A hot press tank temperature offline simulation and online prediction method and system based on hybrid KAN feature fusion

Abstract

The application discloses a kind of based on HybridKAN feature fusion autoclave temperature offline simulation and online prediction method, comprising: collecting different batches of historical temperature time series data and static process parameter data;Carrying out feature screening and pretreatment, construct time series data input sequence and static process parameter input sequence;HybridKAN neural network model is designed, with time series data input sequence and static process parameter input sequence as input, after feature extraction and fusion, output the predicted value of next time autoclave temperature;Minimize loss function as target, train model and obtain optimal model using structured pruning strategy;Deploy optimal model, through rolling time window iteration multi-step prediction, realize the offline simulation and online prediction of autoclave temperature.This method is high in accuracy, wide in universality, while model parameter scale is smaller, reasoning efficiency is high, suitable for complex scene autoclave temperature prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of time series data prediction, in particular to a hot press tank temperature offline simulation and online prediction method and system based on HybridKAN feature fusion. BACKGROUND

[0002] Temperature simulation and prediction of composite curing molding process is an important means to ensure product quality, and the accuracy of temperature prediction directly affects product qualification rate and production capacity. The temperature simulation based on traditional CAE method has high prediction accuracy and can simulate the entire temperature field distribution, but it has problems of difficulty in multi-field coupling modeling, large amount of calculation for solving, and low simulation efficiency. At the same time, since it is offline modeling and simulation, it cannot fuse the measured data to improve the prediction accuracy.

[0003] The data-driven method such as ARIMA and LSTM algorithm autoregressively models the thermocouple temperature time series data, and has relatively high prediction accuracy through iterative prediction. Since it does not fuse static preset static process parameters (such as tooling size, thermocouple position, etc.), it has insufficient generalization ability on multiple parts. The data-driven method such as random forest method constructs multiple decision trees for ensemble learning, and has certain processing ability for static features, but it is not as good as LSTM in capturing the dynamic changes of temperature sequence, and the processing efficiency of high-dimensional data is low. In the face of a large number of static process parameters and temperature data of the hot press tank, both the calculation efficiency and the prediction accuracy need to be improved. Static process parameters reflect the inherent properties of the process system (such as tooling size), and dynamic time series data reflect the dynamic evolution of temperature (such as heating rate). Simple fusion of the two types of features in the stage is easy to lose information, resulting in prediction deviation. Therefore, there is an urgent need for a hot press tank temperature prediction method that can cope with the complex prediction scenario of high coupling between static process parameters and time series data. SUMMARY

[0004] In view of the problems in the prior art, the present application proposes a hot press tank temperature offline simulation and online prediction method and system based on HybridKAN feature fusion. Based on the HybridKAN neural network architecture, the coupling features of static process parameters and time series data are deeply mined, aiming to improve the accuracy of hot press tank temperature prediction and generalization on multiple parts. And relying on the characteristics of small model parameter scale and high reasoning efficiency of the HybridKAN model, temperature offline simulation and real-time prediction can be realized on resource-constrained devices.

[0005] To achieve the above technical purpose, the present application provides the following technical scheme, a hot press tank temperature offline simulation and online prediction method based on HybridKAN feature fusion, which specifically includes:

[0006] Collecting part curing forming data of different batches from the autoclave historical database, the part curing forming data including historical temperature time series data and static process parameter data;

[0007] Feature screening and preprocessing are performed on the part curing forming data, and then the historical temperature time series data and the static process parameter data after feature screening and preprocessing are used to construct time series data input sequences and static process parameter input sequences respectively;

[0008] A HybridKAN neural network model is designed, taking the time series data input sequences and the static process parameter input sequences as model inputs, to extract dynamic time series features and static process features respectively; the two types of features are fused and connected with the autoclave temperature at the current time to obtain the output features of the HybridKAN neural network model; the output features are used to predict the autoclave temperature at the next time;

[0009] The training hyperparameters of the HybridKAN neural network model are set, and the HybridKAN neural network model is trained with the objective of minimizing the loss function, and a structured pruning strategy is used to lightweight the model; the optimal HybridKAN neural network model is obtained after training and pruning;

[0010] The optimal HybridKAN neural network model is deployed to realize offline simulation and online prediction of the autoclave temperature through rolling time window iteration multi-step prediction.

