3.1) Moving average filtering: Assume that the sliding window size is N, then the filtered data x filtered (t) is expressed as:

[0085] (1.3.3.2) Low-pass filter: Design a first-order low-pass filter. Let the original data be x(t). The output y(t) after filtering satisfies the difference equation: y(t) = α·x(t) + (1-α)·y(t-1), where the filter coefficient α is defined as: Among them, τ is the filtering time constant and Δt is the sampling interval.

[0086] By filtering with a low-pass filter, the signal can be effectively smoothed and high-frequency noise can be removed.

[0087] As another implementation, the above data may also be subjected to data standardization processing to unify the data dimensions and ensure the accuracy of subsequent data processing and analysis.

[0088] (1.3.4) Data storage database

[0089] The data after the above cleaning and preprocessing are stored in the database. To ensure the traceability and convenient management of the data, the data storage is divided into two categories: time series database and relational database. The time series database is used to store the cleaned time series data for fast query and processing of multi-dimensional time series data; the relational database is used to store metadata related to the test (such as measurement point information, collection frequency, test conditions, etc.), so as to ensure the structured management of the data and the characteristics of easy traceability.

[0090] (1.4) Delay Identification and Delay Compensation between Data

[0091] Taking into account the existence of time delay relationships between variables in the process industry, this may lead to errors in correlation analysis and data mining. For this reason, in this embodiment, the attention mechanism is firstly used to identify the time delay relationship between different variables, and then delay compensation variable data is generated after delay compensation, so as to ensure the accuracy of subsequent feature extraction and variable screening.

[0092] (1.4.1) Generation of multi-dimensional time-delay data

[0093] All non-flow variables and flow variables are taken as candidate feature variables. Given the candidate feature variable X i (t), generate a set of delayed versions of variable data to simulate the impact of different time delays:

[0094]

[0095] Among them, τ j It is a set of preset time delay values, such as 1 second, 5 seconds, 10 seconds, etc., which is used to explore the characteristics of variables under different time delays.

[0096] (1.4.2) Delay weight allocation based on attention mechanism

[0097] The attention mechanism is used to automatically learn the weight of each delay dimension and determine the impact of each delay on the variable. The calculation of the attention mechanism is as follows:

[0098] Calculating latency data The attention weight vector α (τ) , which represents the importance of different delay dimensions. The vector can be expressed as:

[0099]

[0100] Among them, score(X i (t),X i (t-τ j )) is a scoring function used to quantify the correlation between the current time and different delay times.

[0101] (1.4.3) Feature representation after delay compensation

[0102] Use the attention weights to perform weighted summation on the data of different delay dimensions to obtain the delay compensation feature representation That is, the delay compensation variable data can be expressed as:

[0103]

[0104] Through this step, a feature representation is generated after considering delay compensation for subsequent variable screening and association analysis.

[0105] (1.5) Selection of flow-related variables

[0106] Based on the various variable data after delay compensation, artificial experience weights are introduced, and the characteristic variables most relevant to traffic Q are screened out in combination with data-driven methods, including:

[0107] (1.5.1) Manual experience weight setting

[0108] According to domain knowledge, each candidate variable is assigned an artificial experience weight w i , the weight range is [0,1]. The higher the empirical weight, the greater the potential impact of the variable on traffic prediction in the opinion of domain experts.

[0109] (1.5.2) Joint probability distribution

[0110] The Gaussian mixture model (GMM) is used to fit the delay-compensated traffic data Q and each non-traffic characteristic variable data. The joint probability distribution of From a statistical perspective, the correlation with traffic can be expressed as:

[0111]

[0112] Among them, π k is the weight of the kth Gaussian component, μ k and Σ k are the mean and covariance matrices respectively.

[0113] (1.5.3) Mutual information calculation

[0114] Based on the joint probability distribution obtained by the above calculation, the flow data Q and each non-flow characteristic variable data are calculated. The mutual information between them is used to quantify the nonlinear correlation between different non-flow variables and flow, which can be expressed as:

[0115]

[0116] (1.5.4) Comprehensive score calculation

[0117] Based on the mutual information, the features are sorted from high to low and the features with higher mutual information values ​​are selected, such as selecting the top k feature variables or selecting multiple feature variables with mutual information higher than a certain set value.

