Self-adaptive intelligent cleaning control method and system for open wagon

Abstract

The invention relates to the technical field of industrial automation and intelligent control, and discloses a self-adaptive intelligent cleaning control method and system for a train open wagon, and the method comprises the steps: synchronously obtaining a material apparent image and tool-material interaction data when a target point is cleaned; obtaining an effect image after cleaning, calculating a residual rate, and deciding whether to reclean according to the residual rate; extracting and fusing based on the image and mechanical data to generate a multi-modal dynamic state feature vector; taking the feature vector as an input, taking an actual control parameter and a cleaning effect evaluation parameter as an output, constructing and continuously updating a time-varying inverse model on line, and adopting a model prediction control algorithm to solve an optimal control parameter vector of a next target point to be cleaned in a rolling manner; finally, the cleaning device is driven to execute self-adaptive cleaning, and a new round of control circulation is started. The cleaning device can automatically adapt to spatial changes of material characteristics, and intelligent global optimization of the cleaning effect, efficiency and energy consumption is achieved.

Description

Technical Field

[0001] This application relates to the field of industrial automation and intelligent control technology, and in particular to an adaptive intelligent cleaning control method and system for open train wagons. Background Technology

[0002] After transporting bulk materials, open wagons often accumulate residue on their inner walls and bottoms, requiring efficient cleaning to ensure cleanliness and safety for the next transport. Currently, automated cleaning technology is gradually replacing manual labor. Common automated cleaning devices typically operate based on preset programs and fixed parameters (such as the movement trajectory of the robotic arm, the rotation speed and pressure of the brush head).

[0003] However, due to the significant spatial heterogeneity of material characteristics in different carriages, batches, and even different areas of the same carriage—for example, some areas may have highly adhesive and moist materials, while others may be dry, compacted, or loosely piled—fixed-parameter cleaning methods are inadequate in dealing with such complex variations. This directly leads to two adverse consequences: firstly, insufficient cleaning of highly adhesive areas results in material residue and substandard cleanliness; secondly, over-cleaning of loose areas may cause unnecessary wear and tear on cleaning tools and the carriage body, as well as wasted energy.

[0004] In existing technologies, although there have been attempts to introduce single sensors (such as vision or force) for feedback control, they can often only react to changes in a single dimension. They lack the fusion perception and understanding of the multi-dimensional and deep characteristics of the material state, and cannot build dynamic models that can predict the cleaning effect under different control strategies. Therefore, they cannot achieve truly forward-looking and globally optimal adaptive control. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, the embodiments of this application provide an adaptive intelligent cleaning control method for open wagons of trains to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, this application provides an adaptive intelligent cleaning control method for open wagons, comprising: S1. When the cleaning device performs cleaning action on the target point in the open wagon compartment, it simultaneously acquires the material appearance image of the target point before cleaning by the machine vision sensor, and the interaction force data between the cleaning tool and the material by the force sensor. S2. After the cleaning action is completed, acquire the cleaning effect image of the target point collected by the machine vision sensor, and calculate the residual rate based on the cleaning effect image; determine whether the residual rate exceeds the preset residual threshold: if yes, return to step S1 and perform re-cleaning on the target point; if no, proceed to step S3. S3. Based on the material appearance image before cleaning and the interaction force data, extract the material physical property features and contact state features of the target point, and fuse them to generate a multimodal dynamic state feature vector of the target point; S4. Using the multimodal dynamic state feature vector as input, and the control parameter vector of the cleaning action and the residual evaluation parameters extracted from the cleaning effect image as output, a time-varying inverse model describing the mapping relationship between control input, state and cleaning effect is constructed. S5. Based on the time-varying inverse model, a model predictive control algorithm is adopted, with the goal of achieving the best overall cleaning effect within a finite time domain in the future, and the optimal control parameter vector for the next target point to be cleaned is solved in a rolling manner. S6. Output the optimal control parameter vector to the cleaning device to drive the cleaning device to perform adaptive cleaning on the next target point to be cleaned.