[0011] Further, the historical temperature time series data and static process parameter data specifically include:

[0012] The historical temperature time series data includes: each thermocouple ID, historical time index, and temperature value of each thermocouple at each historical time;

[0013] The static process parameter data includes: tool thermal capacity, part thermal capacity, tool length / width / height, part length / width / height, distance between thermocouple and tool front end, distance between tool and symmetry plane, distance between tool and door, distance between tool and tank wall, and part heat exchange area.

[0014] Further, the feature screening and preprocessing of the part curing forming data, and then the historical temperature time series data and the static process parameter data after feature screening and preprocessing, respectively constructing time series data input sequences and static process parameter input sequences specifically include:

[0015] The static process parameter data is preliminarily screened through heat conduction mechanism, and based on the heat transfer principle in the autoclave, parameters strongly related to heat distribution are preferentially retained, and secondary parameters are removed;

[0016] Finally, the static process parameter data is screened by a data-driven method to construct a random forest model containing full static features, and according to the parameter importance ranking result, the N-dimensional static process parameter input sequence is obtained by selecting the N static process parameters with high importance ranking;

[0017] Data cleaning and exception processing are performed, and the abnormal points in the historical temperature time series data of different batches are removed by the 3σ principle, and then the missing values are filled by linear interpolation;

[0018] The historical temperature time series data after data cleaning and exception processing is rolled with a window size L, and the temperature values of L historical time points and the index i of the last historical time point are combined into a sequence to obtain a L+1-dimensional time series data input sequence; and the negative time historical temperature data before the initial zero time of the sequence is filled as the initial environmental temperature of the autoclave;

[0019] The static process parameter input sequence and the time series data input sequence are used as the input of the subsequent network.

[0020] Further, the HybridKAN neural network model is constructed based on the KAN network, including a time series data processing module, a static process parameter processing module, a feature fusion module and a linear output layer, wherein:

[0021] The time series data processing module and the static process parameter processing module perform feature extraction in parallel; the time series data processing module takes the time series data input sequence as input, passes through three layers of KAN network, and outputs dynamic time series features; the static process parameter processing module takes the static process parameter input sequence as input, also passes through three layers of KAN network, and outputs static process features; the dynamic time series features and the static process features have the same feature dimension;

[0022] Then, the dynamic time series features and the static process features are first spliced and fused in the feature fusion module to obtain fused features, and the fused features are then input into a four-layer KAN network; the dynamic time series features and the static process features are deeply coupled to obtain deep coupling features;

[0023] The deep coupling features output by the feature fusion module are connected to the current time temperature value in the time series data input sequence through a residual network jump, and the next time temperature value prediction result is obtained through a linear output layer.

[0024] Further, the loss function adopts a mean square error loss MSE.

[0025] Further, the structured pruning strategy is adopted to perform lightweight processing on the model, and the structured pruning strategy is specifically:

[0026] The neuron importance is determined based on weight absolute value ranking, and low importance neuron connections are removed according to a preset pruning rate; the pruned network is fine-tuned, and a learning rate decay strategy is adopted to restore the model accuracy.

[0027] The optimal HybridKAN neural network model is deployed to realize offline simulation and online prediction of the autoclave temperature through rolling time window iterative multi-step prediction, and the offline simulation and online prediction are as follows:

[0028] The optimal HybridKAN neural network model is deployed to the computing device connected with the database for offline simulation and online prediction;

[0029] The offline simulation is that, in the case that there is no complete measured temperature data, the static process parameters and the tank temperature are taken as initial inputs of the optimal HybridKAN neural network model, the tank temperature at each time is recursively calculated, the temperature evolution of the curing forming process of the parts under the current process scheme is simulated, and the evolution result is presented in the form of a temperature curve graph;

[0030] The online prediction is that, in the case that the computing device can obtain the measured temperature data collected by the thermocouple from the database, the historical temperature time series data, the static process parameter data and the current time index starting from the current time are taken as inputs of the optimal HybridKAN neural network model, the model outputs the temperature prediction value at the next time, and through rolling time window iterative multi-step prediction, a whole-process temperature process curve evolving with time is generated to reversely guide the adjustment of the heating rate and the holding time of the autoclave.

