Intelligent material path planning method and device in powder conveying system

Abstract

The present application discloses a method and device for intelligent material path planning in a powder conveying system. The method includes: acquiring multi-source sensor data sent by sensors, the sensors are used to monitor the operation of the powder conveying system, the multi-source sensor data include material status data, environmental parameter data and equipment operation data, and extracting features and aligning the multi-source sensor data in time and space based on a pre-established Gaussian edge network to obtain a time-space alignment feature matrix; constructing a generative intervention model based on the time-space alignment feature matrix; training a material path planning strategy based on an equivariance local neural network and a generative intervention model to obtain a path planning strategy model; using the path planning strategy model to generate a powder conveying path, and the powder conveying path is used to control the powder conveying system to perform powder conveying. The present application realizes accurate modeling and intelligent path planning of complex powder conveying environments, and improves conveying efficiency and system stability.

Description

Technical Field

[0001] The present application belongs to the field of artificial intelligence and industrial automation technology, and specifically relates to an intelligent material path planning method and device for a powder conveying system. The present application is mainly used for intelligent path planning and control in the powder material conveying process in industries such as cement, chemical industry, and metallurgy. Background Art

[0002] Powder material transportation is an important part of modern industrial production. Precise control of the transportation process is of great significance to improving production efficiency and product quality. At present, traditional powder transportation systems usually adopt fixed path planning methods based on preset rules, but cannot be dynamically adjusted according to real-time working conditions. Or, the path adjustment method is based on simple sensor feedback, but lacks a systematic optimization strategy.

[0003] With the development of machine learning technology, attempts have been made to apply traditional machine learning algorithms to the path planning of powder conveying systems. Such methods usually train prediction models based on historical data to optimize the conveying path. However, due to the large number of complex dynamic interference factors in the powder conveying process, such as changes in material properties and fluctuations in environmental conditions, traditional machine learning methods are difficult to accurately capture and respond to these dynamic changes.

[0004] Therefore, the main problem with the existing technology is that it is difficult to effectively deal with the dynamic interference in the powder conveying process and lacks in-depth modeling of the spatiotemporal characteristics of the material trajectory, which affects the accuracy of path planning and leads to low conveying efficiency. Summary of the invention

[0005] In view of this, the present application provides an intelligent material path planning method and device in a powder conveying system, which solves the problem in the prior art that it is difficult to effectively deal with dynamic interference in the powder conveying process.

[0006] In a first aspect, an embodiment of the present application provides an intelligent material path planning method in a powder conveying system, comprising:

[0007] Acquire multi-source sensor data sent by a sensor, wherein the sensor is used to monitor the operation of the powder conveying system, and the multi-source sensor data includes material status data, environmental parameter data, and equipment operation data;

[0008] Based on a pre-established Gaussian edge network, feature extraction and spatiotemporal alignment are performed on the multi-source sensor data to obtain a spatiotemporal alignment feature matrix, where the spatiotemporal alignment feature matrix is ​​used to indicate a feature vector at any time point and spatial position;

[0009] constructing a generative intervention model according to the spatiotemporal alignment feature matrix, wherein the generative intervention model is used to indicate interference factors in a powder conveying process;

[0010] Training the material path planning strategy based on the pre-established equivariance local neural network and the generative intervention model to obtain a path planning strategy model;

[0011] The path planning strategy model is used to generate a powder conveying path, and the powder conveying path is used to control the powder conveying system to perform powder conveying.

[0012] In a possible implementation, the feature extraction and spatiotemporal alignment of the multi-source sensor data based on a pre-established Gaussian edge network to obtain a spatiotemporal alignment feature matrix includes:

[0013] Performing standardization processing on the multi-source sensor data, wherein the standardization processing includes calculating the mean and standard deviation of the data and converting the original data into a standard score;

[0014] Performing feature extraction on the standardized multi-source sensor data based on a preset multi-layer perceptron, wherein the multi-layer perceptron includes a ReLU activation function;

[0015] Constructing a Gaussian edge model, wherein the Gaussian edge model is used to indicate that a potential feature representation is obtained by performing agnostic learning on a feature extraction result through a preset edge probability density function;

[0016] The latent feature representations are temporally synchronized and spatially aligned to obtain a spatiotemporally aligned feature matrix.

[0017] In a possible implementation, performing time synchronization and spatial registration on the latent feature representation to obtain a spatiotemporal alignment feature matrix includes:

[0018] Based on the potential feature representation, a spatiotemporal reference system is established with a sampling period of the powder conveying system as a reference time unit;

[0019] Based on the space-time reference system, using a preset dynamic time warping algorithm to perform time synchronization on the potential feature representation;

[0020] A three-dimensional coordinate system is established by taking the center line of the conveying pipeline of the powder conveying system as the reference axis of the spatial coordinate system;

[0021] Based on the three-dimensional coordinate system, the potential feature representation is spatially registered using a preset spatial interpolation algorithm and a preset adaptive weight strategy, wherein the adaptive weight strategy is used to indicate adjusting the weight of the sensor in the interpolation process according to the distance between the point to be interpolated and the sensor corresponding to the potential feature representation;

[0022] The temporally synchronized latent feature representation and the spatially aligned latent feature representation are integrated into a spatiotemporally aligned feature matrix.

[0023] In a possible implementation, the generating an intervention model according to the spatiotemporal alignment feature matrix includes:

[0024] Based on the spatiotemporal alignment feature matrix, construct an initial interference model;

[0025] In the initial interference model, the associated data of the interference factors are introduced to obtain a generated intervention model;

[0026] The step of constructing an initial interference model based on the spatiotemporal alignment feature matrix includes:

[0027] Based on the spatiotemporal alignment feature matrix, construct an encoder, wherein the encoder includes a preset bidirectional LSTM network;

[0028] Using a preset conditional generative network as a decoder, and constructing a model framework based on the encoder and the decoder;

[0029] Establishing an interference recognition module, and establishing an initial model based on the model framework and the interference recognition module, wherein the interference recognition module is used to instruct to obtain a specified interference factor based on a preset attention mechanism;

[0030] Use the preset fully convolutional network structure to build the discriminator;

[0031] Determine the adversarial loss function and the reconstruction loss function of the discriminator and the generator;

[0032] Based on the adversarial loss function and the reconstruction loss function, adversarial training is performed on the initial model to obtain an interference initial model.

[0033] In a possible implementation, the introduction of the associated data of the interference factors into the initial interference model to generate the intervention model includes:

[0034] Determining a cause-and-effect relationship between the interference factor in the powder conveying system;

[0035] Based on the causal relationship, a dynamic causal graph model is constructed, where the dynamic causal graph model is used to describe the temporal dependency relationship between data;

[0036] By using a preset counterfactual reasoning algorithm, an interference impact assessment is performed on the interference factor to obtain an interference assessment result;

[0037] The dynamic causal graph model and the interference assessment result are introduced into the initial interference model to obtain a generated intervention model.

[0038] In a possible implementation, the method further includes:

[0039] Constructing a neural network structure with a weight sharing mechanism, wherein the weight sharing mechanism is used to indicate a weight template in a cyclic shared weight matrix, wherein the weight template is determined by an input of the neural network structure;

[0040] Determining a local connection mode, wherein the local connection mode is used to indicate that an input space of the neural network structure is divided into a plurality of local receptive fields and a local connection matrix is ​​determined;

[0041] Based on the neural network structure and the local connection mode, an equivariance constraint is introduced to obtain an equivariance local neural network.