[0007] To address the aforementioned problems, this application also provides an adaptive intelligent cleaning control system for open wagons, the system comprising: The multimodal perception module is used to simultaneously acquire, when the cleaning device performs cleaning action on the target point inside the open wagon, the appearance image of the material at the target point before cleaning is collected by the machine vision sensor, and the interaction force data between the cleaning tool and the material is collected by the force sensor. The quality inspection and decision-making module is used to acquire the cleaning effect image of the target point collected by the machine vision sensor after the cleaning action is completed, and calculate the residual rate based on the cleaning effect image; determine whether the residual rate exceeds a preset residual threshold, and decide whether to return to cleaning again or perform subsequent processing. The feature extraction and fusion module is used to extract the material physical property features and contact state features of the target point based on the material appearance image before cleaning and the interaction force data, and fuse them to generate a multimodal dynamic state feature vector of the target point. The time-varying inverse model construction module is used to construct a time-varying inverse model describing the mapping relationship between control input, state, and cleaning effect, using the multimodal dynamic state feature vector as input and the control parameter vector of the cleaning action and the residual evaluation parameters extracted from the cleaning effect image as output. The model predictive control module is used to solve the optimal control parameter vector for the next target point to be cleaned by using the model predictive control algorithm based on the time-varying inverse model, with the goal of achieving the best overall cleaning effect in the future finite time domain. The execution control and loop module is used to output the optimal control parameter vector to the controller of the cleaning device to drive the cleaning device to perform adaptive cleaning on the next target point to be cleaned.

[0008] Compared with the prior art, this application has the following beneficial effects: The adaptive intelligent cleaning control method for open wagons provided by this invention effectively solves the problem of adaptive control in the cleaning process caused by the spatial heterogeneity of material characteristics by constructing a closed-loop system integrating multimodal perception, online learning, and model predictive control. Its primary technical effect lies in achieving full-process intelligentization and high adaptability of the cleaning process. The system synchronously collects visual and mechanical information and fuses it to generate feature vectors that comprehensively characterize the instantaneous state of the material, providing accurate input for intelligent decision-making. Furthermore, through online learning, it continuously updates the time-varying inverse model of the "state-control-effect" mapping relationship, enabling the system to continuously learn and evolve from historical operational experience and adapt to different working conditions. Finally, it uses a model predictive control framework for rolling optimization, dynamically solving customized optimal control parameters for each cleaning point, thereby achieving a globally optimal balance between cleaning effect and resource consumption while ensuring cleaning quality.

[0009] Specifically, the beneficial effects of this method are reflected in the following aspects: First, the multimodal perception and feature fusion mechanism makes the system's judgment on the physical properties of materials more comprehensive and accurate, overcoming the limitations of a single sensor; Second, the use of a neural network with dual output branches to construct a time-varying inverse model online creatively integrates control strategy learning and effect prediction learning, enabling the model to not only recommend actions but also predict results; Third, embedding the data-driven time-varying inverse model into the model predictive control algorithm realizes a forward-looking optimization control based on prediction and combining feedforward and feedback, avoiding the lag and local optimization defects of traditional feedback control; Fourth, through real-time analysis of cleaning effect images and residual rate judgment, a strict quality closed-loop inspection is formed, ensuring that any point that fails to meet the standard will trigger re-cleaning, thus forcibly guaranteeing the final operation quality from the process perspective. Attached Figure Description

[0010] Figure 1 A flowchart illustrating an adaptive intelligent cleaning control method for open wagons provided in an embodiment of this application; Figure 2 A functional block diagram of an adaptive intelligent cleaning control system for open wagons provided in an embodiment of this application; The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0011] It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0012] This application provides an adaptive intelligent cleaning control method for open train wagons. The executing entity of the adaptive intelligent cleaning control method for open train wagons includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application embodiment: a server, a terminal, etc. In other words, the adaptive intelligent cleaning control method for open train wagons can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks, and big data and artificial intelligence platforms.

