State space model identification method and device, electronic equipment and storage medium

Abstract

The invention provides a state space model identification method and device, electronic equipment and a storage medium, and is applied to the technical field of computers, and the method comprises the steps: obtaining input time sequence data and output time sequence data of a target industrial system, generating a training data set based on the input time sequence data and the output time sequence data, and carrying out the recognition of a state space model; based on an initial state space model created by a neural network, initial parameters of the initial state space model are optimized by using a particle swarm optimization algorithm and based on the training data set to obtain target initial parameters, and then the target state space model is obtained based on the target initial parameters and the training data set. And performing network training on the initial state space model by taking minimization of the total loss function as a target to obtain a state space model used for representing the target industrial system, the state space model identification method provided by the invention can efficiently and reliably construct the state space model suitable for the complex industrial environment, and the identification accuracy of the state space model is improved. And a basis is provided for system analysis prediction and advanced control of an industrial system.

Description

Technical Field

[0001] This application relates to the field of computer technology, specifically to a method, apparatus, electronic device, and storage medium for identifying state-space models. Background Technology

[0002] State-space models are one of the core mathematical models in modern control theory and system identification. These models typically consist of state equations and output equations, and can comprehensively characterize the dynamic evolution and response characteristics of a system. In industrial production, state-space models play an irreplaceable role: they are the cornerstone of advanced control algorithms such as model predictive control and optimal control. Based on state-space models, future system behavior can be predicted, thereby formulating optimal control strategies and improving product quality and production efficiency.

[0003] However, despite the great theoretical value and application potential of state-space models, there is still a lack of systematic identification methods for constructing a high-precision, interpretable, and practical model for a multi-input multi-output system with strong coupling, nonlinearity, and time-varying characteristics in actual industrial scenarios. Therefore, how to provide a state-space model identification method to efficiently and reliably construct a state-space model suitable for complex industrial environments, and provide a foundation for system analysis, prediction, and advanced control of industrial systems, has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] In view of this, this application aims to provide a state-space model identification method, device, electronic device and storage medium to efficiently and reliably construct state-space models suitable for complex industrial environments, and to provide a foundation for system analysis, prediction and advanced control of industrial systems.

[0005] Firstly, this application provides a method for identifying state-space models, including: Acquire the input time-series data and output time-series data of the target industrial system, and generate a training dataset based on the input time-series data and the output time-series data; An initial state space model is created based on a neural network, wherein the parameter matrix of the initial state space model is defined by the neuron nodes of the neural network; The initial parameters of the initial state-space model are optimized using the particle swarm optimization algorithm to obtain the target initial parameters; Based on the target initial parameters and the training dataset, the initial state-space model is trained to minimize the total loss function, thereby obtaining a state-space model that represents the target industrial system.

[0006] In one optional implementation, the initial state-space model created based on the neural network includes: Based on the physical coupling relationship between the input and output channels of the target industrial system, a relationship mask matrix is ​​generated; Based on the relational mask matrix, structural constraints are applied to the parameter matrix of the initial state space model, and the positions of non-zero elements in the parameter matrix are limited, so as to construct a structured initial state space model that matches the physical coupling relationship.

[0007] In one optional implementation, the structural constraints include: input channel constraints, output channel constraints, and state transition constraints, wherein, The input channel constraints are used to limit the state variables associated with each input channel in the input matrix of the state-space model; The output channel constraint is used to limit the state variables associated with each output channel in the output matrix of the state-space model; The state transition constraints are used to define the interaction relationships between different state subspaces in the state transition matrix of the state-space model.

[0008] In one optional implementation, the total loss function includes a prediction error term and a constraint penalty term; The constraint penalty term includes at least one of the following constraints: Stability constraint terms based on the eigenvalues ​​of the state transition matrix; Parameter boundary constraints are used to limit the range of parameter values; Output gain constraint terms used to limit the system's output gain; L1 regularization constraints are used to promote parameter sparsity.

[0009] In one optional implementation, a training dataset is generated based on the input time-series data and the output time-series data, including: The input timing data and the output timing data are normalized respectively. The normalized input time-series data and output time-series data are segmented using a sliding window method to generate a training dataset. The training dataset includes an input dataset and an output dataset. The input dataset includes multiple input data distributed in time sequence, and the output dataset includes multiple output data distributed in time sequence. The width of the sliding window is greater than the time required for the state change of the target industrial system.