[0042] In a possible implementation, the material path planning strategy is trained based on the pre-established equivariance local neural network and the generated intervention model to obtain a path planning strategy model, including:

[0043] constructing an environmental model based on the generative intervention model;

[0044] Based on the equivariance local neural network and the environment model, determining a deep reinforcement learning framework, wherein the deep reinforcement learning framework includes a value network structure, a policy network structure, and a reward function;

[0045] Based on the environment model and the deep reinforcement learning framework, a preset proximal strategy optimization algorithm is used to optimize the decision network;

[0046] Based on the optimized decision network, the path planning strategy model is obtained.

[0047] In a possible implementation, the step of using the path planning strategy model to generate a powder conveying path includes:

[0048] Based on the path planning strategy model, calculating a speed adjustment amount and a direction adjustment amount for powder conveying;

[0049] The speed adjustment amount and the direction adjustment amount are converted into control instructions acting on the powder conveying system, and the control instructions are used to instruct the powder conveying system to perform powder conveying.

[0050] In a possible implementation, it also includes:

[0051] Monitoring the operating status of the powder conveying system;

[0052] When it is detected that the operating state is in an abnormal state, the control instruction is adjusted based on a preset abnormality handling strategy, and the abnormal state includes at least one of material blockage, pressure fluctuation and temperature abnormality.

[0053] In a second aspect, an embodiment of the present application provides an intelligent material path planning device in a powder conveying system, comprising:

[0054] A data acquisition module, used to acquire multi-source sensor data sent by sensors, wherein the sensors are used to monitor the operation of the powder conveying system, and the multi-source sensor data includes material status data, environmental parameter data, and equipment operation data;

[0055] A feature processing module, used for performing feature extraction and spatiotemporal alignment on the multi-source sensor data based on a pre-established Gaussian edge network to obtain a spatiotemporal alignment feature matrix, wherein the spatiotemporal alignment feature matrix is ​​used to indicate a feature vector at any time point and spatial position;

[0056] An intervention model module, used for constructing a generative intervention model according to the spatiotemporal alignment feature matrix, wherein the generative intervention model is used for indicating interference factors in the powder conveying process;

[0057] A strategy optimization module, used for training the material path planning strategy based on the pre-established equivariance local neural network and the generative intervention model to obtain a path planning strategy model;

[0058] A control execution module is used to generate a powder conveying path using the path planning strategy model, and the powder conveying path is used to control the powder conveying system to perform powder conveying.

[0059] The intelligent material path planning method and device in the powder conveying system provided in the embodiment of the present application, by introducing a Gaussian edge network to extract features and align time and space of multi-source sensor data, achieves accurate modeling of complex powder conveying environments, ensures the time and space consistency of data, and improves the perception and adaptability of the system. Subsequently, based on the generative intervention model, the interference factors are modeled and evaluated, which enhances the system's ability to cope with dynamic interference and significantly improves the stability of the conveying process. The use of equivariance local neural networks and generative intervention models for path planning reduces the computational complexity while ensuring model performance, improves the system's real-time response capability and control accuracy, thereby improving the accuracy of path planning and improving conveying efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the disclosed embodiments in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without paying any creative work.

[0061] Figure 1A schematic diagram of a flow chart of an intelligent material path planning method in a powder conveying system provided in an embodiment of the present application;

[0062] Figure 2 A schematic diagram of the structure of an intelligent material path planning device in a powder conveying system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0063] The embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0064] It should be clear that the following embodiments of the present disclosure are described by specific specific examples, and those skilled in the art can easily understand other advantages and effects of the present disclosure from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all of the embodiments. The present disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that the following embodiments and features in the embodiments can be combined with each other in the absence of conflict. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present disclosure.

[0065] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein may be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on the present disclosure, it should be understood by those skilled in the art that an aspect described herein may be implemented independently of any other aspect, and two or more of these aspects may be combined in various ways. For example, any number of aspects described herein may be used to implement the device and / or practice the method. In addition, other structures and / or functionalities other than one or more of the aspects described herein may be used to implement this device and / or practice this method.

[0066] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present disclosure. The drawings only show components related to the present disclosure rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0067] Additionally, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, it will be understood by those skilled in the art that the aspects described may be practiced without these specific details.

[0068] The specific implementation of the present application is described in detail below in conjunction with the accompanying drawings. It should be noted that the drawings in the specification of the present application are only used to explain and illustrate the present application, and are not intended to limit the scope of protection of the present application.

[0069] Figure 1 A schematic diagram of a flow chart of an intelligent material path planning method in a powder conveying system provided in an embodiment of the present application. Figure 1 The intelligent material path planning method in the powder conveying system provided in the embodiment of the present application includes steps S101 to S105.

[0070] S101. Acquire multi-source sensor data sent by a sensor, where the sensor is used to monitor the operation of a powder conveying system, and the multi-source sensor data includes material status data, environmental parameter data, and equipment operation data.

[0071] Specifically, monitoring data is obtained from multiple sensors arranged in the powder conveying system. The sensors include but are not limited to: sensors for monitoring material status, such as sensors for measuring material particle size, moisture content and temperature; sensors for monitoring environmental parameters, such as sensors for measuring ambient temperature, humidity and air pressure; and sensors for monitoring equipment operation status, such as sensors for measuring motor speed, vibration frequency and pipeline pressure. These sensors are connected to the control system through a preset communication interface and collect data in real time according to a predetermined sampling frequency (e.g., 100Hz). Material status data includes material particle size, moisture content, temperature, etc., environmental parameter data includes ambient temperature, humidity, air pressure, etc., and equipment operation data includes motor speed, vibration frequency, pipeline pressure, etc.

[0072] S102: performing feature extraction and spatiotemporal alignment on the multi-source sensor data based on a pre-established Gaussian edge network to obtain a spatiotemporal alignment feature matrix, where the spatiotemporal alignment feature matrix is ​​used to indicate a feature vector at any time point and spatial position.

[0073] In this application, the potential features of multi-source sensor data are extracted through Gaussian edge network and agnostic learning for spatiotemporal alignment processing, thereby achieving the alignment of multi-source data and finally outputting a regularized feature matrix for subsequent analysis, namely the spatiotemporal alignment feature matrix. This hierarchical processing method ensures both the effectiveness of feature extraction and the spatiotemporal consistency of the data.

[0074] S103: constructing a generative intervention model according to the spatiotemporal alignment feature matrix, wherein the generative intervention model is used to indicate interference factors in the powder conveying process.

[0075] In this application, the generative intervention model provides a basic framework for material trajectory modeling, and introduces a causal inference mechanism to achieve accurate modeling of interference factors. This method of combining the generative intervention model and causal inference not only ensures the expressive power of the model, but also provides an explainable causal structure, providing a reliable environmental model for subsequent path planning.

[0076] S104, training the material path planning strategy based on the pre-established equivariance local neural network and the generated intervention model to obtain a path planning strategy model.

[0077] In the present application, the equivariance local neural network provides an efficient and stable model foundation. Based on the model structure, the strategy optimization is performed in combination with the generated intervention model of step S103. This not only ensures the expressiveness and stability of the model, but also achieves effective adaptation to complex environments. The optimized path planning strategy can be directly applied to the actual powder conveying system control.

[0078] S105 . Generate a powder conveying path using the path planning strategy model, where the powder conveying path is used to control the powder conveying system to perform powder conveying.