[0118] Furthermore, based on mutual information and combined with the preset artificial experience weight, the comprehensive score S of each feature variable is calculated. i , the formula is:

[0119]

[0120] Among them, w i is the artificial experience weight, is the mutual information of data-driven computation.

[0121] Based on the final score calculated above, select the score S i The higher the variables, the more relevant the variables are as the key feature variables, and the more relevant the variables are.

[0122] (1.5.5) Storage and subsequent application of delay compensation features

[0123] The extracted delay compensation feature variables and their corresponding attention weights are stored in the database, and the scoring function, weight vector and other parameters in the attention mechanism are recorded to facilitate the tracing and verification of the delay relationship in the future.

[0124] (2) Actual system testing and data-driven model building stage

[0125] (2.1) Experimental data collection

[0126] Experiments under various working conditions (including startup, steady state and transient) were carried out on the actual system A to collect non-flow process variable data such as temperature and pressure at each measuring point. During this process, the measuring point name and location information were fully recorded to ensure the comprehensiveness and consistency of the data.

[0127] (2.2) Data preprocessing and storage

[0128] After data collection, the non-flow variable data collected by the prototype system and the actual system during the test phase are cleaned and preprocessed, including erroneous data deletion, missing value processing, timestamp transformation, data filtering, data standardization, etc., to ensure data consistency and validity; the cleaned data are stored in time series and relational databases respectively to ensure that the data has efficient query and traceability, laying the foundation for subsequent modeling.

[0129] (2.3) Encoder-Decoder Model Training

[0130] (2.3.1) Encoder

[0131] like Figure 2 In the encoder-decoder training phase shown, the preprocessed data is input into the encoder, which extracts domain-independent feature representations from the input non-flow process variable data. Preferably, the encoder can adopt the following network structures: long short-term memory network LSTM, gated recurrent unit GRU, time domain convolution network TCN or Transformer, etc., to effectively capture the time dependency of the input features.

[0132] The encoder converts the input sequence X of other non-flow process variables into non-flow , including the data of the prototype system and the actual system, is converted into a feature vector z, namely:

[0133] z=Encoder(X non-flow );

[0134] (2.3.2) Decoder

[0135] The extracted feature vector is then input into the encoder to reconstruct the non-flow variable data, that is, the decoder uses the feature vector z generated by the encoder to reconstruct the non-flow process variable data, ensuring that the features learned by the encoder have strong expressive power. The decoder can use the same network structure as the encoder to decode the feature vector z, that is:

[0136]

[0137] Through training, the output of the decoder As close as possible to the input data X non-flow , thereby ensuring that the feature representation learned by the encoder is comprehensive and stable.

[0138] (2.3.3) Training process

[0139] The training loss function of Encoder-Decoder uses reconstruction error to construct the loss function through the error between input data and reconstructed data, and optimizes the model parameters by minimizing the difference between input data and decoded output. The loss function is:

[0140]

[0141] The above design ensures that the encoder can extract stable domain-independent feature vector z, which is convenient for subsequent migration and traffic prediction tasks.

[0142] In this embodiment, the feature extraction model uses an encoder-decoder structure to extract features of non-traffic variables for traffic prediction. As another implementation, the encoder-decoder structure may not be used, but features may be extracted through self-supervised learning. Specifically, the latent feature representation is learned on non-traffic variables through a self-supervised model and directly used as input to the traffic prediction model. This method eliminates the need for a decoder to perform reconstruction tasks, thereby reducing model complexity and computational costs, while still ensuring feature stability and domain independence.

[0143] (2.4) Traffic prediction model training (based on Encoder Vector)

[0144] like Figure 3 In the traffic prediction model training phase shown in FIG, after completing the training of the encoder-decoder model, the feature vector z generated by the prototype system data is used A As the input of the traffic prediction model, the real traffic data of the prototype system is used as the supervision signal (i.e., output data) to build and train the traffic prediction model. To this end, this embodiment selects a fully connected neural network (FCNN) as the architecture of the traffic prediction model. In this traffic prediction model:

[0145] Input: feature vector z generated by the encoder A ;

[0146] Output: Traffic prediction value

[0147] The loss function is: Based on the mean square error between the predicted flow value and the actual flow value of the prototype system, a loss function is constructed, and the model is optimized using the mean square error (MSE) loss. The loss function is:

[0148]

[0149] in, is the actual flow value, is the traffic prediction value.