[0010] In one optional implementation, based on the target initial parameters and the training dataset, the initial state-space model is trained as a network with the objective of minimizing the total loss function to obtain a state-space model representing the target industrial system, including: Based on the target initial parameters and the training dataset, the initial state space model is trained with the goal of minimizing the total loss function. During the training process, the model parameters are sparsified using a dynamic pruning algorithm to obtain a state space model that represents the target industrial system.

[0011] In an optional implementation, the state-space model identification method provided in the first aspect of this application further includes: Using the state-space model as the teacher model, and based on the training dataset, a student model with a smaller parameter scale is trained through knowledge distillation. The student model is determined to be the target state-space model of the target industrial system.

[0012] Secondly, this application provides a state-space model identification device, comprising: A data processing unit is used to acquire input time-series data and output time-series data of the target industrial system, and generate a training dataset based on the input time-series data and the output time-series data. A model creation unit is used to create an initial state space model based on a neural network, wherein the parameter matrix of the initial state space model is defined by the neuron nodes of the neural network; The parameter optimization unit is used to optimize the initial parameters of the initial state space model using the particle swarm optimization algorithm to obtain the target initial parameters. The model training unit is used to train the network on the initial state space model based on the target initial parameters and the training dataset, with the objective of minimizing the total loss function, to obtain a state space model that represents the target industrial system.

[0013] Thirdly, this application provides an electronic device including a memory, a processor, and a computer program stored in the memory and executed by the processor, wherein the processor executes the computer program to implement the steps of the state-space model identification method as described in any embodiment of the first aspect of this application.

[0014] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the state-space model identification method as described in any of the first aspects of this application.

[0015] Based on the above, the state-space model identification method provided in this application, after acquiring the input and output time-series data of the target industrial system and generating a training dataset based on the input and output time-series data, uses a particle swarm optimization algorithm to optimize the initial parameters of the initial state-space model based on the training dataset to obtain target initial parameters. Then, based on the target initial parameters and the training dataset, the initial state-space model is trained on a network with the objective of minimizing the total loss function to obtain a state-space model representing the target industrial system. The state-space model identification method provided in this application can efficiently and reliably construct state-space models suitable for complex industrial environments, providing a foundation for system analysis, prediction, and advanced control of industrial systems. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of a state-space model identification method provided in an embodiment of this application.

[0018] Figure 2 This is a flowchart of another state-space model identification method provided in the embodiments of this application.

[0019] Figure 3 This is a structural block diagram of a state-space model identification device provided in an embodiment of this application.

[0020] Figure 4 This is a structural block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0021] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] To address the lack of a systematic method for constructing state-space models for multi-input multi-output systems and complex industrial systems in existing technologies, this application provides a state-space model identification method that can efficiently and reliably construct state-space models suitable for complex industrial environments, providing a foundation for system analysis, prediction, and advanced control of industrial systems.

[0023] The state-space model identification method provided in this application is applied to electronic devices, such as laptops, personal computers (PCs), tablets, and other electronic devices capable of running applications corresponding to the state-space model identification method provided in this application. These will not be listed here. In some cases, it can also be applied to network-side servers. See [link to relevant documentation] Figure 1 As shown in the embodiments of this application, the state space model identification method includes the following steps.

[0024] S100. Obtain the input time-series data and output time-series data of the target industrial system, and generate a training dataset based on the input time-series data and output time-series data.

[0025] First, it should be noted that the target industrial system mentioned in the embodiments of this application refers to any industrial system that needs to be identified in state space model in actual application. It can be the multi-input multi-output system mentioned above, or other industrial systems with complex architectures. This application does not limit the specific selection of the target industrial system.

[0026] The state-space model identification method provided in this embodiment is based on relevant data during the operation of the target industrial system. Therefore, it is first necessary to obtain the input time-series data and output time-series data of the target industrial system, and then generate a training dataset for training the neural network created in subsequent steps based on the obtained input time-series data and output time-series data.

[0027] Understandably, in practical applications, the operation of the target industrial system depends on a large number of input time series data of various types. At the same time, as the target industrial system operates, it will also produce a large number of output time series data of different types. The dimensions (units) of the various time series data are very different. Directly using these time series data will greatly increase the difficulty of training the neural network. Therefore, it is necessary to normalize the obtained input time series data and output time series data separately.