[0031] The application further discloses an application system of the above-mentioned autoclave temperature offline simulation and online prediction method based on HybridKAN feature fusion, and specifically comprises:

[0032] A data acquisition unit is used for curing forming data of different batches of parts;

[0033] A data processing unit is used for feature screening and preprocessing of the curing forming data of the parts;

[0034] A model construction unit is used for constructing the HybridKAN neural network model;

[0035] A training optimization unit is used for training and pruning optimization of the HybridKAN neural network model;

[0036] A prediction application unit is used for deploying the model and performing offline simulation and online prediction.

[0037] In addition, the application further provides an electronic device comprising a memory and a processor, wherein:

[0038] Memory is used to store computer programs that can run on a processor;

[0039] The processor is configured to execute, while running the computer program, a method for offline simulation and online prediction of autoclave temperature based on HybridKAN feature fusion as described above.

[0040] A computer-readable storage medium is also disclosed, which stores computer instructions for causing a processor to execute an offline simulation and online prediction method for autoclave temperature based on HybridKAN feature fusion as described above.

[0041] Based on the above technical solution, the present invention has at least the following beneficial effects:

[0042] The method proposed in this invention achieves high accuracy in predicting autoclave temperature while having a small number of model parameters, high inference efficiency, and good adaptability and generalization on various composite material components.

[0043] The HybridKAN neural network architecture and input data format are compatible with both offline and online modes, supporting both offline temperature simulation and online prediction, facilitating flexible deployment and switching in practical engineering systems. Offline simulation can predict the heating curves of specific process parameters in advance before production scheduling, providing support for process and scheduling optimization; online prediction can guide the adjustment of parameters such as the heating rate and holding time of the autoclave, avoiding the impact of uneven temperature distribution on product quality; and it enables autoclave temperature prediction in complex prediction scenarios where static process parameters and time-series data are highly coupled. Attached Figure Description

[0044] Figure 1 This is a flowchart of an offline simulation and online prediction method for autoclave temperature based on HybridKAN feature fusion proposed in this invention;

[0045] Figure 2 This is a framework diagram of the HybridKAN network autoclave temperature prediction method of the present invention.

[0046] Figure 3 This is a diagram of the HybridKAN network architecture of the present invention;

[0047] Figure 4 This is a graph showing the loss function during the training process of the HybridKAN network in this invention.

[0048] Figure 5 A comparison curve of offline simulation and actual temperature measurement of a batch of parts in an autoclave;

[0049] Figure 6 This is a graph showing the online temperature prediction for the autoclave. Detailed Implementation

[0050] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the following description is provided in conjunction with the accompanying drawings. Figures 1-6 The present invention will be further described in detail with reference to the embodiments. This will allow for a full understanding of how the present application uses technical means to solve technical problems and achieve technical effects, and to facilitate its implementation.

[0051] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0052] To improve the accuracy of temperature prediction in composite material manufacturing and its generalizability across multiple parts, in this embodiment, such as Figure 1 As shown, this invention proposes an offline simulation and online prediction method for autoclave temperature based on HybridKAN feature fusion, which specifically includes the following steps:

[0053] The curing and molding data of parts from different batches (at least 30 batches in this embodiment) are collected from the autoclave historical database. The curing and molding data of parts includes historical temperature time series data and static process parameter data.

[0054] In a preferred embodiment, the historical temperature time series data and static process parameter data specifically include:

[0055] The historical temperature time series data includes: each thermocouple ID, historical time index, and the temperature value of each thermocouple at each historical time.

[0056] The static process parameter data includes more than twenty items, including: tooling heat capacity, part heat capacity, tooling length / width / height, part length / width / height, distance of thermocouple from tooling front end, distance of tooling from symmetry plane, distance of tooling from tank door, distance of tooling from tank wall, and part heat exchange area.