[0150] The traffic prediction model after the above training has the ability to estimate the traffic from the encoded feature vector and can be migrated to the actual system to realize traffic estimation.

[0151] (2.5) Transfer learning and actual system traffic prediction

[0152] In the actual system B, since there is no flow sensor, only non-flow process variables are collected. Through transfer learning, the feature vector generated by the encoder is applied to estimate the flow of the actual system, including:

[0153] (2.5.1) The non-flow process variable input of the actual system B is: Collect the non-flow process variable data of the actual system B

[0154] (2.5.2) Generate the flow prediction model of the actual system: Use the encoder to convert the non-flow variable data of the actual system B into a feature vector:

[0155] (2.5.3) Traffic prediction: The generated feature vector z B Input into the traffic prediction model to obtain the traffic estimate of system B

[0156] Through the above design, sensorless estimation of the flow of system B is achieved.

[0157] (2.6) Model storage

[0158] During the model training and updating process, all model versions, parameters, and training losses are stored to ensure the traceability and efficient management of the model. The model is stored in a dedicated database, which supports the rapid loading and application of the model in the actual system.

[0159] (3) Actual system operation stage

[0160] (3.1) Online data collection and preprocessing

[0161] During the normal operation of the actual system, non-flow process variable data such as temperature and pressure are collected in real time. Furthermore, in order to ensure the stability and robustness of the model, the collected data is preprocessed, including error data deletion, missing value processing, timestamp transformation, data filtering, data standardization, etc., and the preprocessed data is stored in the database.

[0162] (3.2) Traffic prediction

[0163] like Figure 4 In the actual system operation stage shown in the figure, the Encoder-Decoder model and flow prediction model built and trained previously are used to predict the flow value of the actual system online to achieve flow estimation under sensorless conditions. That is, the pre-processed non-flow process variable data is used as input data and input into the trained feature extraction model and flow prediction model. The feature extraction model extracts the feature vector that is not related to the field. The feature vector is then passed through the flow prediction model to predict and output the flow value of the actual system online.

[0164] (3.3) Online adaptive learning trigger conditions

[0165] In the normal operation of the actual system, considering that the distribution of input data may drift with changes in the environment and working conditions, in order to ensure the adaptability and accuracy of the model, this embodiment also designs reasonable trigger conditions during the entire prediction process to determine when to start adaptive transfer learning to update the model.

[0166] In this embodiment, the input and output data of the real-time monitoring model are combined with the trigger conditions of self-transfer learning to make a judgment, wherein multiple trigger conditions are set, including that the distribution change of the input data is greater than the first set value, the flow prediction error is greater than the second set value, and the feature space similarity is less than the third set value, specifically:

[0167] (3.3.1) Data distribution change detection

[0168] Based on the distribution changes of the input data monitored in real time, determine whether there is an offset. Furthermore, the following statistical methods can be used to quantify the distribution changes of the input data, including the maximum mean difference between the input data in the current time window and the previous time window, or the mean and variance change values ​​of the input data in the current window and the previous window extracted based on the sliding window. Specifically:

[0169] Maximum Mean Difference (MMD): MMD measures the distribution difference between the current running data and the historical data. Assume that the data in the current time window is X current , the data of the previous time window is X prev , then MMD is defined as: δ MMD =MMD(X current ,X prev ); when δ MMD Exceeding the set threshold τ MMD (i.e., the first set value), triggering the adaptive transfer learning update;

[0170] Changes in the mean and variance of the sliding window: The mean and variance of the input features are monitored through the sliding window. When the mean and variance change exceeds the set threshold, the data distribution may change. Assume that the mean and variance of the sliding window are μ and σ 2 , when |μ current -μ prev |>τ μ or , it triggers a model update.

[0171] (3.3.2) Predictive performance monitoring

[0172] The model accuracy is judged based on the real-time monitored traffic prediction error. If the error gradually increases, it means that the model may no longer be suitable for the current operating data and needs to be adaptively updated. The traffic prediction error is: the mean square error of the pseudo-label prediction error at the previous moment and the current moment in steady state. The pseudo-label refers to the predicted traffic value output by the traffic prediction model.

[0173] Specifically, the mean square error (MSE) of the prediction error of the sliding window in steady state is: In the absence of real traffic labels, the model performance can be estimated by the prediction error generated by the pseudo-label. For example, the mean square error of the pseudo-label prediction error at the previous moment and the current moment is calculated. If the error gradually increases, it means that the accuracy of the model may decrease. The error can be expressed as:

[0174]

[0175] When MSE pseudo Exceeding the set threshold τ MSE (i.e. the second set value), the model update is triggered.