[0028] Specifically, for any input time series data u, normalization can be performed according to the following formula: (1) in, This represents the result obtained after normalizing the input time series data u; This represents the average value of all input time series data; This represents the standard deviation of all input time series data.

[0029] Accordingly, for any output time series data y, it can be normalized using the following formula: (2) in, This indicates the result obtained after normalizing the output time series data; This represents the average value of all output time-series data y; This represents the standard deviation of all output time-series data.

[0030] Through the above normalization process, time series data with different physical dimensions can be transformed into similar numerical ranges, avoiding the dominance of training process by some time series data due to their excessive magnitude.

[0031] Furthermore, a sliding window method is used to segment the normalized input and output time-series data to generate a training dataset. Alternatively, a portion of the normalized input and output time-series data can be used as a validation dataset at this step. This validation dataset verifies the prediction performance of the state-space model obtained in subsequent steps, providing a reference for further optimization of the state-space model or adjustment of the training process.

[0032] It should be noted that, in this embodiment of the application, the width of the sliding window is required to be greater than the time required for the state change of the target industrial system, so as to ensure that the data obtained by dividing according to the sliding window can represent the complete system response process, rather than the intermediate data in the operation of the target process system. As for the specific process of dividing the normalized input time series data and output time series data according to the sliding window method, it can be implemented with reference to relevant technologies, and will not be described in detail here.

[0033] The resulting training dataset after segmentation includes the input dataset and the output dataset, which can be represented by the following formula: (3) Where w represents the width of the sliding window; This represents the i-th window; This represents the movement step size of the sliding window; This represents each input data point in the input dataset; This represents each output data in the output dataset.

[0034] As mentioned above, the embodiments of this application require the width of the sliding window to be greater than the time required for the state change of the target industrial system. Based on this, in practical applications, for typical industrial processes, the width of the sliding window is generally set to 3-5 times the main time constant of the target industrial system, and the moving step size can be adjusted according to the actual amount of data and computing resources, usually set to 1 / 5 to 1 / 10 of the sliding window width. In addition, according to formula (3), the obtained input dataset includes multiple input data distributed in time sequence, and correspondingly, the output dataset includes multiple output data distributed in time sequence. The final training dataset will be used as the basic data for subsequent model training.

[0035] S110, Initial state-space model created based on neural network.

[0036] A typical state-space model is represented by four parameter matrices: state transition matrix A, input matrix B, output matrix C, and perturbation matrix D. In the state-space model identification method provided in this application embodiment, the architecture of the state-space model is integrated with a neural network, making the neural network parameters physically interpretable, and the stability and engineering adaptability of the final model are enhanced by setting constraints.

[0037] Specifically, the parameter matrix of the initial state space model is defined as the neuron node of the neural network, such that each neuron node in the neural network corresponds to an element in the parameter matrix of the state space model, and the set of all neuron nodes constitutes the parameter matrix of the state space model.

[0038] It should also be noted that the state-space model strictly follows the state-space recursion mechanism in the process of predicting the system state, that is, for each time step, the state at the next time step is... Based on the current state and the current input, according to The recursive method yields the output, which is then obtained from... The calculation is as follows. During backpropagation, the gradient is recursively calculated according to the chain rule. Taking matrix parameters as an example, the recursive process can be followed by the following formula: (4) in, .

[0039] Similarly, other parameter matrices in the state-space model , , The backpropagation process of gradients also follows this process. When fusing neural networks with state-space models, the aforementioned characteristics of the state-space model need to be considered comprehensively. As for the details of the above recursive process not elaborated upon, please refer to relevant technical implementations; they will not be detailed here.

[0040] As an optional implementation, this application provides a method based on a masking mechanism that ensures the creation of an initial state-space model based on a neural network. Specifically, a relational mask matrix is ​​generated based on the physical coupling relationship between the input and output channels of the target industrial system. Furthermore, based on the obtained relational mask matrix, structural constraints are applied to the parameter matrix of the initial state-space model to limit the positions of non-zero elements in the parameter matrix, and finally a structured initial state-space model that matches the physical coupling relationship of the target industrial system is constructed.