[0079] The intelligent material path planning method in the powder conveying system provided in this embodiment realizes accurate modeling of complex powder conveying environment by introducing Gaussian edge network to extract features and align time and space of multi-source sensor data, ensures the time and space consistency of data, and improves the perception and adaptability of the system. Then, based on the generative intervention model, the interference factors are modeled and evaluated, which enhances the system's ability to cope with dynamic interference and significantly improves the stability of the conveying process. The use of equivariance local neural network and generative intervention model for path planning reduces the computational complexity while ensuring model performance, improves the real-time response capability and control accuracy of the system, thereby improving the accuracy of path planning and improving conveying efficiency.

[0080] In some embodiments, the feature extraction and spatiotemporal alignment of the multi-source sensor data based on the pre-established Gaussian edge network to obtain the spatiotemporal alignment feature matrix includes:

[0081] Performing standardization processing on the multi-source sensor data, wherein the standardization processing includes calculating the mean and standard deviation of the data and converting the original data into a standard score;

[0082] Performing feature extraction on the standardized multi-source sensor data based on a preset multi-layer perceptron, wherein the multi-layer perceptron includes a ReLU activation function;

[0083] Constructing a Gaussian edge model, wherein the Gaussian edge model is used to indicate that a potential feature representation is obtained by performing agnostic learning on a feature extraction result through a preset edge probability density function;

[0084] The latent feature representations are temporally synchronized and spatially aligned to obtain a spatiotemporally aligned feature matrix.

[0085] Specifically, the acquired multi-source sensor data is subjected to Z-score standardization processing to convert data of different dimensions into a unified scale space. Exemplarily, the standardization processing is performed by the following calculation formula:

[0086] x' = (x - μ) / σ;

[0087] Among them, x' is the standardized multi-source sensor data, x is the multi-source sensor data, μ is the data mean, and σ is the data standard deviation. The standardization process in this embodiment helps to eliminate the dimensional differences between different types of data and provide a unified data basis for subsequent feature extraction.

[0088] Then, a multi-layer perceptron is used to extract features from the standardized multi-source sensor data. Exemplarily, the multi-layer perceptron includes three hidden layers, with 256, 128 and 64 neurons in each layer, and a ReLU activation function is connected after each layer. Through this multi-layer structure, deep feature extraction of multi-source sensor data is achieved, and data of different dimensions are mapped to a unified feature space.

[0089] Furthermore, a Gaussian marginal model is introduced for agnostic learning. The model predicts the mean and covariance matrix of the data distribution through a neural network, thereby realizing adaptive modeling of the unknown distribution. In the specific implementation process, the model uses a variational inference method for parameter optimization. Specifically, the variational distribution and prior distribution are first defined, and then parameter optimization is achieved by maximizing the lower bound of evidence. The advantage of this method is that it can effectively handle the uncertainty of data distribution and improve the adaptability of the model to unknown distributions. In addition, by introducing KL divergence as a regularization term, the model can be prevented from overfitting and its generalization ability can be enhanced. In practical applications, the model can automatically adjust its internal parameters according to the distribution characteristics of the input data, thereby realizing adaptive processing of different types of data.

[0090] Furthermore, a unified spatiotemporal reference system and three-dimensional coordinate system are established through latent feature representation, the time dimension alignment is achieved through the DTW algorithm, and the spatial dimension alignment is completed using the Kriging interpolation method, and finally the spatiotemporal alignment feature matrix is ​​output.

[0091] Therefore, this embodiment standardizes the multi-source sensor data through the processing of the Gaussian edge model, and then performs feature extraction, and then realizes preliminary feature extraction through the multi-layer perceptron, and then uses the unknowable learning mechanism of the Gaussian edge model to realize adaptive modeling of unknown distribution and obtain potential feature representation, thereby realizing the alignment of multi-source data, which not only ensures the effectiveness of feature extraction, but also ensures the temporal and spatial consistency of data.

[0092] Based on the above embodiment, in some embodiments, performing time synchronization and spatial registration on the potential feature representation to obtain a spatiotemporal alignment feature matrix includes:

[0093] Based on the potential feature representation, a spatiotemporal reference system is established with a sampling period of the powder conveying system as a reference time unit;

[0094] Based on the space-time reference system, using a preset dynamic time warping algorithm to perform time synchronization on the potential feature representation;

[0095] A three-dimensional coordinate system is established by taking the center line of the conveying pipeline of the powder conveying system as the reference axis of the spatial coordinate system;

[0096] Based on the three-dimensional coordinate system, the potential feature representation is spatially registered using a preset spatial interpolation algorithm and a preset adaptive weight strategy, wherein the adaptive weight strategy is used to indicate adjusting the weight of the sensor in the interpolation process according to the distance between the point to be interpolated and the sensor corresponding to the potential feature representation;

[0097] The temporally synchronized latent feature representation and the spatially aligned latent feature representation are integrated into a spatiotemporally aligned feature matrix.

[0098] In this embodiment, a unified spatiotemporal reference system is established. In the time dimension, the sampling period of the system is used as the reference time unit to ensure that all sensor data are consistent in time scale. In the spatial dimension, the center line of the conveying pipeline is used as the reference axis of the spatial coordinate system to establish a three-dimensional coordinate system, and the positions of all sensors are mapped to this unified coordinate system. This unified reference system provides a basic framework for subsequent spatiotemporal alignment.

[0099] Then, time synchronization is achieved. The dynamic time warping algorithm is used to time align the data collected by different sensors. In the specific implementation, a distance matrix is ​​first constructed to represent the distance relationship between corresponding points in different time series. Then the cumulative distance matrix is ​​calculated, and the optimal time alignment path is found through the dynamic programming method. This method can effectively deal with the sampling frequency differences and time delay problems between different sensors.

[0100] Furthermore, spatial registration is performed. The Kriging spatial interpolation method is used to achieve the alignment of spatial dimensions. This method is based on the theory of spatial statistics and describes spatial correlation by constructing a variogram. In practical applications, it is first necessary to analyze the spatial distribution characteristics of the data, and then select a suitable variogram model based on these characteristics.

[0101] In order to improve the accuracy of interpolation, an adaptive weight strategy is introduced to dynamically adjust the weight of each sensor in the interpolation process according to its reliability. Specifically, sensors closer to the interpolation point are assigned larger weights, while sensors farther away are assigned smaller weights. In addition, factors such as the measurement accuracy and historical reliability of the sensor are also considered to further optimize the weight allocation.

[0102] Exemplarily, the weight w_i in the adaptive weight strategy is calculated by the following formula:

[0103] w_i = exp(-d_i² / 2σ²);

[0104] Where d_i is the distance from the point to be interpolated to the i-th sensor, and σ is an adjustable scale parameter.

[0105] Furthermore, the results of time synchronization and space registration are integrated. The alignment results of the time dimension and the space dimension are fused to output a unified time-space alignment feature matrix. Each row of the matrix represents a time point, and each column represents a feature vector of a spatial position. This matrix structure enables the system to easily obtain the feature information of any time and space point, providing a regularized data foundation for subsequent interference modeling and path planning. In actual applications, the matrix will be continuously updated with the arrival of new data to keep the system accurately grasp the real-time status.

[0106] Therefore, through the processing of the above steps, this embodiment finally obtains a feature representation with good consistency in both time and space dimensions, which provides reliable data support for subsequent interference modeling and path planning. The advantage of this processing method is that it can effectively handle the alignment problem of multi-source heterogeneous data and improve the perception accuracy and reliability of the system.

[0107] In some embodiments, generating an intervention model according to the spatiotemporal alignment feature matrix comprises:

[0108] Based on the spatiotemporal alignment feature matrix, construct an initial interference model;

[0109] In the initial interference model, the associated data of the interference factors are introduced to obtain a generated intervention model.