[0057] In this embodiment, historical temperature time series data and static process parameters are collected for each thermocouple in the autoclave. Subsequently, the HybridKAN neural network model will be used to predict the temperature value of each thermocouple, and the temperature of the entire autoclave will be reflected based on the local temperature reflected by the temperature value of each thermocouple.

[0058] Feature filtering and preprocessing are performed on the solidification and molding data of the parts. Then, the time-series data input sequence and the static process parameter input sequence are constructed respectively using the historical temperature time-series data and static process parameter data after feature filtering and preprocessing.

[0059] In a preferred embodiment, the step of performing feature filtering and preprocessing on the curing data of the parts, and then constructing a time-series data input sequence and a static process parameter input sequence using the historical temperature time-series data and static process parameter data after feature filtering and preprocessing, specifically includes:

[0060] First, the static process parameter data are initially screened through the heat conduction mechanism. Then, based on the heat transfer principle in the autoclave, parameters that are strongly correlated with heat distribution are retained first, while secondary parameters are eliminated.

[0061] Then, the static process parameter data is finally filtered using a data-driven method to construct a random forest model containing all static features. Based on the parameter importance ranking results, the top N static process parameters are selected to obtain an N-dimensional static process parameter input sequence.

[0062] In this embodiment, six static process parameters were ultimately selected: the distance between the thermocouple and the front end of the tooling, the width of the tooling, the distance between the tooling and the plane of symmetry, the length of the tooling, the distance between the thermocouple and the door of the autoclave, and the distance between the thermocouple and the wall of the autoclave, to obtain a six-dimensional static process parameter input sequence. , represented as:

[0063] ;

[0064] in, This indicates the six selected static process parameters;

[0065] Data cleaning and anomaly handling were carried out. Outliers were removed from different batches of historical temperature time series data using the 3σ principle, and then linear interpolation was used to fill in the missing values.

[0066] The historical temperature time series data, after data cleaning and anomaly processing, is processed by a sliding window of size L. The temperature values ​​of the L historical moments and the index i of the last historical moment (i.e. the current moment) are combined into a sequence to obtain an L+1 dimensional time series data input sequence. The negative historical temperature data before the initial zero moment of the sequence are filled with the initial ambient temperature of the autoclave.

[0067] In this embodiment, temperature values ​​from 10 historical moments and the index of the current moment i are selected to obtain an 11-dimensional time-series data input sequence. , represented as

[0068] ;

[0069] in, The temperature values ​​are from 10 historical moments (the initial zero moment in this sequence is...). time, The time before a given time is considered a negative time.

[0070] Static process parameter input sequence and time series data input sequence Used as input for subsequent networks.

[0071] Design a HybridKAN neural network model, such as Figure 2 As shown, using time-series data input sequences and static process parameter input sequences as model inputs, dynamic time-series features and static process features are extracted respectively; the two types of features are fused and connected with the current autoclave temperature jump to obtain the output features of the HybridKAN neural network model; the output features are used to predict the autoclave temperature at the next moment.

[0072] In a preferred embodiment, the HybridKAN neural network model is constructed based on the KAN network, which is based on the Kolmogorov-Arnold network. For a multivariate input sequence... The KAN network represents it as the following multivariate function:

[0073] ;

[0074] in, It is a single-variable function. and The parameters of a univariate function For the number of univariate functions, For single-variable function indexing, Given the length of the input sequence, The index is used as the element index in the sequence; in this embodiment, the KAN network takes the sequence as input from static process parameters. and time series data input sequence In this process, the complex nonlinear relationships between tooling dimensions, distances, and other process parameters are captured.

[0075] The HybridKAN neural network model built based on the KAN network is as follows: Figure 3 As shown, it includes a time-series data processing module, a static process parameter processing module, a feature fusion module, and a linear output layer, wherein:

[0076] The time-series data processing module and the static process parameter processing module perform feature extraction in parallel. The time-series data processing module takes an 11-dimensional time-series data input sequence as input, passes it through a three-layer KAN network, and outputs 16-dimensional dynamic time-series features, capturing nonlinear relationships such as time dependence and trend changes in temperature time-series data.