[0041] Specifically, the physical coupling relationship between the input and output channels mentioned above can be determined based on historical experimental data or the engineering architecture of the target industrial system. Furthermore, based on the masking principle, the finite state dimension is divided into different state subspaces according to the channels, achieving structured modeling of the state-space network. The masking not only limits the parameter matrix in the state-space model... , , , The non-zero position can also ensure that each input-output channel has a clear state assignment.

[0042] In practical applications, the construction of relation mask matrices falls into two categories: one directly uses the original relation mask to ensure strict single-channel coupling; the other, while allowing collaboration between outputs, adds block connectivity to all corresponding state subspaces of the same output, enabling... A matrix can propagate information between multiple state sub-blocks of the same output. Specifically, this may include the following steps: A. Setting the relation matrix: Construct a relation matrix ,in, Indicates input channel There is a causal relationship with output channel p, and correspondingly, if If this happens, all subsequent states and parameters associated with this input-output pair will remain at zero.

[0043] B. State allocation: Let the total number of states be For each satisfied Assign a non-zero value to the node such that Further, record each one according to the order of relation enumeration. Corresponding index range ,in, And all intervals do not overlap.

[0044] C. Mask construction: In this application, the structural constraints involved in mask construction are divided into three types: input channel constraints, output channel constraints, and state transition constraints. The input channel constraints are used to limit the state variables associated with each input channel in the input matrix of the state space model. The output channel constraints are used to limit the state variables associated with each output channel in the output matrix of the state space model. The state transition constraints are used to limit the interaction relationships between different state subspaces in the state transition matrix of the state space model.

[0045] Specifically, regarding input channel constraints, for any Only allowed State and input in Connect, that is, connect the input matrix Apply a mask so that Regarding output channel constraints, that is, constraints on the output matrix... Apply a mask so that Thus, the output channel p can only observe intermediate states shared with its associated input. Regarding state transition constraints, for The block structure of the matrix is ​​restricted, firstly ensuring... The internal state sub-blocks are allowed to evolve freely. Furthermore, under a configuration that allows output coupling, multiple sets of s-indexes that share the same output channel can be joined together. And enable interactions between states within the union, otherwise these interactions are set to 0.

[0046] Through the masking mechanism described above, it is possible to effectively ensure that the network structure of the initial state space model is consistent with the physical prior and matches the physical coupling relationship between the channels in the target industrial system. This structured initial state space model not only preserves the physical coupling relationship in the target industrial system, but also provides a stable topological foundation for subsequent particle swarm optimization, dynamic pruning, and distillation compression, which is conducive to improving the controllability and interpretability of the identification process of complex multi-input multi-output systems.

[0047] As an optional implementation, to further improve the trainability of the initial spatial state model, or in other words, to further reduce the training difficulty of the initial spatial state model, a state transfer mechanism can be introduced based on the obtained initial state space model to further emphasize the correlation of temporal data. The state transfer mechanism ensures the continuity of states in consecutive data segments, thereby constructing a complete temporal recursive structure. Referring to the aforementioned related content, in each step of the loop, the new state... It is obtained recursively from the state at the previous time step, that is This allows for the linking of historical states with current inputs, providing a complete description of the dynamic process of the target industrial system over time. Therefore, in subsequent training phases, the state of the previous data can be used as a reference. recursion At any given moment, obtain the current system state. The state recursion mechanism allows for intervals between two training data sets. This approach ensures both a complete modeling of the dynamic process and enhances the flexibility of training and inference.

[0048] Furthermore, as another optional implementation, multiple constraint terms are introduced into the initial state-space model, including stability constraints, parameter boundary constraints, output gain constraints, and L1 regularization constraints, to ensure the stability and physical meaning of the network optimization results.

[0049] Specifically, the stability constraint term is determined based on the eigenvalues ​​of the state transition matrix in the state-space model, and is used to adjust the state transition matrix. The penalty constraint is applied to the real part of the eigenvalues, which can be expressed as the following formula: (5) in, for The first of the matrix 1 eigenvalue, This indicates taking the real part.

[0050] The parameter boundary constraint term, used to impose penalty constraints based on the range of parameter values, can be expressed by the following formula: (6) in, The fourth parameter in the four parameter matrices representing the spatial state model One parameter, , These are the upper and lower bounds of the parameter.

[0051] The output gain constraint term is used to penalize the portion of the gain of each output channel that exceeds a threshold. It can be expressed by the following formula: (7) in, For the first One output item, This is the upper limit of the gain for this output item.