[0110] Based on the above embodiment, in some embodiments, constructing an initial interference model based on the spatiotemporal alignment feature matrix includes:

[0111] Based on the spatiotemporal alignment feature matrix, construct an encoder, wherein the encoder includes a preset bidirectional LSTM network;

[0112] Using a preset conditional generative network as a decoder, and constructing a model framework based on the encoder and the decoder;

[0113] Establishing an interference recognition module, and establishing an initial model based on the model framework and the interference recognition module, wherein the interference recognition module is used to instruct to obtain a specified interference factor based on a preset attention mechanism;

[0114] Use the preset fully convolutional network structure to build the discriminator;

[0115] Determine the adversarial loss function and the reconstruction loss function of the discriminator and the generator;

[0116] Based on the adversarial loss function and the reconstruction loss function, adversarial training is performed on the initial model to obtain an interference initial model.

[0117] In this embodiment, it is first necessary to build an encoder-decoder framework, that is, a model framework. On the encoder side, a bidirectional LSTM network is used to process time series features. The network receives a spatiotemporal alignment feature matrix as input, which contains information in three dimensions: time step, number of spatial points, and feature dimension. The bidirectional LSTM layer processes sequence data in both forward and backward directions. This bidirectional processing mechanism can fully capture the contextual dependencies of the data. During the processing, the state of each time step contains information about the two hidden states, forward and backward. In order to highlight important features, an attention mechanism is introduced after the bidirectional LSTM layer. This mechanism enables the model to adaptively focus on the information of key time points by calculating the importance weights of different time steps.

[0118] At the decoder end, a conditional generative network is used for processing. In specific implementation, the context vector output by the encoder is first concatenated with the randomly sampled noise vector to form the input of the decoder. Then, the interference sequence is generated through a multi-layer transposed convolutional network. The transposed convolutional network contains multiple transposed convolutional layers, each layer uses convolution kernels of different sizes, and achieves feature upsampling and reconstruction through layer-by-layer processing. This generation mechanism can effectively simulate the generation process of various types of interference.

[0119] Then, an interference identification module is established. This module is mainly designed based on the attention mechanism and includes two branches: spatial attention and channel attention. The spatial attention branch performs average pooling and maximum pooling operations on the feature map, and then sends the pooling result to the convolution layer for processing, and finally obtains the attention weight of the spatial dimension through the sigmoid function. The channel attention branch performs global average pooling on the feature map, and then uses a multi-layer perceptron for processing, and also obtains the attention weight of the channel dimension through the sigmoid function. The attention features of the two branches are fused with the original features through element-wise multiplication to enhance the important features. Thus, an initial model is established based on the model framework and the interference identification module.

[0120] Then, adversarial training is implemented. A discriminator based on a fully convolutional network is constructed to distinguish between generated samples and real samples, thereby defining the adversarial loss function of the generator and the discriminator, and introducing the reconstruction loss function. During the training process, the generator and the discriminator are continuously optimized in an adversarial manner. The generator tries to generate more realistic samples to deceive the discriminator, while the discriminator strives to improve its ability to distinguish. In addition to the adversarial loss, the reconstruction loss is also introduced to ensure the consistency of the generated results with the input data. These two losses are balanced by weight coefficients to jointly guide the optimization process of the model.

[0121] Therefore, this embodiment constructs an encoder-decoder framework through spatiotemporal alignment features, enhances the ability to capture key interference factors through the attention mechanism, and uses adversarial training to improve the generation performance of the model.

[0122] Based on the above embodiment, in some embodiments, the associated data of the interference factors are introduced into the initial interference model to generate an intervention model, including:

[0123] Determining a cause-and-effect relationship between the interference factor in the powder conveying system;

[0124] Based on the causal relationship, a dynamic causal graph model is constructed, where the dynamic causal graph model is used to describe the temporal dependency relationship between data;

[0125] By using a preset counterfactual reasoning algorithm, an interference impact assessment is performed on the interference factor to obtain an interference assessment result;

[0126] The dynamic causal graph model and the interference assessment result are introduced into the initial interference model to obtain a generated intervention model.

[0127] In this embodiment, it is first necessary to determine the causal relationship in the system. The initial causal graph is constructed by a conditional independence test, which determines the independence between variables by calculating conditional mutual information. In actual operation, the observation data of each variable in the system is first collected, and then the conditional mutual information between the variable pairs is calculated based on these data. If the conditional mutual information value is lower than a preset threshold, it is considered that the pair of variables is independent under given conditions. By testing all possible variable pairs, a preliminary causal network is finally obtained.

[0128] Then, a dynamic causal graph model is constructed to describe the temporal dependencies between variables. The model adopts the form of time-varying structural equations and describes the dependencies between each variable and its parent node set through functional relationships. In the specific implementation, a neural network is used to learn this nonlinear causal mechanism, and the parameters of the network are optimized by the gradient descent method. This method can capture the complex dynamic dependencies between variables.

[0129] Next, the impact of the disturbance is evaluated. Counterfactual reasoning is used to evaluate the impact of the disturbance. First, the intervention operation is defined to simulate the occurrence of the disturbance by changing the value of a specific variable. Then a counterfactual predictor is constructed to predict the response of the system under the intervention condition. By comparing the system state before and after the intervention, the impact of the disturbance can be quantified. This evaluation method can help the system better understand and deal with various types of disturbances.

[0130] Finally, an uncertainty quantification mechanism is introduced. A Bayesian neural network is used to model parameter uncertainty. The parameters of the network are no longer fixed values, but obey a certain probability distribution. Through the Monte Carlo sampling method, the probability distribution of the model prediction results can be obtained to evaluate the uncertainty of the prediction. This mechanism can provide more reliable interference prediction results and help the system make more robust decisions.

[0131] Therefore, through the above steps, a complete generative intervention model is finally obtained. This generative intervention model can not only accurately capture the characteristics of various types of interference, but also evaluate the impact of interference and the uncertainty of prediction, providing a reliable environmental model support for subsequent path planning. Therefore, the causal inference mechanism is introduced on the basis of the initial interference model to achieve accurate modeling of interference factors. This method of combining generative models and causal inference not only ensures the expressive power of the model, but also provides an interpretable causal structure, and realizes the simulation of the powder conveying system and the environment during the powder conveying process, such as environmental interference factors, providing a reliable environmental model for subsequent path planning.

[0132] In some embodiments, the method further comprises:

[0133] Constructing a neural network structure with a weight sharing mechanism, wherein the weight sharing mechanism is used to indicate a weight template in a cyclic shared weight matrix, wherein the weight template is determined by an input of the neural network structure;

[0134] Determining a local connection mode, wherein the local connection mode is used to indicate that an input space of the neural network structure is divided into a plurality of local receptive fields and a local connection matrix is ​​determined;

[0135] Based on the neural network structure and the local connection mode, an equivariance constraint is introduced to obtain an equivariance local neural network.

[0136] In this embodiment, a network structure with weight sharing characteristics is constructed. The core of this structure is a shared weight matrix, and the dimension of the matrix is ​​determined by the input dimension and the output dimension. In the actual implementation, a series of basic weight templates are first designed, which contain different types of feature extraction modes. During the operation of the network, these templates are used cyclically to process the input data, thereby achieving weight sharing. This sharing mechanism not only reduces the number of parameters of the model, but also improves the efficiency of feature extraction. Specifically, when the input data passes through the network, each weight template is reused at different positions, similar to a convolution operation, but with a more flexible sharing method. This design enables the model to efficiently extract local features while keeping the number of parameters within a controllable range.