[0176] (3.3.3) Similarity measure in feature space

[0177] In the process of transfer learning, the feature space similarity is used to determine whether the current feature vector has changed significantly. If the change is obvious, an adaptive update is performed. The feature space similarity is the cosine similarity of the feature vectors generated in the current window and the previous window extracted based on the sliding window.

[0178] Specifically, the encoder output feature changes as follows: calculate the similarity (such as cosine similarity) of the feature vector z generated by the encoder in the sliding window. When the similarity is lower than the set threshold, it means that the feature space has changed. Assume that the feature vectors generated twice in the sliding window are z prev and z current , then the cosine similarity is defined as:

[0179] If CosSim<τ CosSim (i.e., the third set value), then the transfer learning update is triggered.

[0180] (3.3.4) Time trigger condition

[0181] In order to avoid frequent updates from burdening the system, the above trigger conditions also include a time trigger condition, that is, adding a time interval constraint. The time trigger condition has the highest priority, and the time trigger condition requires that the time interval for model updates is not less than the set time threshold. In this embodiment, the minimum update time interval T is set. min , even if the trigger condition is met, the time interval between updates must not be less than T min .

[0182] (3.3.5) Conditional combination for adaptive transfer learning

[0183] In order to improve the sensitivity and stability of the trigger conditions, a combination of multiple conditions is used:

[0184] The above multiple trigger conditions are set, and only when at least two conditions are met will the adaptive transfer learning be triggered to fine-tune the model parameters and update them, so as to avoid false triggering caused by fluctuations in a single condition. The joint condition can be expressed as: (δ MMD >τ MMD )∧(MSE pseudo >τ MSE )∨(CosSim<τ CosSim ).

[0185] (3.4) Update process of adaptive transfer learning

[0186] (3.4.1) Data collection and preparation

[0187] After the trigger condition is met, non-flow process variable data (such as temperature, pressure, etc.) for a recent period of time are collected from the actual system and used for the adaptively updated data set.

[0188] (3.4.2) Encoder feature fine-tuning

[0189] During the updating process, the encoder is first fine-tuned to adapt to the distribution changes of the latest data in the actual system. The feature fine-tuning of the encoder is mainly achieved through self-supervised reconstruction loss.

[0190] (3.4.2.1) Retrain the encoder

[0191] The encoder is retrained using the most recently acquired actual system data to minimize the encoding reconstruction error between the input data and the decoder reconstructed output, allowing the encoder to capture the characteristic changes under the current working conditions.

[0192] The encoding reconstruction loss is: using the most recently collected data As input, the reconstruction loss function of the encoder-decoder is defined as:

[0193]

[0194] By minimizing the loss function, the parameters of the encoder are updated so that the feature vector generated by the encoder is more in line with the latest data distribution of the actual system.

[0195] (3.4.2.2) Maintaining domain-independent features

[0196] When fine-tuning the encoder, add contrastive learning or feature alignment loss so that the new feature vector z new and the original eigenvector z original Maintain consistency in the high-level feature space. Specifically, it can be achieved using the maximum mean difference (MMD) or contrast loss to ensure the domain independence of the feature vector. The feature alignment loss can be expressed as:

[0197] L alignment =‖E[z original ]-E[z new ]‖ 2 ;

[0198] The final fine-tuning loss is a combination of the two:

[0199] L encoder =λ1L reconstruct +λ2L alignment .

[0200] In this embodiment, in order to maintain the domain independence of the features, contrastive learning or feature alignment loss is used to maintain the stability of the features extracted by the encoder. As another implementation method, multi-task learning can be used instead of feature alignment loss, and flow prediction and other non-flow process variable predictions can be completed simultaneously during training, so that the features extracted by the encoder can take into account multiple tasks, thereby improving its domain independence and cross-system applicability. Through this method, the encoder can extract robust feature representations for flow prediction in various systems without feature alignment loss.

[0201] (3.4.3) Pseudo-label fine-tuning of traffic prediction model

[0202] Since the actual system does not have a flow sensor and cannot directly obtain the real label of the flow, the feature vector generated by the encoder and the pseudo label generation method are used to fine-tune the flow prediction model. It is noted that when the actual system is in a steady state, the fluctuation of the flow is small, so the model can be trained to make the predicted flow fluctuation smaller.