[0052] The L1 regularization constraint term is used to promote parameter sparsity and reduce redundant parameters. It can be expressed by the following formula: (8) in, Represents the state transition matrix The j-th element in the i-th row, The state transition matrix is ​​represented by the given information. The sum of the absolute values ​​of the elements in the middle. , , Similarly.

[0053] Furthermore, based on the above constraints, the total loss function of the initial state-space model is constructed, which can be expressed by the following formula: (9) in, , , , These are weights used to balance the penalties of each constraint term and prevent any single constraint from dominating the gradient descent process. Lmse represents the prediction error term, which is the mean squared error between the predicted and true values. It measures the deviation between the current model output and the actual system output. The gradient of Lmse with respect to the model parameters is calculated using the backpropagation algorithm, allowing the model output to iteratively optimize in the direction of reducing deviation.

[0054] Based on the above operations, an initial state-space model can be obtained that is created based on a neural network and matches the physical coupling relationship with the target industrial system. At the same time, the constraint terms and total loss function of the initial state-space model are defined to meet the requirements of subsequent model training.

[0055] S120. Using the particle swarm optimization algorithm, the initial parameters of the initial state space model are optimized to obtain the target initial parameters.

[0056] Understandably, the training process of a neural network model depends not only on the training data, but also on the initial parameters used at the beginning of training. Properly setting the initial parameters can effectively improve the training efficiency of the model and avoid local optima.

[0057] Therefore, after constructing the initial state-space model, it is necessary to optimize the initial parameters of the initial state-space model based on the training dataset to obtain the selected target initial parameters. The identification method provided in this embodiment uses the particle swarm optimization algorithm to determine the target initial parameters. That is, in the initial training stage, the particle swarm optimization (PSO) algorithm is used to initialize and optimize the parameter matrix (A, B, C, D). By iteratively updating the state of the particles, the combination of initial parameters that minimizes the objective function is found. This combination of parameters will be used as the target initial parameters of the initial state-space model.

[0058] Furthermore, to fully utilize the historical data of the target industrial system, the optimization process can be based on historical data, with multiple groups of particles initialized by small-scale randomization, and the model re-optimized based on the existing historical experience model. In this mechanism, the initialization of the particle swarm is divided into two categories: the first group of particles directly adopts the historically optimal model parameters, while the remaining particles are based on the historical model parameters with small-scale random perturbations superimposed. This ensures both the preservation of historically optimal solutions and enhances the diversity of the search.

[0059] Assume the historical optimal model parameters are The total number of particle groups is ,but: (10) in, It is a set of random values ​​with small variance.

[0060] This initialization method ensures that at least one particle in the particle swarm is located at the historical optimal solution, allowing it to directly inherit the performance advantages of existing models. Simultaneously, the perturbation initialization of other particles provides the algorithm with the ability to explore new solutions, helping it escape local optima and improve global optimization performance. The specific execution process of the PSO algorithm can be found in relevant technical implementations and will not be detailed here.

[0061] S130. Based on the target initial parameters and training dataset, the network is trained on the initial state-space model with the goal of minimizing the total loss function, so as to obtain the state-space model used to represent the target industrial system.

[0062] After determining the target initial parameters of the initial state-space model, the network can be trained based on the target initial parameters in the training dataset generated in the previous steps. The training process aims to minimize the total loss function until a state-space model that meets the requirements is obtained. This state-space model represents the target industrial system. The total loss function comprehensively considers prediction error terms, stability constraints, parameter boundary constraints, output gain constraints, and L1 regularization constraints to ensure that the model balances prediction accuracy and robustness.

[0063] In one optional implementation, to further enhance the predictive power of the final state-space model, the model parameters can be sparsified during the training process of the neural network based on a dynamic pruning algorithm. First, neuron pruning can be initiated after a period of network optimization, resetting the values ​​of neurons below a threshold to zero. Then, the pruning ratio is gradually increased by dynamically adjusting the threshold until the target pruning ratio is reached. The threshold for each pruning step is determined by a ratio that linearly transitions from the initial pruning ratio to the final target pruning ratio. During this process, the role of effective parameters is gradually strengthened.

[0064] Specifically, the general process of dynamic pruning can be based on the following steps: A. Determine the pruning ratio for this round: (11) in, This represents the pruning ratio for this round. To achieve the target pruning ratio, The total number of pruning steps, This represents the current pruning step.