[0013] Reference Figure 1 The diagram shown is a flowchart illustrating an adaptive intelligent cleaning control method for open wagons provided in an embodiment of this application. In this embodiment, the adaptive intelligent cleaning control method for open wagons includes: S1. When the cleaning device performs cleaning action on the target point in the open wagon compartment, it simultaneously acquires the material appearance image of the target point before cleaning by the machine vision sensor, and the interaction force data between the cleaning tool and the material by the force sensor.

[0014] In the embodiments of this application, step S1 is a fundamental step for realizing multimodal perception and quantitative characterization of the material state at the cleaning target point.

[0015] In this embodiment, the target point refers to a specific local area inside the carriage where the cleaning tool of the cleaning device is acting at the current moment; the machine vision sensor refers to an industrial digital camera installed on the robotic arm of the cleaning device; the force sensor refers to a six-dimensional force / torque sensor installed between the cleaning tool and the end effector of the robotic arm; and the interactive force data refers to a mechanical feature vector that can comprehensively characterize the mechanical properties of the material after a specific signal processing procedure, rather than the original voltage or current signal.

[0016] In some embodiments, the acquisition of interaction force data between the cleaning tool and the material by the force sensor includes: Acquire raw interactive force signals continuously collected by mechanical sensors during the cleaning process; The original interactive force signal is decomposed into force amplitude, force change rate, and force direction change data. Based on the force amplitude, the force change rate, and the force direction change data, a mechanical feature vector characterizing the material adhesion strength and shear resistance is calculated and generated. The mechanical feature vector is used as the interaction force data.

[0017] In this embodiment, step S1 primarily involves the coordinated triggering of hardware and the synchronous acquisition of data. At the same moment the control system of the cleaning device drives the cleaning tool (such as a rotating brush) to begin cleaning a target point, the system sends synchronous trigger signals to the machine vision sensor and the force sensor. The machine vision sensor is triggered the instant before the cleaning tool contacts the material surface, acquiring a high-resolution RGB image, which is defined as the material's appearance image before cleaning. For example, an industrial camera with a resolution of two megapixels is used, capturing images that clearly record the material's accumulation contour, surface texture, and color information. The force sensor continuously acquires raw interactive force signals throughout the cleaning process at a sampling rate much higher than the frequency of mechanical movement. This raw signal is a time-varying six-dimensional data sequence containing force components in three directions and torque components in three directions. For example, the sampling frequency is set to one thousand times per second to ensure that rapid dynamic changes in the force signal can be captured.

[0018] In this embodiment, the processing of mechanical data is the core, aiming to extract low-dimensional features directly related to the physical properties of the material from the raw, high-dimensional time-series signal. The process begins by acquiring a continuous-time raw interaction force signal from a mechanical sensor. Then, a digital signal processing algorithm decomposes this raw signal into three physically meaningful dimensions. The first dimension is the force amplitude, obtained by calculating the instantaneous magnitude of the resultant force vector, which directly reflects the force required for the cleaning tool to overcome the overall resistance of the material. The second dimension is the rate of force change, approximated by differential calculation of the force amplitude, which characterizes the drastic change in force and can distinguish between plastic deformation and brittle fracture behavior of the material. The third dimension is the change in force direction, obtained by calculating the angle between the resultant force vectors at continuous time points, revealing the relative slippage and friction between the tool and the material during the cleaning process.

[0019] In this embodiment, based on the data from the three dimensions mentioned above, a quantitative mechanical feature vector is further calculated and generated. This is accomplished by extracting statistical features from the force amplitude, rate of change, and direction change data within a representative cleaning time window. For example, the system calculates the average and standard deviation of the force amplitude within the time window to describe the fluctuation of the material's average adhesion strength and resistance; calculates the maximum value of the force change rate to capture the instantaneous characteristics of material agglomeration or brittle peeling; and calculates the average value of the force direction change to assess the material's shear resistance and flowability. Combining these statistical features in a predetermined order forms a multidimensional mechanical feature vector. For example, a four-dimensional mechanical feature vector can be represented as an array containing the average force, force standard deviation, maximum rate of change, and average direction change. Finally, this calculated mechanical feature vector is defined as the interactive force data output in step S1, replacing the original, difficult-to-use force signal sequence.