[0203] (3.4.3.1) Pseudo-label generation

[0204] In the absence of true labels, the existing traffic prediction model is used to generate pseudo labels. Assume that the current feature vector is z new , then the traffic pseudo-label generated by the traffic prediction model is:

[0205]

[0206] (3.4.3.2) Self-supervised training using pseudo-labels

[0207] The generated pseudo-label is used as the training target of the traffic prediction model, the pseudo-label loss function is defined, and the pseudo-label data is used to perform self-supervision updates on the model parameters, so that the traffic prediction model gradually adapts to the new input features. Among them, the constructed pseudo-label loss function is:

[0208]

[0209] in, is the traffic forecast value at the current time step, is the pseudo label of the previous time step.

[0210] (3.4.3.3) In addition, it is noted that when in steady state, a more accurate flow value can be inferred from other states, so the prediction accuracy at this time is also taken as one of the optimization goals. The prediction loss is:

[0211]

[0212] in, is the actual flow value, is the predicted value.

[0213] (3.4.4) Joint optimization objectives

[0214] To ensure the robustness of the update, the overall loss function is obtained by weighted summing the encoder’s reconstruction loss, feature alignment loss, pseudo label loss, and prediction loss, which can be expressed as:

[0215] L total =λ1L reconstruct +λ2L alignment +λ3L pseudo +λ4L predict :

[0216] By minimizing L total , the encoder and traffic prediction model are jointly optimized to ensure the adaptability and prediction accuracy of the model on the latest data.

[0217] (3.4.5) Delay adjustment of attention mechanism

[0218] If the encoder-decoder structure includes an attention mechanism, the attention weights can be dynamically adjusted during the update process to adapt to data changes with different delays.

[0219] Specifically, the delay adjustment is: according to the new input features, the delay weight in the attention mechanism is recalculated, so that the model pays more attention to the current relevant delay features during decoding. The calculation formula of the attention weight is:

[0220]

[0221] Using this dynamic weight, the decoder can accurately capture the temporal dependencies between different inputs.

[0222] (3.4.6) Updated version control and performance monitoring

[0223] After each update, the model version, corresponding parameters, and current data distribution information are stored in the database for easy tracing in the future. The prediction performance of the model on the actual system is monitored by the mean square error (MSE) of the sliding window. When the error continues to exceed the threshold, the adaptive update can be triggered again.

[0224] In this embodiment, in order to achieve adaptive updating, a fixed-size sliding window is used to monitor the distribution changes and prediction errors of the input data. As another implementation, a dynamic window adjustment mechanism can also be used to automatically adjust the sliding window size according to the rate of data change. For example, when the data changes quickly, the window becomes smaller to increase the update frequency; when the data is relatively stable, the window becomes larger to reduce the computational burden. Through this dynamic window adjustment mechanism, not only can it adapt to different environmental changes, but it also further saves resources, thereby ensuring the real-time and high efficiency of the system.

[0225] Embodiment 2

[0226] This embodiment provides a sensorless flow estimation system, including:

[0227] Data acquisition and preprocessing module, used to collect non-flow process variable data in real time during the normal operation of the actual reactor cooling system, and perform data preprocessing;

[0228] The flow prediction module is used to input the preprocessed non-flow process variable data into the trained feature extraction model and flow prediction model, extract the feature vector that is not related to the field through the feature extraction model, and then pass the feature vector through the flow prediction model to predict and output the flow value of the actual system online;

[0229] In the prediction process, the input and output data of the model are monitored in real time. When the monitoring data meets the trigger conditions of adaptive transfer learning, the non-flow process variable data of the actual system collected recently are used to fine-tune the model parameters of the feature extraction model and the flow prediction model based on encoding reconstruction loss, feature alignment loss, pseudo-label loss, and prediction loss. The model is updated and the updated model is used to perform online prediction of the flow value.