[0065] B. Determine the pruning threshold: Pruning threshold The threshold is determined by the quantiles of the absolute values ​​of all parameters, that is, by arranging the absolute values ​​of all parameters in an ordered manner and selecting the threshold from them according to the proportions calculated in step A above: (12) in, This is a quantile function used to return the set of absolute values ​​of the arguments. The Middle Value at the quantile As a threshold. For example, if Then the set 60% of the values ​​are less than .

[0066] C. Element pruning: For all parameters in the parameter matrix, based on the current threshold Pruning: After pruning, the influence of key parameters on the final training result can be further improved, ultimately yielding a state-space model to characterize the target industrial system. For a more detailed explanation of the dynamic pruning algorithm's execution process, please refer to relevant technical documentation; it will not be elaborated upon here.

[0067] The state-space model identification method provided in this application can efficiently and reliably construct state-space models suitable for complex industrial environments, providing a foundation for system analysis, prediction, and advanced control of industrial systems.

[0068] Compared to traditional identification methods (such as recursive least squares, maximum likelihood estimation, subspace identification, etc.), which rely on model assumptions and human experience and are difficult to adapt to complex dynamic systems, this method uses particle swarm optimization algorithm to initialize and optimize model parameters, which significantly improves the robustness and convergence speed of parameter optimization, reduces reliance on human experience, and makes the identification process more automated.

[0069] To address the issues of black-box training, lack of physical constraints, and poor model interpretability in deep learning identification methods, this method integrates state-space models with neural networks and introduces an input-output relationship masking mechanism. During training, it continuously constrains state allocation and parameter updates, giving each parameter in the network a clear physical meaning. Furthermore, by integrating various physical and engineering constraints (such as stability constraints and output gain constraints), it optimizes the model structure and parameters, thereby improving the model's interpretability and engineering applicability.

[0070] To address the issue of high parameter redundancy in neural networks, this method also employs L1 regularization constraints and dynamic pruning to dynamically filter model parameters, continuously enhancing the effectiveness of model parameters. This ensures that the final model possesses both the generalization performance of large-scale network models and the simplicity of small network structures, enabling the identification model to not only have strong expressive power but also improve model stability and robustness.

[0071] Furthermore, this method supports online iterative identification. As can be seen from the foregoing, this method has a complete parameter retraining and optimization mechanism based on historical model parameters, supports online data stream processing and real-time training and updating of model parameters, enhances the dynamic adaptability of the model, and achieves continuous optimization of structural parameters.

[0072] In summary, the state-space model identification method provided in this application not only solves the weaknesses of existing identification methods such as local optima, parameter initial value dependence, poor model interpretability, and weak online update capability, but also achieves automation of the identification process, model adaptation and compression, dynamic parameter optimization, and knowledge transfer through intelligent optimization and deep iteration mechanisms, significantly improving identification accuracy, efficiency, and practical engineering application capabilities.

[0073] Furthermore, this application also provides another method for state-space model identification, see [link to relevant documentation]. Figure 2 As shown above, in the aforementioned Figure 1 Based on the illustrated embodiment, the state-space model identification method provided in this embodiment further includes: S140. Using the state-space model as the teacher model, a student model with a smaller parameter scale is trained based on the training dataset through knowledge distillation, and the student model is determined as the target state-space model of the target industrial system.

[0074] It should be noted that the state-space model mentioned in this step refers to the final output of S130, which represents the target industrial system. The purpose of this step is to further simplify the obtained state-space model, so that the target state-space model is smaller in size while maintaining the generalization performance of large-scale network models, thereby reducing the performance requirements of the hardware running the target state-space model. During knowledge distillation, the loss function of the small network includes not only output consistency loss but also stability constraints and parameter boundary constraints, ensuring a balance between predictive performance and robustness.

[0075] As for the specific process of knowledge distillation, it can be implemented by combining the aforementioned content and related technologies, and will not be detailed here.

[0076] The state-space model identification method provided in this embodiment further simplifies the model based on the state-space model used to characterize the target industrial system. The resulting target state-space model not only has the generalization performance of a large-scale model, but also has a smaller model size and lower requirements for hardware computing power, which is conducive to the promotion and application of the obtained target state-space model.