[0077] The static process parameter processing module takes a 6-dimensional static process parameter input sequence as input, and also passes through a three-layer KAN network to output a 16-dimensional static process feature (although the time series data input sequence and the static process parameter input sequence have different dimensions, they can be upgraded by setting different univariate functions in each KAN layer, and the last layer uses a univariate function of the same dimension to reduce the dimension to 16). This captures the complex nonlinear spatial relationship between process parameters (such as the relationship between tooling size and heat distribution).

[0078] It should also be noted that in the three-layer KAN network of the time series data processing module and the static process parameter processing module of this application, the first and second layers of KAN are used to process the original input features and extract basic local features (such as size-position correlation in static parameters and adjacent temperature dependence in time series data). The second and third layers of KAN are responsible for combining and transforming the features of the current layer, capturing more complex correlations and generating the final abstract features. Finally, the third layer of KAN integrates the local / middle-layer features of the first two layers into high-dimensional, globally representative features, providing input for the subsequent fusion module.

[0079] Dynamic time-series features and static process features are first concatenated and fused in the feature fusion module to obtain 32-dimensional fused features. The fused features are then input into a four-layer KAN network (the input dimension of each KAN layer is determined by the output of the previous layer; overall, the dimensionality is first increased and then reduced to fit the feature dimension of the linear output layer), deeply coupling the dynamic time-series features and static process features to obtain deeply coupled features. These deeply coupled features better reflect the combined effect of process parameters and time-series temperature on autoclave temperature prediction.

[0080] Additionally, it should be noted that in the four-layer KAN network of the feature fusion module: the first and second layers of KAN are used to capture the interaction between static process parameter features and dynamic time series features; the second and third layers of KAN are used to generate higher-order comprehensive features to abstractly describe the global correlation; and finally, the fourth layer of KAN outputs the feature dimension adapted to the linear output layer to prepare for the final temperature prediction.

[0081] The deeply coupled features output by the feature fusion module are connected to the time series data input sequence through a residual network to obtain the current temperature value at the next time step. The result is then passed through a linear output layer to obtain the predicted temperature value at the next time step.

[0082] The training hyperparameters of the HybridKAN neural network model are set to minimize the loss function. In this embodiment, the loss function is the mean squared error loss (MSE), which is expressed as follows:

[0083] ;

[0084] Where n is the number of samples, , These are the actual value and the network's predicted value, respectively.

[0085] The optimizer was chosen as Adam (with a learning rate of 0.001 and a decay rate of 0.9). During training, early stopping was used to prevent overfitting, ensuring the model's MSE value on the input dataset was close to 0, thus completing the initial training. The specific loss function curve during training is shown below. Figure 4 As shown;

[0086] After initial training, a structured pruning strategy is used to lightweight the model. As a preferred implementation, in this embodiment, the lightweighting of the model using the structured pruning strategy specifically involves:

[0087] Neuron importance is determined by ranking based on the absolute value of weights, and low-importance neuron connections are removed according to a preset pruning rate. The pruned network is then fine-tuned and trained, and a learning rate decay strategy is used to restore model accuracy.

[0088] The optimal HybridKAN neural network model was obtained through training and pruning. Compared with the initial model, the optimal model obtained in this embodiment has 35% fewer parameters, an inference speed that is about 1.2 times faster, and a cross-tank batch test accuracy of over 95%.

[0089] Deploy the optimal HybridKAN neural network model and use rolling time windows for iterative multi-step prediction to achieve offline simulation and online prediction of autoclave temperature;

[0090] In a preferred embodiment, the deployment of the optimal HybridKAN neural network model, through iterative multi-step prediction via a rolling time window, to achieve offline simulation and online prediction of the autoclave temperature specifically involves:

[0091] The optimal HybridKAN neural network model is deployed to a computing device that interfaces with a database for offline simulation and online prediction.

[0092] The offline simulation involves using static process parameters and tank temperature as initial inputs to the optimal HybridKAN neural network model in the absence of complete measured temperature data. The tank temperature is then recursively calculated at each moment to simulate the temperature evolution of the part during the curing process under the current process scheme. The evolution results are presented as temperature curves, such as... Figure 5 As shown in the figure, the offline simulation and actual measurement curves of a batch of parts are compared. It can be seen that the offline simulation and actual measurement results of this method are extremely consistent.