[0137] Then, a local connection mode is implemented to reduce the complexity of the model. The core idea of ​​this mode is to divide the input space into multiple local receptive fields, and each neuron only establishes connections with the input within its receptive field. In the specific implementation, it is necessary to first determine the position of each neuron, and then determine the range in which it can receive input based on the preset receptive field radius. This local connection method is implemented through a local connection matrix, which specifies the connection relationship between each neuron and the input. During the forward propagation process, the input signal is transmitted through this local connection method, so that each neuron only processes the information within its receptive field. This design greatly reduces the number of connections in the network while retaining sensitivity to local features.

[0138] Then, equal variance constraints are introduced to ensure the stability of the model. This process first requires the network weights to be initialized reasonably, and a specific initialization scheme is used to keep the variance of each layer output within an appropriate range. During the network training process, the variance distribution of each layer output is constrained by adding equal variance regularization terms. In specific implementation, the variance of each layer output is calculated and compared with the preset target variance value, and the difference is added to the loss function as a regularization term. In addition, special variance normalization layers are introduced in the network, which can adaptively adjust the distribution of the data to ensure that the signal maintains an appropriate scale during propagation.

[0139] Specifically, this step trains the material path planning strategy based on the pre-established equivariance local neural network and generative intervention model, including:

[0140] (1) Construct a weight sharing mechanism. Design a shared weight matrix W∈R^{m×n}, where m is the output dimension and n is the input dimension, and implement weight sharing by cyclically using weight templates.

[0141] (2) Implement the local connection mode. Specifically, the input space is divided into multiple local receptive fields, and then the local connection matrix is ​​defined and the forward propagation calculation is modified.

[0142] The equal variance constraint is introduced to ensure the stability of the model.

[0143] Therefore, this embodiment realizes the unification of weight sharing, local connection and equivariance characteristics by constructing a neural network structure with a weight sharing mechanism, provides an efficient and stable model basis for subsequent strategy training, and improves the accuracy of path planning.

[0144] Based on the above embodiments, in some embodiments, the material path planning strategy is trained based on the pre-established equivariance local neural network and the generated intervention model to obtain a path planning strategy model, including:

[0145] constructing an environmental model based on the generative intervention model;

[0146] Based on the equivariance local neural network and the environment model, determining a deep reinforcement learning framework, wherein the deep reinforcement learning framework includes a value network structure, a policy network structure, and a reward function;

[0147] Based on the environment model and the deep reinforcement learning framework, a preset proximal strategy optimization algorithm is used to optimize the decision network;

[0148] Based on the optimized decision network, the path planning strategy model is obtained.

[0149] In this embodiment, an environmental model is constructed based on a generative intervention model. The environmental model needs to define a complete state space and action space. The state space contains state quantities of multiple dimensions such as the position information, speed information, and environmental parameters of the material. The action space defines the operations that the system can perform, mainly including two dimensions: speed adjustment and direction adjustment. When constructing the transfer function, it is necessary to consider the dynamic characteristics of the system and the influence of external interference. Among them, the basic transfer characteristics of the system are described by the physical model, while the influence of external interference is provided by the generative intervention model constructed above. This combination can more completely describe the dynamic characteristics of the system.

[0150] Then, a deep reinforcement learning framework is constructed. The framework contains two core components: the value network and the policy network. The value network is responsible for evaluating the value of different states. It uses the previously constructed equivariance local neural network as a function approximator. The policy network is responsible for generating specific control actions. It uses a similar network structure but independent parameters. In actual training, the two networks continuously improve the quality of decision-making through interaction. The design of the reward function takes into account multiple factors, including path deviation, energy consumption, and constraint violation. These factors are balanced through weight coefficients to jointly guide the optimization direction of the strategy.

[0151] Then, the decision network is optimized. The proximal policy optimization algorithm is used to train the network. This algorithm ensures the stability of training by limiting the policy update step size. In the specific implementation, it is necessary to first collect enough interaction data, including state, action, reward and other information. Then the advantage function is calculated, which reflects the degree of advantage of a certain action relative to the average level. When updating the strategy, a truncated objective function is used to prevent excessive policy changes. In addition to the policy loss, the value function network needs to be optimized at the same time, and the entropy regularization term is used to encourage the exploration of the strategy. In order to improve the efficiency of sample utilization, an experience replay mechanism is introduced to select more valuable experience for learning through priority sampling.

[0152] Finally, the output and application of the strategy. The trained strategy model consists of a deterministic part and a random exploration part. The deterministic part directly outputs the optimal action, while the random exploration part adds appropriate noise based on the optimal action. This design ensures the stability of the strategy while retaining a certain degree of exploration capability. In practical applications, the degree of exploration can be dynamically adjusted according to the state of the system, reducing randomness in scenarios that require stability and increasing exploration in scenarios that require adaptability. This flexible strategy output method enables the system to better balance the needs of stability and adaptability.

[0153] Thus, this embodiment finally obtains a complete path planning strategy optimization framework through the above steps. This framework realizes the effective solution of path planning problems in complex environments by combining the characteristics of equivariance local neural networks and the advantages of deep reinforcement learning. Based on the equivariance local neural network, the strategy optimization is carried out in combination with the generative intervention model, which not only ensures the expressive power and stability of the model, but also realizes the effective adaptation to the complex environment, so that the optimized path planning strategy can be directly applied to the actual powder conveying system control.

[0154] In some embodiments, the step of using the path planning strategy model to generate a powder conveying path includes:

[0155] Based on the path planning strategy model, calculating a speed adjustment amount and a direction adjustment amount for powder conveying;

[0156] The speed adjustment amount and the direction adjustment amount are converted into control instructions acting on the powder conveying system, and the control instructions are used to instruct the powder conveying system to perform powder conveying.

[0157] In this embodiment, the application of the path planning strategy (i.e., the powder conveying path) generated by the path planning strategy model mainly includes three key links: strategy deployment, real-time control, and exception handling. In the strategy deployment link, the trained strategy model must first be deployed to the real-time control system. This process requires reasonable configuration of the system's operating parameters, including sampling frequency, state update cycle, and control instruction delay. Normally, the system sampling frequency is set to 100Hz, the state vector is updated every 10 milliseconds, and the control instruction issuance delay needs to be controlled within 5 milliseconds. The setting of these parameters can ensure that the system has a fast enough response speed.

[0158] In the real-time control link, the system needs to continuously build and update the state vector. This state vector contains data in multiple dimensions, such as the current position of the material, speed information, and environmental parameters. Based on the current state, the strategy model will output the corresponding control action, mainly including the speed adjustment amount and the direction adjustment angle. In order to ensure the smoothness of control, the system sets reasonable range limits for these adjustments, usually limiting the speed adjustment amount to between plus or minus 0.5 meters per second and the direction adjustment angle to between plus or minus 15 degrees. These restrictions can effectively prevent the control action from being too drastic, resulting in system instability.

[0159] Based on the above embodiments, in some embodiments, the method further includes:

[0160] Monitoring the operating status of the powder conveying system;

[0161] When it is detected that the operating state is in an abnormal state, the control instruction is adjusted based on a preset abnormality handling strategy, and the abnormal state includes at least one of material blockage, pressure fluctuation and temperature abnormality.

[0162] In the abnormality handling stage, the system needs to monitor the operating status in real time and respond promptly when abnormal conditions are detected. Common abnormal conditions include material blockage, pressure fluctuations, and temperature abnormalities. The system will formulate corresponding handling strategies for different types of abnormalities. For example, when a blockage warning is detected, the system will take measures to reduce the conveying speed and adjust the conveying angle; when a temperature abnormality occurs, the temperature control system will be activated for compensation.