[0077] Understandably, in practical applications, Figure 1 and Figure 2 The state-space models provided in the corresponding embodiments can all be applied to actual industrial systems. The specific selection needs to be combined with the performance of the hardware device loading the model, and this application does not make any specific limitations in this regard.

[0078] Furthermore, after obtaining the state-space model of the target industrial system, the application effect of the obtained state-space model can be tested based on the aforementioned validation dataset. The prediction results of the state-space model can be displayed through visualization methods, and the parameters can be fine-tuned based on the test results to finally obtain a state-space model that conforms to the actual situation of the industrial system.

[0079] The state-space model identification device provided by this invention is described below. This device belongs to the same concept as the state-space model identification method provided in the embodiments of this application, and can execute the state-space model identification method provided in any embodiment of this application. It possesses the corresponding functional modules and beneficial effects of the state-space model identification method. Technical details not described in detail in this embodiment can be found in the state-space model identification method provided in the embodiments of this application, and will not be repeated here.

[0080] See Figure 3 The state-space model identification device provided in this application includes: The data processing unit 10 is used to acquire the input time-series data and output time-series data of the target industrial system, and generate a training dataset based on the input time-series data and the output time-series data. The model creation unit 20 is used to create an initial state space model based on a neural network, wherein the parameter matrix of the initial state space model is defined by the neuron nodes of the neural network. The parameter optimization unit 30 is used to optimize the initial parameters of the initial state space model using the particle swarm optimization algorithm to obtain the target initial parameters; The model training unit 40 is used to train the network on the initial state space model based on the target initial parameters and the training dataset, with the goal of minimizing the total loss function, to obtain a state space model that represents the target industrial system.

[0081] Below, for reference Figure 4 The electronic device provided in this embodiment of the invention may include: at least one processor 100, at least one communication interface 200, at least one memory 300, and at least one communication bus 400. In this embodiment of the invention, the number of processor 100, communication interface 200, memory 300, and communication bus 400 is at least one, and the processor 100, communication interface 200, and memory 300 communicate with each other through communication bus 400; obviously, Figure 4 The communication connections shown for the processor 100, communication interface 200, memory 300, and communication bus 400 are optional. Optionally, the communication interface 200 can be an interface of a communication module, such as the interface of a GSM module; the processor 100 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention.

[0082] The memory 300 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0083] Specifically, the processor 100 is used to execute the application program in the memory to implement the steps of the state space model identification method described above.

[0084] In some embodiments, this embodiment also provides a computer-readable storage medium, such as a floppy disk, optical disk, hard disk, flash memory, USB flash drive, SD (Secure Digital Memory Card), MMC (Multimedia Card), etc., in which one or more instructions implementing the above steps are stored. When these one or more instructions are executed by one or more processors, the processors perform the state-space model identification method described above. For specific implementation details, please refer to the foregoing description; further elaboration is not provided here.

[0085] In addition to the methods and apparatus described above, embodiments of this application may also be computer program products, which include computer program instructions that, when executed by a processor, cause the processor to perform the steps in the state space model identification methods according to various embodiments of this application as described above.

[0086] Computer program products can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of this application. The programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0087] Those skilled in the art will understand that the contents disclosed herein can be varied and modified in many ways. For example, the various devices or components described above can be implemented in hardware, or in software, firmware, or a combination of some or all of the three.

[0088] Furthermore, while this disclosure makes various references to certain elements of systems according to embodiments of this disclosure, any number of different elements may be used and operated on clients and / or servers. Elements are merely illustrative, and different aspects of the system and method may use different elements.

[0089] This disclosure uses flowcharts to illustrate the steps of a method according to embodiments of this disclosure. It should be understood that the preceding or following steps are not necessarily performed in exact order. Instead, the steps can be processed in reverse order or simultaneously. Furthermore, other operations can be added to these processes.

[0090] Those skilled in the art will understand that all or part of the steps in the above methods can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium, such as a read-only memory. Optionally, all or part of the steps in the above embodiments can also be implemented using one or more integrated circuits. Accordingly, each module / unit in the above embodiments can be implemented in hardware or as a software functional module. This disclosure is not limited to any particular combination of hardware and software.

[0091] Unless otherwise defined, all terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains. It should also be understood that terms such as those defined in a common dictionary should be interpreted as having a meaning consistent with their meaning in the context of the relevant art, and not as having an idealized or highly formalized meaning, unless expressly defined herein.