[0230] In some embodiments, the sensorless flow estimation system can be implemented in software, for example, in the form of programs and plug-ins, including the following software modules: data acquisition and preprocessing module, flow prediction module, etc. These modules are logical, so they can be arbitrarily combined or further split according to the functions implemented. For example, the flow prediction module can be further split into: a feature extraction module, which is used to extract feature vectors that are independent of the domain; a model update module, which is used to determine whether the monitoring data meets the trigger conditions for adaptive transfer learning. If so, the non-flow process variable data of the actual system collected recently is used to fine-tune the model parameters of the feature extraction model and the flow prediction model based on the encoding reconstruction loss, feature alignment loss, pseudo-label loss, and prediction loss to update the model; and other modules, which can all be implemented by running computer programs to achieve more accurate sensorless flow estimation implemented in this embodiment.

[0231] Furthermore, the computer program can be a native program or software module in the operating system; it can be a native application (APP, Application), that is, a program that needs to be installed in the operating system to run, such as the smart shooting APP; it can also be a small program, that is, a program that can be run only by downloading it to the browser environment; it can also be a small program that can be embedded in any APP. In short, the above-mentioned computer program can be any form of application, module or plug-in.

[0232] As another implementation, the sensorless flow estimation system provided in the embodiment of the present application can also be implemented in hardware. As an example, a processor in the form of a hardware decoding processor can be used, which is programmed to execute the sensorless flow estimation method provided in the embodiment of the present application. For example, the processor in the form of a hardware decoding processor can use one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs) or other electronic components.

[0233] The steps involved in the above embodiment 2 correspond to those in the method embodiment 1. For the specific implementation method, please refer to the relevant description part of the embodiment 1, which will not be repeated here.

[0234] Embodiment 3

[0235] This embodiment provides an electronic device, including a memory and a processor, wherein the memory is used to store executable instructions; the processor is used to implement the sensorless flow estimation method provided in the first embodiment when executing the executable instructions stored in the memory.

[0236] Furthermore, as an implementation method of an electronic device, a computing device has a processor that reads the computer executable instructions from a computer readable storage medium, and the processor executes the computer executable instructions, so that the computing device executes the sensorless flow estimation method proposed in the first embodiment.

[0237] Embodiment 4

[0238] This embodiment provides a computer-readable storage medium storing executable instructions for causing a processor to execute instructions to complete the sensorless flow estimation method provided in the first embodiment and to complete the steps in the sensorless flow estimation method described above.

[0239] The steps involved in the above embodiments 2 to 4 correspond to the method embodiment 1, and the specific implementation methods can refer to the relevant description part of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0240] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EP ROM, EEPROM, flash memory, magnetic surface storage, optical disk, or CD-ROM; or it may be various devices including one or any combination of the above memories.

[0241] In some embodiments, executable instructions may be in the form of a program, software, software module, script or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a stand-alone program or as a module, component, subroutine or other unit suitable for use in a computing environment.

[0242] As an example, executable instructions may, but need not, correspond to a file in a file system, may be stored as part of a file that stores other programs or data, such as in one or more scripts in a HyperText Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple coordinated files (e.g., files storing one or more modules, subroutines, or code portions).

[0243] By way of example, executable instructions may be deployed to be executed on one computing device, or on multiple computing devices located at one site, or on multiple computing devices distributed across multiple sites and interconnected by a communication network.

[0244] Further, those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, and optionally, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0245] The above description is only a preferred embodiment of the present invention. Although the specific implementation mode of the present invention is described in conjunction with the accompanying drawings, it is not a limitation of the protection scope of the present invention. Those skilled in the art should understand that on the basis of the technical solution of the present invention, various modifications or variations that can be made by those skilled in the art without creative work are still within the protection scope of the present invention.

Claims

1. A sensorless flow estimation method, characterized in that: include: Real-time acquisition of non-flow process variable data during the normal operation of the actual reactor cooling system and data preprocessing; The pre-processed non-flow process variable data is used as input data and input into the trained feature extraction model and flow prediction model. The feature extraction model extracts the feature vector that is not related to the field. The feature vector is then passed through the flow prediction model to predict and output the flow value of the actual system online. In the prediction process, the input and output data of the model are monitored in real time. When the monitoring data meets the trigger conditions of adaptive transfer learning, the non-flow process variable data of the actual system collected recently are used to fine-tune the model parameters of the feature extraction model and the flow prediction model based on encoding reconstruction loss, feature alignment loss, pseudo-label loss, and prediction loss. The model is updated and the updated model is used to perform online prediction of the flow value.