[0020] In this embodiment, the technical effect of this step is to provide the entire adaptive control system with synchronous and multi-angle raw quantitative information about the material state at the target point. By synchronously acquiring visual images and mechanical signals, the system simultaneously grasps the material's apparent morphology and internal mechanical response. This lays a reliable data foundation for solving control problems caused by the uneven spatial distribution of material characteristics. Transforming the raw, high-noise mechanical signals into feature vectors reflecting the material's adhesion and shear resistance characteristics allows subsequent steps to directly utilize these physically meaningful features for decision-making, improving information processing efficiency.

[0021] S2. After the cleaning action is completed, acquire the cleaning effect image of the target point collected by the machine vision sensor, and calculate the residual rate based on the cleaning effect image; determine whether the residual rate exceeds the preset residual threshold: if yes, return to step S1 and perform re-cleaning on the target point; if no, proceed to step S3.

[0022] In the embodiments of this application, step S2 is a key step in realizing the closed-loop inspection and decision-making process for cleaning quality.

[0023] In this embodiment, the cleaning effect image is a frame of digital image captured again by the machine vision sensor after the cleaning action at the current target point is completed, used to intuitively display the final cleaning status of the point; the residual rate is a dimensionless ratio calculated by image analysis, used to quantitatively describe the proportion of visual area occupied by residual material at the target point; the preset residual threshold is a numerical limit preset according to the cleanliness process requirements, used to determine whether the result of a cleaning action meets the standard.

[0024] In some embodiments, after the cleaning action is completed and the cleaning effect image of the target point is acquired by the machine vision sensor, the method further includes: The cleaning effect image is processed to distinguish and quantify the areas of residual material in the image; Based on the quantification results of the residual material area, residual assessment parameters for characterizing the cleanup completion rate are calculated and generated; The residual evaluation parameters are used in the construction of the time-varying inverse model.

[0025] In this embodiment, step S2 begins with the triggering of image acquisition after the cleaning action is completed. When the cleaning device controller detects that the planned cleaning action sequence for the current target point has been completed, it immediately sends an image acquisition command to the machine vision sensor. The machine vision sensor responds to this command and acquires a new image frame, which is defined as the cleaning effect image. For example, the system controls an industrial camera to take a picture immediately after the robotic arm removes the cleaning tool to ensure that the image captures the true state after cleaning.

[0026] In this embodiment, calculating the residual rate based on the cleaning effect image is the core image processing step. To achieve this, the cleaning effect image must first be input into an image processing algorithm unit. This algorithm unit performs image segmentation, distinguishing pixels in the image into regions belonging to residual material and regions belonging to the clean carriage floor. A typical implementation uses an adaptive threshold segmentation algorithm, which automatically determines the optimal segmentation threshold based on the local grayscale distribution of the image, thus adapting to uneven reflections or stains that may exist on the carriage floor. After successfully segmenting the binary image, the system calculates the residual rate through pixel statistics. Specifically, the algorithm counts the total number of pixels marked as residual material and the total number of pixels within the region of interest of the target point. The residual rate is the ratio of these two values. For example, by performing connected component analysis and pixel counting on the segmented binary image, the pixel area occupied by residual material can be obtained, and thus the residual rate can be calculated.