[0092] The foregoing description is intended to illustrate the present disclosure and should not be construed as limiting it. While several exemplary embodiments of the present disclosure have been described, those skilled in the art will readily understand that many modifications may be made to the exemplary embodiments without departing from the novel teachings and advantages of the present disclosure. Therefore, all such modifications are intended to be included within the scope of the present disclosure as defined by the claims. It should be understood that the foregoing description is intended to illustrate the present disclosure and should not be construed as limiting it to the specific embodiments disclosed, and modifications to the disclosed embodiments and other embodiments are intended to be included within the scope of the appended claims. The present disclosure is defined by the claims and their equivalents.

Claims

1. A method for identifying state-space models, characterized in that, include: Acquire the input time-series data and output time-series data of the target industrial system, and generate a training dataset based on the input time-series data and the output time-series data; An initial state space model is created based on a neural network, wherein the parameter matrix of the initial state space model is defined by the neuron nodes of the neural network; The initial parameters of the initial state-space model are optimized using the particle swarm optimization algorithm to obtain the target initial parameters; Based on the target initial parameters and the training dataset, the initial state-space model is trained to minimize the total loss function, thereby obtaining a state-space model that represents the target industrial system.

2. The method according to claim 1, characterized in that, The initial state-space model created based on the neural network includes: Based on the physical coupling relationship between the input and output channels of the target industrial system, a relationship mask matrix is ​​generated; Based on the relational mask matrix, structural constraints are applied to the parameter matrix of the initial state space model, and the positions of non-zero elements in the parameter matrix are limited, so as to construct a structured initial state space model that matches the physical coupling relationship.

3. The method according to claim 2, characterized in that, The structural constraints include: input channel constraints, output channel constraints, and state transition constraints, wherein, The input channel constraints are used to limit the state variables associated with each input channel in the input matrix of the state-space model; The output channel constraint is used to limit the state variables associated with each output channel in the output matrix of the state-space model; The state transition constraints are used to define the interaction relationships between different state subspaces in the state transition matrix of the state-space model.

4. The method according to claim 1, characterized in that, The total loss function includes a prediction error term and a constraint penalty term; The constraint penalty term includes at least one of the following constraints: Stability constraint terms based on the eigenvalues ​​of the state transition matrix; Parameter boundary constraints are used to limit the range of parameter values; Output gain constraint terms used to limit the system's output gain; L1 regularization constraints are used to promote parameter sparsity.

5. The method according to claim 1, characterized in that, A training dataset is generated based on the input time-series data and the output time-series data, including: The input timing data and the output timing data are normalized respectively. The normalized input time-series data and output time-series data are segmented using a sliding window method to generate a training dataset. The training dataset includes an input dataset and an output dataset. The input dataset includes multiple input data distributed in time sequence, and the output dataset includes multiple output data distributed in time sequence. The width of the sliding window is greater than the time required for the state change of the target industrial system.

6. The method according to claim 1, characterized in that, Based on the initial target parameters and the training dataset, the initial state-space model is trained to minimize the total loss function, resulting in a state-space model characterizing the target industrial system, including: Based on the target initial parameters and the training dataset, the initial state space model is trained with the goal of minimizing the total loss function. During the training process, the model parameters are sparsified using a dynamic pruning algorithm to obtain a state space model that represents the target industrial system.

7. The method according to any one of claims 1 to 6, characterized in that, Also includes: Using the state-space model as the teacher model, and based on the training dataset, a student model with a smaller parameter scale is trained through knowledge distillation. The student model is determined to be the target state-space model of the target industrial system.

8. A state-space model identification device, characterized in that, include: A data processing unit is used to acquire input time-series data and output time-series data of the target industrial system, and generate a training dataset based on the input time-series data and the output time-series data. A model creation unit is used to create an initial state space model based on a neural network, wherein the parameter matrix of the initial state space model is defined by the neuron nodes of the neural network; The parameter optimization unit is used to optimize the initial parameters of the initial state space model using the particle swarm optimization algorithm to obtain the target initial parameters. The model training unit is used to train the network on the initial state space model based on the target initial parameters and the training dataset, with the objective of minimizing the total loss function, to obtain a state space model that represents the target industrial system.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executed by the processor, characterized in that, When the processor executes the computer program, it implements the steps of the state-space model identification method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the state-space model identification method as described in any one of claims 1 to 7.

Images