[0163] This process is illustrated by a specific application example. Assume that in a cement production line, powder materials need to be transported from the grinder to the storage bin. Several key points are set in the entire conveying path, including the starting point (exit of the grinder) at the origin of the coordinates, the end point (entrance of the storage bin) at the coordinates (50, 30, 15), and equipment such as dust collectors and metal detectors are also set in the middle. At the initial stage of system operation, the material starts to move at a horizontal speed of 2.0 meters per second, a vertical speed of 1.2 meters per second, and a lateral speed of 0.5 meters per second. The environmental conditions are a temperature of 25 degrees Celsius, a relative humidity of 45%, and an air pressure of 101.3 kPa.

[0164] As the system runs, when the time reaches 0.1 second, the material moves to a new position (0.23, 0.12, 0.05), and the speed changes slightly, reaching 2.3, 1.3 and 0.5 meters per second in each direction. At this time, the environmental parameters also change slightly, the temperature rises to 25.1 degrees Celsius, the relative humidity drops to 44.8%, and the air pressure drops to 101.2 kPa. The system can smoothly cope with these minor changes and maintain the stability of the conveying process.

[0165] When the system was running for 1.5 seconds, it suddenly encountered a pressure fluctuation interference, and the pressure at position (15, 8, 4) dropped by 2.1 kPa. The system responded immediately, reducing the conveying speed by 0.2 m / s and adjusting the conveying direction by minus 8 degrees to adapt to the pressure change. The response time of this adjustment process was less than 50 milliseconds, which fully demonstrated the system's rapid response capability.

[0166] Through the application of this intelligent control strategy, the overall performance of the powder conveying system has been significantly improved. The conveying efficiency has increased from the original 85% to 94%, the probability of blockage has decreased from 5% to 0.5%, and the energy consumption has decreased from 1.2 kWh per conveying to 0.8 kWh. The system has shown excellent dynamic adaptability and can maintain stable operation under the influence of various interference factors such as pressure fluctuations, temperature changes, and humidity fluctuations. Especially when dealing with pressure fluctuations, the system's response time is controlled within 50 milliseconds, the compensation range for temperature changes reaches plus or minus 2 degrees Celsius, and it can maintain stable operation within the range of plus or minus 10% relative humidity. These performance indicators fully demonstrate the practical value of this intelligent path planning method.

[0167] It should be noted that for the powder conveying path, based on the current system state, the speed adjustment amount and the direction adjustment amount are calculated by the strategy model. For example, the speed adjustment amount ranges from [-0.5m / s, 0.5m / s], and the direction adjustment angle ranges from [-15°, 15°].

[0168] Then, for execution control, the calculated speed adjustment and direction adjustment are converted into specific control instructions and sent to the actuator of the powder conveying system. The sampling period of the system is 10ms, and the delay of sending the control instruction does not exceed 5ms.

[0169] For abnormality handling, the system operating status is monitored in real time. When an abnormal state is detected (such as material blockage, pressure fluctuation, temperature abnormality, etc.), the control instructions are adjusted according to the preset abnormality handling strategy. For example, when a blockage warning is detected (pressure gradient greater than 0.5kPa / m), the handling measures of reducing the speed by 50% and increasing the conveying angle are taken; when a temperature abnormality is detected (temperature change exceeds ±5℃), the auxiliary heating or cooling system is started.

[0170] Therefore, the method provided in the embodiment of the present application can effectively handle various types of dynamic interference in the powder conveying process, realize intelligent path planning and control, and significantly improve the conveying efficiency and stability of the system.

[0171] In order to better illustrate the technical effect of the present application, a specific application example is given below.

[0172] Example: Powder conveying system in a cement production line:

[0173] The basic parameter configuration of the system is as follows:

[0174] Conveying system configuration:

[0175] Starting point (mill exit): coordinates (0, 0, 0);

[0176] End point (storage warehouse entrance): coordinates (50, 30, 15);

[0177] Intermediate equipment: dust collector, coordinates (20, 10, 5); metal detector, coordinates (35, 20, 10);

[0178] Control parameter configuration:

[0179] Sampling period: 10ms;

[0180] Control cycle: 100ms;

[0181] Prediction window: 2s;

[0182] Cache depth: 1000;

[0183] Safety margin: 0.2m;

[0184] During the operation of the system, the initial state (t=0s) is:

[0185] Material position: [0, 0, 0];

[0186] Material speed: [2.0, 1.2, 0.5] m / s;

[0187] Environmental parameters: temperature 25°C, relative humidity 45%, air pressure 101.3kPa;

[0188] When the system is running, it first performs feature extraction and spatiotemporal alignment based on multi-source sensor data and Gaussian edge network. For example, at t=0.1s, the system state is updated to:

[0189] Material position: [0.23, 0.12, 0.05];

[0190] Material speed: [2.3, 1.3, 0.5] m / s;

[0191] Environmental parameters: temperature 25.1°C, relative humidity 44.8%, air pressure 101.2 kPa;

[0192] When a pressure fluctuation disturbance (amplitude -2.1 kPa, position [15, 8, 4]) is detected at t = 1.5 s, the generative intervention model is able to quickly capture the disturbance and make corresponding adjustments through the path planning strategy:

[0193] Speed ​​adjustment: -0.2m / s;

[0194] Direction adjustment angle: -8°;

[0195] Through the application of this application method, the powder conveying system has achieved significant performance improvement in actual operation:

[0196] 1. Path execution performance: average deviation is controlled within ±0.15m, running time is 120s, and energy consumption is reduced to 0.8kWh;

[0197] 2. Improvement of key indicators: transportation efficiency increased from 85% to 94%, congestion probability decreased from 5% to 0.5%, and energy consumption decreased from 1.2kWh to 0.8kWh;

[0198] 3. Dynamic adaptability: The response time to pressure fluctuations is less than 50ms, temperature change compensation is automatically adjusted within the range of ±2°C, and the relative humidity changes within the range of ±10% to maintain stable operation.

[0199] Figure 2 This is a schematic diagram of the structure of an intelligent material path planning device in a powder conveying system provided in an embodiment of the present application. Figure 2 , the intelligent material path planning device 200 in the powder conveying system includes:

[0200] A data acquisition module 201 is used to acquire multi-source sensor data sent by a sensor, wherein the sensor is used to monitor the operation of the powder conveying system, and the multi-source sensor data includes material status data, environmental parameter data, and equipment operation data;

[0201] A feature processing module 202 is used to perform feature extraction and spatiotemporal alignment on the multi-source sensor data based on a pre-established Gaussian edge network to obtain a spatiotemporal alignment feature matrix, wherein the spatiotemporal alignment feature matrix is ​​used to indicate a feature vector at any time point and spatial position;

[0202] An intervention model module 203, used to construct a generated intervention model according to the spatiotemporal alignment feature matrix, wherein the generated intervention model is used to indicate interference factors in the powder conveying process;

[0203] A strategy optimization module 204 is used to train the material path planning strategy based on the pre-established equivariance local neural network and the generated intervention model to obtain a path planning strategy model;

[0204] The control execution module 205 is used to generate a powder conveying path using the path planning strategy model, and the powder conveying path is used to control the powder conveying system to perform powder conveying.

[0205] In some embodiments, the feature processing module 202 includes:

[0206] A standardization processing unit, used for performing standardization processing on the multi-source sensor data, wherein the standardization processing includes calculating the mean and standard deviation of the data and converting the original data into a standard score;

[0207] A feature extraction unit, used for extracting features from the standardized multi-source sensor data based on a preset multi-layer perceptron, wherein the multi-layer perceptron includes a ReLU activation function;

[0208] A Gaussian edge model construction unit, used to construct a Gaussian edge model, wherein the Gaussian edge model is used to indicate that a potential feature representation is obtained by performing agnostic learning on a feature extraction result through a preset edge probability density function;

[0209] The alignment unit is used to perform temporal synchronization and spatial registration on the latent feature representation to obtain a spatiotemporal alignment feature matrix.