[0027] In this embodiment, the judgment logic and process control are executed based on the comparison between the calculated residual rate and the preset residual threshold. The control system compares the calculated residual rate value with the preset residual threshold stored in memory in real time. If the comparison result shows that the residual rate exceeds the preset residual threshold, the system determines that the cleaning is substandard and generates a control command to return the process to step S1. After returning to S1, the system will initiate a new cleaning action for the same target point, possibly with adjusted parameters (the adjustment stems from the optimization of the subsequent S5 step). If the residual rate does not exceed the preset residual threshold, the system determines that the cleaning is qualified and generates a control command to continue the process to step S3. This judgment mechanism constitutes a direct quality feedback loop, ensuring that no target point is skipped before the cleaning standard is met.

[0028] In the described embodiment, after acquiring the cleaning effect image, a further step of generating residual evaluation parameters is included. This step provides data support for more refined model learning. After processing the cleaning effect image to distinguish residual material areas, the system not only calculates the total area ratio (i.e., the residual rate) but also performs further quantitative analysis. For example, the system identifies and calculates the area of ​​all residual connected regions in the image, finding the area of ​​the largest residual connected region. This parameter reflects whether there are large pieces of scale that have not been cleaned properly. Another example is that the system analyzes the spatial distribution dispersion of residual pixels and calculates a distribution entropy value using the information entropy formula. This value characterizes whether the residue is concentrated in one place or scattered in various locations. Based on these quantitative results, the system calculates and generates a multi-dimensional residual evaluation parameter vector. This vector contains richer cleaning effect information than a single residual rate. Finally, this residual evaluation parameter vector is stored and labeled as the actual output result corresponding to this "state-control" pair. It will be used in the subsequent S4 step to construct and train the time-varying inverse model as a supervision signal for model learning.

[0029] In the above embodiment, the technical effect of this step is to establish an instant quality verification and decision feedback mechanism for the cleaning process. It replaces subjective human judgment with objective image analysis, making the cleaning effect measurable and verifiable. Its judgment logic directly drives whether the process should be "reworked" or "enter the next stage", thereby forcibly ensuring that the final quality of each cleaned point is not lower than the preset standard, fundamentally avoiding the problem of local cleaning failure caused by improper control.

[0030] S3. Based on the material appearance image before cleaning and the interaction force data, extract the material physical property features and contact state features of the target point, and fuse them to generate a multimodal dynamic state feature vector of the target point.

[0031] In this embodiment, step S3 is the core step in realizing a unified and digital representation of the complex working conditions of the target point. It extracts multi-source heterogeneous sensing data into a feature vector that can comprehensively describe the state of the material.

[0032] In this embodiment, the material physical property features refer to the digital features that reflect the inherent properties of the material itself, which are parsed from the material appearance image before cleaning; the contact state features refer to the digital features that reflect the dynamic process of the interaction between the cleaning tool and the material, which are parsed from the interaction force data; and the multimodal dynamic state feature vector is a fixed-dimensional digital state descriptor that combines the above visual features and mechanical features through fusion rules.

[0033] In some embodiments, the step of extracting material physical property features and contact state features of the target point based on the material appearance image before cleaning and the interaction force data, and fusing them to generate a multimodal dynamic state feature vector of the target point, includes: Based on the material appearance image before cleaning, extract the material appearance visual feature vector of the target point; Based on the interaction force data, extract the material mechanics response feature vector of the target point; Based on the material's apparent visual feature vector and the material's mechanical response feature vector, the multimodal dynamic state feature vector is generated by weighted fusion through feature fusion weights.

[0034] In this embodiment, step S3 begins with a parallel feature extraction process of the two types of data obtained in step S1. The system first inputs the material appearance image before cleaning and the interaction force data into two independent feature extraction modules.

[0035] In this embodiment, extracting the visual feature vector of the material appearance based on the pre-cleaning material appearance image is accomplished using a pre-trained deep convolutional neural network model. Specifically, the system scales the RGB image to a fixed size and inputs it into a convolutional neural network such as ResNet-18. The network's feedforward propagation process automatically performs multi-level feature abstraction, from low-level edges to high-level texture semantics. Instead of using the network's final classification output, the system extracts the activation values ​​before its last global pooling layer as the original high-dimensional visual features; for example, this feature might be a one-dimensional vector of length 512. To reduce the dimensionality and match the scale of subsequent mechanical features, a fully connected layer is typically used to linearly transform and reduce the dimensionality of this 512-dimensional feature, ultimately outputting a low-dimensional visual feature vector of the material appearance, for example, reduced to 32 dimensions, resulting in a feature vector where each element represents the intensity of a visual pattern abstracted from the image.