2. A sensorless flow estimation method as claimed in claim 1, characterized in that: Setting multiple adaptive transfer learning trigger conditions, including that the distribution change of input data is greater than a first set value, the flow prediction error is greater than a second set value, and the feature space similarity is less than a third set value; Based on the monitoring data, when at least two trigger conditions are met, adaptive transfer learning is triggered to perform fine-tuning training and update of model parameters.

3. A sensorless flow estimation method as claimed in claim 2, characterized in that: Determine whether there is an offset based on the distribution change of the input data monitored in real time, where the distribution change of the input data is: the maximum mean difference between the input data in the current time window and the previous time window, or the mean and variance change values ​​of the input data in the current window and the previous window extracted based on the sliding window; The model accuracy is determined based on the traffic prediction error monitored in real time, where the traffic prediction error is: the mean square error of the pseudo-label prediction errors at the previous moment and the current moment in steady state; the pseudo-label refers to the predicted traffic value output by the traffic prediction model; Whether the current feature vector has changed is determined based on the feature space similarity, wherein the feature space similarity is the cosine similarity of the feature vectors generated in the current window and the previous window extracted based on the sliding window.

4. A sensorless flow estimation method as claimed in claim 2, characterized in that: The trigger condition also includes a time trigger condition, which has the highest priority. The time trigger condition is that the time interval for model update is not less than a set time threshold.

5. A sensorless flow estimation method as claimed in claim 1, characterized in that: The data preprocessing includes: erroneous data deletion, missing value filling, timestamp transformation and data filtering.

6. A sensorless flow estimation method as claimed in claim 1, characterized in that: The non-flow process variables use key characteristic variables for flow prediction, and the key characteristic variables include pressure and temperature. The screening of the key characteristic variables includes: Install flow measurement sensors and various types of sensors in the prototype system to collect flow variable data and non-flow variable data at each measuring point under various working condition experiments; the non-flow variables include temperature and pressure; Preprocess the collected data of different variables, use the attention mechanism to identify the delay relationship between different variables, and generate delay compensation variable data; Based on the delay-compensated data, the Gaussian mixture model is used to fit the joint probability distribution of traffic and each non-traffic variable, so as to calculate the mutual information between traffic and each non-traffic variable; Based on mutual information and the preset artificial experience weights, the comprehensive score of each variable is calculated, and the variables most relevant to traffic are selected as key feature variables according to the comprehensive score.

7. A sensorless flow estimation method as claimed in claim 1, characterized in that: The feature extraction model adopts an encoder-decoder structure, and the traffic prediction model adopts a fully connected neural network architecture; The training process of the feature extraction model and the traffic prediction model includes: Collect non-flow variable data of each measuring point of the actual system under various working condition experiments; the non-flow variables include temperature and pressure; Preprocess the non-flow variable data collected by the prototype system and the actual system during the test phase, input the preprocessed data into the encoder to extract feature vectors, and then input the feature vectors into the encoder to reconstruct the non-flow variable data; Train the encoder-decoder model by constructing a loss function based on the error between the input data and the reconstructed data; The trained encoder is used to extract the feature vector of the non-flow variable data of the prototype system, and the feature vector is input into the flow prediction model to output the predicted flow value; A loss function is constructed based on the mean square error between the predicted flow value and the actual flow value of the prototype system to train the flow prediction model.

8. A sensorless flow estimation system, characterized in that: include: Data acquisition and preprocessing module, used to collect non-flow process variable data in real time during the normal operation of the actual reactor cooling system, and perform data preprocessing; The flow prediction module is used to input the pre-processed non-flow process variable data into the trained feature extraction model and flow prediction model, extract the feature vector that is not related to the field through the feature extraction model, and then pass the feature vector through the flow prediction model to predict and output the flow value of the actual system online; In the prediction process, the input and output data of the model are monitored in real time. When the monitoring data meets the trigger conditions of adaptive transfer learning, the non-flow process variable data of the actual system collected recently are used to fine-tune the model parameters of the feature extraction model and the flow prediction model based on encoding reconstruction loss, feature alignment loss, pseudo-label loss, and prediction loss. The model is updated and the updated model is used to perform online prediction of the flow value.

9. An electronic device, characterized in that: include: A memory for storing executable instructions; The processor is used to implement a sensorless flow estimation method as described in any one of claims 1 to 7 when executing the executable instructions stored in the memory.

10. A computer-readable storage medium, characterized in that: Executable instructions are stored, which are used to cause the processor to execute and complete a sensorless flow estimation method as described in any one of claims 1-7.

Images