[0093] The online prediction is as follows: When the computing device can obtain measured temperature data collected by thermocouples from the database, historical temperature time-series data from the current moment backwards, static process parameter data, and the current moment index are used as inputs to the optimal HybridKAN neural network model. The model outputs the predicted temperature value for the next moment, and through iterative multi-step prediction via a rolling time window, a full-process temperature process curve evolving over time is generated. The online predicted temperature curve is shown below. Figure 6 As shown, the online prediction results can guide the adjustment of parameters such as the heating rate and holding time of the autoclave.

[0094] This concludes the description of the method proposed in this invention. This embodiment also describes the application system of the above method, which specifically includes:

[0095] The data acquisition unit is used to solidify molding data of different batches of parts;

[0096] The data processing unit is used to perform feature filtering and preprocessing on the solidified molding data of the parts;

[0097] The model building unit is used to build the HybridKAN neural network model.

[0098] The training optimization unit is used to train and prune the HybridKAN neural network model.

[0099] The prediction application unit is used to deploy models and perform offline simulations and online predictions.

[0100] In addition, an electronic device is disclosed, which includes a memory and a processor, wherein:

[0101] Memory is used to store computer programs that can run on a processor;

[0102] The processor is configured to execute, while running the computer program, a method for offline simulation and online prediction of autoclave temperature based on HybridKAN feature fusion as described above.

[0103] A computer-readable storage medium is also disclosed, which stores computer instructions for causing a processor to execute an offline simulation and online prediction method for autoclave temperature based on HybridKAN feature fusion as described above.

[0104] In summary, the method proposed in this invention achieves high accuracy in predicting autoclave temperature while having a small number of model parameters, high inference efficiency, and good adaptability and generalization on various composite material components; it realizes autoclave temperature prediction in complex prediction scenarios where static process parameters and time-series data are highly coupled.

[0105] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0106] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. Collect curing and molding data of different batches of parts from the autoclave historical database, wherein the curing and molding data of parts includes historical temperature time series data and static process parameter data; The historical temperature time-series data and static process parameter data specifically include: The historical temperature time series data includes: each thermocouple ID, historical time index, and the temperature value of each thermocouple at each historical time. The static process parameter data includes: tooling heat capacity, part heat capacity, tooling length / width / height, part length / width / height, thermocouple distance from tooling front end, tooling distance from symmetry plane, tooling distance from tank door, tooling distance from tank wall, and part heat exchange area. Feature filtering and preprocessing are performed on the solidification and molding data of the parts. Then, the time-series data input sequence and the static process parameter input sequence are constructed respectively using the historical temperature time-series data and static process parameter data after feature filtering and preprocessing. A HybridKAN neural network model is designed, using time-series data input sequences and static process parameter input sequences as model inputs to extract dynamic time-series features and static process features, respectively. These two types of features are fused and connected to the current autoclave temperature jump to obtain the output features of the HybridKAN neural network model. These output features are used to predict the autoclave temperature at the next time step. The HybridKAN neural network model is built based on a KAN network and includes a time-series data processing module, a static process parameter processing module, a feature fusion module, and a linear output layer, wherein: The time-series data processing module and the static process parameter processing module perform feature extraction in parallel. The time-series data processing module takes the time-series data input sequence as input, passes it through a three-layer KAN network, and outputs dynamic time-series features. The static process parameter processing module takes the static process parameter input sequence as input, passes it through a three-layer KAN network, and outputs static process features. The dynamic time-series features and the static process features have the same feature dimension. Then, the dynamic time-series features and static process features are first spliced ​​and fused in the feature fusion module to obtain fused features. The fused features are then input into a four-layer KAN network to deeply couple the dynamic time-series features and static process features to obtain deeply coupled features. The deeply coupled features output by the feature fusion module are passed through a residual network to jump-connect the current temperature value in the time series data input sequence, and then the predicted temperature value for the next time step is obtained through a linear output layer. The training hyperparameters of the HybridKAN neural network model are set to minimize the loss function. The HybridKAN neural network model is trained and a structured pruning strategy is used to lightweight the model. The optimal HybridKAN neural network model is obtained after training and pruning. Deploying the optimal HybridKAN neural network model, and using a rolling time window for iterative multi-step prediction, enables offline simulation and online prediction of autoclave temperature. Specifically: The optimal HybridKAN neural network model is deployed to a computing device that interfaces with a database for offline simulation and online prediction. The offline simulation is as follows: In the absence of complete measured temperature data, the static process parameters and tank temperature are used as the initial inputs of the optimal HybridKAN neural network model. The tank temperature at each moment is recursively calculated to simulate the temperature evolution of the curing and molding process of the part under the current process scheme. The evolution results are presented in the form of temperature curves. The online prediction is as follows: when the computing device can obtain the measured temperature data collected by thermocouples from the database, the historical temperature time series data from the current moment backward, the static process parameter data, and the current moment index are used as the input of the optimal HybridKAN neural network model. The model outputs the temperature prediction value for the next moment, and generates a full-process temperature process curve that evolves over time by iterating through multiple steps in a rolling time window. This curve is used to guide the adjustment of the heating rate and holding time of the autoclave.