[0036] In this embodiment, extracting the material mechanical response feature vector based on interaction force data is a direct mapping or lightweight processing procedure. The interaction force data itself is already a mechanical feature vector calculated in step S1, for example, a 4-dimensional vector containing statistics such as average force and standard deviation of force. In this step, this vector is directly defined as the material mechanical response feature vector, or it undergoes simple scale normalization according to the model input requirements. For example, each dimension of the 4-dimensional vector is divided by a predetermined reference value to make it fall within a similar numerical range, thus forming the mechanical response feature vector.

[0037] In this embodiment, weighted fusion using feature fusion weights to generate multimodal dynamic state feature vectors is a creative step in achieving modal information integration. This step requires addressing how to combine visual and mechanical feature vectors from different sources, with different physical meanings, and different numerical scales into a unified vector with optimal representational capabilities. First, feature fusion weights need to be determined, which dictate the contribution ratio of each modal feature in the final fused vector. In one embodiment, fixed weights are used; for example, the visual feature fusion weight coefficient is set to α, and the mechanical feature fusion weight coefficient is set to β, where the sum of α and β is a fixed value. The weight assignment can be based on prior knowledge; for example, if historical data analysis indicates that visual features are more important for predicting clearing effects, then α can be assigned a value greater than β. In another more creative embodiment, the weight coefficients are dynamically generated by a small learning network based on the current two feature vectors, thereby achieving adaptive fusion.

[0038] In the specific execution of the weighted fusion operation, the system first multiplies each element of the material's apparent visual feature vector by the visual feature fusion weight coefficient α, and multiplies each element of the material's mechanical response feature vector by the mechanical feature fusion weight coefficient β. Then, the system concatenates these two weighted feature vectors in terms of dimensions. For example, the weighted 32-dimensional visual feature vector and the weighted 4-dimensional mechanical feature vector are joined end-to-end to form a new composite vector of dimension 36. This newly generated composite vector is the aforementioned multimodal dynamic state feature vector, which organically integrates the material's apparent visual attributes and real-time interactive mechanical attributes, constituting a complete digital signature of the current target point's state.

[0039] In this embodiment, the technical effect of this step is to create a core data structure that can comprehensively and unambiguously characterize the dynamic state of a target point. It transforms unstructured image data and relatively abstract mechanical data into structured, machine-processable high-dimensional feature vectors. This multimodal fusion feature vector can more comprehensively reflect the complex characteristics of materials (such as the correspondence between visual dark spots and high force amplitudes in moist clay) than any single-modal feature, laying the foundation for establishing a precise "state-control-effect" mapping model.

[0040] S4. Using the multimodal dynamic state feature vector as input, and the control parameter vector of the cleaning action and the residual evaluation parameters extracted from the cleaning effect image as output, a time-varying inverse model describing the mapping relationship between control input, state, and cleaning effect is constructed.

[0041] In this embodiment, step S4 is a key step in building the core cognitive and decision-making capabilities of the entire adaptive intelligent cleaning system.

[0042] In some embodiments, the step of constructing a time-varying inverse model describing the mapping relationship between control input, state, and cleanup effect, using the multimodal dynamic state feature vector as input and the control parameter vector of the cleanup action and the residual evaluation parameters extracted from the cleanup effect image as output, includes: Obtain the multimodal dynamic state feature vector, the control parameter vector of the cleaning action, and the residue evaluation parameters; Based on the first mapping relationship between the multimodal dynamic state feature vector and the control parameter vector, and the second mapping relationship between the multimodal dynamic state feature vector and the residual evaluation parameter, a comprehensive mapping relationship model is established through an adaptive update mechanism. The integrated mapping model is updated to the current time-varying inverse model for the rolling solution.