[0210] In some embodiments, the alignment unit includes:

[0211] A space-time reference system establishment unit, used to establish a space-time reference system with a sampling period of the powder conveying system as a reference time unit based on the potential feature representation;

[0212] A time synchronization unit, configured to perform time synchronization on the potential feature representation based on the space-time reference system using a preset dynamic time warping algorithm;

[0213] A three-dimensional coordinate system establishing unit, used to establish a three-dimensional coordinate system using the center line of the conveying pipeline of the powder conveying system as the reference axis of the spatial coordinate system;

[0214] A spatial registration unit, configured to perform spatial registration on the potential feature representation based on the three-dimensional coordinate system using a preset spatial interpolation algorithm and a preset adaptive weight strategy, wherein the adaptive weight strategy is used to indicate adjusting the weight of the sensor in the interpolation process according to the distance between the point to be interpolated and the sensor corresponding to the potential feature representation;

[0215] The spatiotemporal alignment feature matrix integration unit is used to integrate the latent feature representation after time synchronization and the latent feature representation after spatial registration into a spatiotemporal alignment feature matrix.

[0216] In some embodiments, the intervention model module 203 includes:

[0217] An interference initial model building unit, used to build an interference initial model based on the spatiotemporal alignment feature matrix;

[0218] The generated intervention model construction unit is used to introduce the associated data of the interference factors into the initial interference model to obtain the generated intervention model.

[0219] Wherein, the initial disturbance model building unit includes:

[0220] An encoder construction unit, used to construct an encoder based on the spatiotemporal alignment feature matrix, wherein the encoder includes a preset bidirectional LSTM network;

[0221] A decoder construction unit, used to generate a network with preset conditions as a decoder, and to construct a model framework based on the encoder and the decoder;

[0222] An initial model building unit, used to build an interference recognition module, and to build an initial model based on the model framework and the interference recognition module, wherein the interference recognition module is used to instruct to obtain a specified interference factor based on a preset attention mechanism;

[0223] A discriminator construction unit, used to construct a discriminator using a preset full convolutional network structure;

[0224] A loss determination unit, used to determine the adversarial loss function and the reconstruction loss function of the discriminator and the generator;

[0225] The adversarial training unit is used to perform adversarial training on the initial model based on the adversarial loss function and the reconstruction loss function to obtain an interfered initial model.

[0226] In some embodiments, generating an intervention model building unit includes:

[0227] a causal relationship determination unit, used to determine the causal relationship of the interference factor in the powder conveying system;

[0228] A dynamic causal graph model construction unit, used to construct a dynamic causal graph model based on the causal relationship, wherein the dynamic causal graph model is used to describe the temporal dependency relationship between data;

[0229] An interference impact assessment unit, used to perform interference impact assessment on the interference factor through a preset counterfactual reasoning algorithm to obtain an interference assessment result;

[0230] The generated intervention model acquisition unit is used to introduce the dynamic causal graph model and the interference assessment result into the interference initial model to obtain the generated intervention model.

[0231] In some embodiments, the apparatus further comprises:

[0232] A network structure construction unit, used to construct a neural network structure with a weight sharing mechanism, wherein the weight sharing mechanism is used to indicate a weight template in a cyclic shared weight matrix, and the weight template is determined by an input of the neural network structure;

[0233] A local connection mode determination unit, used to determine a local connection mode, wherein the local connection mode is used to indicate that an input space of the neural network structure is divided into a plurality of local receptive fields and a local connection matrix is ​​determined;

[0234] The equivariance local neural network construction unit is used to introduce equivariance constraints based on the neural network structure and the local connection mode to obtain an equivariance local neural network.

[0235] In some embodiments, the policy optimization module 204 includes:

[0236] An environmental model building unit, used to build an environmental model based on the generated intervention model;

[0237] A deep reinforcement learning framework construction unit, used to determine a deep reinforcement learning framework based on the equivariance local neural network and the environment model, wherein the deep reinforcement learning framework includes a value network structure, a policy network structure, and a reward function;

[0238] A decision network optimization unit, configured to optimize the decision network using a preset proximal strategy optimization algorithm based on the environment model and the deep reinforcement learning framework;

[0239] The path planning strategy model building unit is used to obtain the path planning strategy model based on the optimized decision network.

[0240] In some embodiments, the control execution module 205 includes:

[0241] An adjustment unit, used for calculating a speed adjustment amount and a direction adjustment amount for powder conveying based on the path planning strategy model;

[0242] The instruction generating unit is used to convert the speed adjustment amount and the direction adjustment amount into a control instruction acting on the powder conveying system, wherein the control instruction is used to instruct the powder conveying system to perform powder conveying.

[0243] In some embodiments, the apparatus further comprises:

[0244] An operating status monitoring unit, used to monitor the operating status of the powder conveying system;

[0245] The abnormality unit is used to adjust the control instruction based on a preset abnormality handling strategy when it is detected that the operating state is in an abnormal state, and the abnormal state includes at least one of material blockage, pressure fluctuation and temperature abnormality.

[0246] The device of the embodiments of the present application can execute the method provided by the embodiments of the present application, and the implementation principles are similar. The actions performed by each module in the device of each embodiment of the present application correspond to the steps in the method of each embodiment of the present application. For the detailed functional description of each module of the device, please refer to the description in the corresponding method shown in the previous text, which will not be repeated here.

[0247] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. For the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0248] The above description is only a preferred embodiment of the present application and is not intended to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application are included in the protection scope of the present application.

[0249] The system of the embodiments of the present disclosure can execute the method provided by the embodiments of the present disclosure, and the implementation principles are similar. The actions performed by each module in the system of each embodiment of the present disclosure correspond to the steps in the method of each embodiment of the present disclosure. For the detailed functional description of each module of the system, please refer to the description in the corresponding method shown in the previous text, which will not be repeated here.

[0250] The above is only an optional implementation method for some implementation scenarios of the present disclosure. It should be pointed out that for ordinary technicians in this technical field, without departing from the technical concept of the scheme of the present disclosure, other similar implementation methods based on the technical ideas of the present disclosure are also within the protection scope of the embodiments of the present disclosure.