2. The method for offline simulation and online prediction of autoclave temperature based on HybridKAN feature fusion according to claim 1, characterized in that, The process of feature filtering and preprocessing the solidification and molding data of the parts, and then constructing the time-series data input sequence and the static process parameter input sequence using the historical temperature time-series data and static process parameter data after feature filtering and preprocessing, specifically includes: First, the static process parameter data are initially screened through the heat conduction mechanism. Then, based on the heat transfer principle in the autoclave, parameters that are strongly correlated with heat distribution are retained first, while secondary parameters are eliminated. Then, the static process parameter data is finally filtered using a data-driven method to construct a random forest model containing all static features. Based on the parameter importance ranking results, the top N static process parameters are selected to obtain an N-dimensional static process parameter input sequence. Data cleaning and anomaly handling were carried out. Outliers were removed from different batches of historical temperature time series data using the 3σ principle, and then linear interpolation was used to fill in the missing values. The historical temperature time series data, after data cleaning and anomaly processing, is processed by a sliding window of size L. The temperature values ​​of the L historical moments and the index i of the last historical moment are combined into a sequence to obtain an L+1 dimensional time series data input sequence. The negative historical temperature data before the initial zero moment of the sequence are filled with the initial ambient temperature of the autoclave. The static process parameter input sequence and timing data input sequence are used as inputs for subsequent networks.

3. The method for offline simulation and online prediction of autoclave temperature based on HybridKAN feature fusion according to claim 1, characterized in that, The loss function used is the mean squared error loss (MSE).

4. The method for offline simulation and online prediction of autoclave temperature based on HybridKAN feature fusion as described in claim 1, characterized in that, The specific steps of using a structured pruning strategy to lightweight the model are as follows: Neuron importance is determined by ranking based on the absolute value of weights, and low-importance neuron connections are removed according to a preset pruning rate. The pruned network is then fine-tuned and trained, and a learning rate decay strategy is used to restore model accuracy.

5. An application system for an offline simulation and online prediction method of autoclave temperature based on HybridKAN feature fusion according to any one of claims 1-4, characterized in that, Specifically, it includes: The data acquisition unit is used to acquire curing data of different batches of parts; The data processing unit is used to perform feature filtering and preprocessing on the solidified molding data of the parts; The model building unit is used to build the HybridKAN neural network model. The training optimization unit is used to train and prune the HybridKAN neural network model. The prediction application unit is used to deploy models and perform offline simulations and online predictions.

6. An electronic device, characterized in that, The electronic device includes a memory and a processor, wherein: Memory is used to store computer programs that can run on a processor; A processor is configured to, when running the computer program, execute the offline simulation and online prediction method for autoclave temperature based on HybridKAN feature fusion as described in any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause a processor to execute the offline simulation and online prediction method for autoclave temperature based on HybridKAN feature fusion as described in any one of claims 1-4.

Images