Claims

1. An intelligent material path planning method in a powder conveying system, characterized in that: include: Acquire multi-source sensor data sent by a sensor, wherein the sensor is used to monitor the operation of the powder conveying system, and the multi-source sensor data includes material status data, environmental parameter data, and equipment operation data; Based on a pre-established Gaussian edge network, feature extraction and spatiotemporal alignment are performed on the multi-source sensor data to obtain a spatiotemporal alignment feature matrix, where the spatiotemporal alignment feature matrix is ​​used to indicate a feature vector at any time point and spatial position; constructing a generative intervention model according to the spatiotemporal alignment feature matrix, wherein the generative intervention model is used to indicate interference factors in a powder conveying process; Training the material path planning strategy based on the pre-established equivariance local neural network and the generative intervention model to obtain a path planning strategy model; The path planning strategy model is used to generate a powder conveying path, where the powder conveying path is used to control the powder conveying system to perform powder conveying; Wherein, constructing and generating an intervention model according to the spatiotemporal alignment feature matrix includes: Based on the spatiotemporal alignment feature matrix, construct an initial interference model; In the initial interference model, the associated data of the interference factors are introduced to obtain a generated intervention model; The step of constructing an initial interference model based on the spatiotemporal alignment feature matrix includes: Based on the spatiotemporal alignment feature matrix, construct an encoder, wherein the encoder includes a preset bidirectional LSTM network; Using a preset conditional generative network as a decoder, and constructing a model framework based on the encoder and the decoder; Establishing an interference recognition module, and establishing an initial model based on the model framework and the interference recognition module, wherein the interference recognition module is used to instruct to obtain a specified interference factor based on a preset attention mechanism; Use the preset fully convolutional network structure to build the discriminator; Determine the adversarial loss function and the reconstruction loss function of the discriminator and the generator; Based on the adversarial loss function and the reconstruction loss function, performing adversarial training on the initial model to obtain an interference initial model; The associated data of the interference factors are introduced into the initial interference model to generate an intervention model, including: Determining a cause-and-effect relationship between the interference factor in the powder conveying system; Based on the causal relationship, a dynamic causal graph model is constructed, where the dynamic causal graph model is used to describe the temporal dependency relationship between data; By using a preset counterfactual reasoning algorithm, an interference impact assessment is performed on the interference factor to obtain an interference assessment result; The dynamic causal graph model and the interference assessment result are introduced into the initial interference model to obtain a generated intervention model.

2. The intelligent material path planning method in the powder conveying system according to claim 1 is characterized in that: The method of performing feature extraction and spatiotemporal alignment on the multi-source sensor data based on a pre-established Gaussian edge network to obtain a spatiotemporal alignment feature matrix includes: Performing standardization processing on the multi-source sensor data, wherein the standardization processing includes calculating the mean and standard deviation of the data and converting the original data into a standard score; Performing feature extraction on the standardized multi-source sensor data based on a preset multi-layer perceptron, wherein the multi-layer perceptron includes a ReLU activation function; Constructing a Gaussian edge model, wherein the Gaussian edge model is used to indicate that a potential feature representation is obtained by performing agnostic learning on a feature extraction result through a preset edge probability density function; The latent feature representations are temporally synchronized and spatially aligned to obtain a spatiotemporally aligned feature matrix.

3. The intelligent material path planning method in the powder conveying system according to claim 2 is characterized in that: The step of performing temporal synchronization and spatial registration on the latent feature representation to obtain a spatiotemporal alignment feature matrix includes: Based on the potential feature representation, a spatiotemporal reference system is established with a sampling period of the powder conveying system as a reference time unit; Based on the space-time reference system, using a preset dynamic time warping algorithm to perform time synchronization on the potential feature representation; A three-dimensional coordinate system is established by taking the center line of the conveying pipeline of the powder conveying system as the reference axis of the spatial coordinate system; Based on the three-dimensional coordinate system, the potential feature representation is spatially registered using a preset spatial interpolation algorithm and a preset adaptive weight strategy, wherein the adaptive weight strategy is used to indicate adjusting the weight of the sensor in the interpolation process according to the distance between the point to be interpolated and the sensor corresponding to the potential feature representation; The temporally synchronized latent feature representation and the spatially aligned latent feature representation are integrated into a spatiotemporally aligned feature matrix.

4. The intelligent material path planning method in a powder conveying system according to claim 1, characterized in that: The method further comprises: Constructing a neural network structure with a weight sharing mechanism, wherein the weight sharing mechanism is used to indicate a weight template in a cyclic shared weight matrix, wherein the weight template is determined by an input of the neural network structure; Determining a local connection mode, wherein the local connection mode is used to indicate that an input space of the neural network structure is divided into a plurality of local receptive fields and a local connection matrix is ​​determined; Based on the neural network structure and the local connection mode, an equivariance constraint is introduced to obtain an equivariance local neural network.

5. The intelligent material path planning method in a powder conveying system according to claim 4, characterized in that: The material path planning strategy is trained based on the pre-established equivariance local neural network and the generated intervention model to obtain a path planning strategy model, including: constructing an environmental model based on the generative intervention model; Based on the equivariance local neural network and the environment model, determining a deep reinforcement learning framework, wherein the deep reinforcement learning framework includes a value network structure, a policy network structure, and a reward function; Based on the environment model and the deep reinforcement learning framework, a preset proximal strategy optimization algorithm is used to optimize the decision network; Based on the optimized decision network, the path planning strategy model is obtained.

6. The intelligent material path planning method in a powder conveying system according to claim 1, characterized in that: The step of using the path planning strategy model to generate a powder conveying path includes: Based on the path planning strategy model, calculating a speed adjustment amount and a direction adjustment amount for powder conveying; The speed adjustment amount and the direction adjustment amount are converted into control instructions acting on the powder conveying system, and the control instructions are used to instruct the powder conveying system to perform powder conveying.

7. The intelligent material path planning method in a powder conveying system according to claim 6, characterized in that: Also includes: Monitoring the operating status of the powder conveying system; When it is detected that the operating state is in an abnormal state, the control instruction is adjusted based on a preset abnormality handling strategy, and the abnormal state includes at least one of material blockage, pressure fluctuation and temperature abnormality.

8. An intelligent material path planning device in a powder conveying system, characterized in that: include: A data acquisition module, used to acquire multi-source sensor data sent by sensors, wherein the sensors are used to monitor the operation of the powder conveying system, and the multi-source sensor data includes material status data, environmental parameter data, and equipment operation data; A feature processing module, used for performing feature extraction and spatiotemporal alignment on the multi-source sensor data based on a pre-established Gaussian edge network to obtain a spatiotemporal alignment feature matrix, wherein the spatiotemporal alignment feature matrix is ​​used to indicate a feature vector at any time point and spatial position; An intervention model module, used for constructing a generative intervention model according to the spatiotemporal alignment feature matrix, wherein the generative intervention model is used for indicating interference factors in the powder conveying process; A strategy optimization module, used for training a material path planning strategy based on a pre-established equivariance local neural network and the generative intervention model to obtain a path planning strategy model; A control execution module, used for generating a powder conveying path by using the path planning strategy model, wherein the powder conveying path is used for controlling the powder conveying system to perform powder conveying; The intervention model module is used to construct and generate an intervention model according to the spatiotemporal alignment feature matrix, including: Based on the spatiotemporal alignment feature matrix, construct an initial interference model; In the initial interference model, the associated data of the interference factors are introduced to obtain a generated intervention model; The step of constructing an initial interference model based on the spatiotemporal alignment feature matrix includes: Based on the spatiotemporal alignment feature matrix, construct an encoder, wherein the encoder includes a preset bidirectional LSTM network; Using a preset conditional generative network as a decoder, and constructing a model framework based on the encoder and the decoder; Establishing an interference recognition module, and establishing an initial model based on the model framework and the interference recognition module, wherein the interference recognition module is used to instruct to obtain a specified interference factor based on a preset attention mechanism; Use the preset fully convolutional network structure to build the discriminator; Determine the adversarial loss function and the reconstruction loss function of the discriminator and the generator; Based on the adversarial loss function and the reconstruction loss function, performing adversarial training on the initial model to obtain an interference initial model; The associated data of the interference factors are introduced into the initial interference model to generate an intervention model, including: Determining a cause-and-effect relationship between the interference factor in the powder conveying system; Based on the causal relationship, a dynamic causal graph model is constructed, where the dynamic causal graph model is used to describe the temporal dependency relationship between data; By using a preset counterfactual reasoning algorithm, an interference impact assessment is performed on the interference factor to obtain an interference assessment result; The dynamic causal graph model and the interference assessment result are introduced into the initial interference model to obtain a generated intervention model.

Images