[0043] In this embodiment, step S4 is implemented by first retrieving the complete data samples required for model construction from system memory. These data are derived from the execution results of previous steps. The system retrieves the multimodal dynamic state feature vector generated in step S3, which represents the comprehensive state of the target point before cleaning. The system retrieves the control parameter vector that actually drives the cleaning device to perform this cleaning action; this vector is the actual instruction sent to the actuator. The system retrieves the residual evaluation parameters calculated from the cleaning effect image in step S2; this vector quantifies the actual effect of this cleaning action. These three elements together constitute a complete data triplet of "state-action-effect".

[0044] In this embodiment, the most innovative technical core of the entire method is the establishment of a comprehensive mapping relationship model based on two mapping relationships and through an adaptive update mechanism. The purpose of this model is to learn two crucial relationships simultaneously. The first one to be learned is the first mapping relationship between the multimodal dynamic state feature vector and the control parameter vector. This relationship is essentially learning the experience and strategy of "what control action should be taken when seeing material in a certain state". The second one to be learned is the second mapping relationship between the multimodal dynamic state feature vector and the residual evaluation parameters. This relationship is learning the predictive knowledge of "what the expected cleaning effect will be when facing material in a certain state".

[0045] To simultaneously learn these two relationships, this embodiment employs a neural network structure with dual output branches as the foundation for the comprehensive mapping relationship model. Specifically, the number of nodes in the input layer of this neural network equals the dimension of the multimodal dynamic state feature vector. The network has a shared hidden layer used to extract high-order abstract features from the input state features. After the shared hidden layer, the network branches into two independent output branches. The first output branch is dedicated to establishing the first mapping relationship; its output layer number equals the dimension of the control parameter vector, and its output is defined as the predicted value of the control parameters. The second output branch is dedicated to establishing the second mapping relationship; its output layer number equals the dimension of the residual evaluation parameters, and its output is defined as the predicted value of the residual evaluation parameters.

[0046] The model's learning relies on an adaptive update mechanism driven by a composite loss function and an online learning algorithm. The composite loss function is constructed as a weighted sum of a first mapping error and a second mapping error. The first mapping error is the difference between the actual control parameter vector and the predicted control parameters output by the first branch of the neural network, and is typically calculated using the mean squared error (MSE) function. The second mapping error is the difference between the actual residual evaluation parameters and the predicted residual evaluation parameters output by the second branch of the neural network, and is also calculated using the MSE function.

[0047] The core of the adaptive update mechanism lies in employing an online learning algorithm. This algorithm iteratively optimizes the parameters of the neural network model using newly generated data. Whenever a target point completes cleaning and passes the qualification check in step S2, its corresponding "state-action-effect" triplet is added to the training dataset. The learning algorithm, such as mini-batch stochastic gradient descent, calculates the model's prediction error based on the composite loss function L and backpropagates this error to adjust all connection weights and bias parameters in the neural network. This process is continuous and incremental, enabling the model to continuously absorb new operational experience and dynamically adapt to changes in material properties under different conditions (different carriages, batches, or humidity levels). This embodies the "time-varying" characteristic.

[0048] In this embodiment, the comprehensive mapping model after one iteration update is immediately deployed as the time-varying inverse model currently used for rolling solution. The system loads this updated model parameters into the predictive control module, replacing the old model version. This updated time-varying inverse model encapsulates all operational experience and knowledge up to the present, and can output, based on the input state characteristics, the currently considered optimal control parameter suggestions and an evaluation of the expected effect of applying these parameters, thereby providing the latest and most accurate predictive basis for model predictive control in step S5.

[0049] The technical effect of this step is to generate an intelligent kernel that can continuously evolve with the operation process and has the ability to recommend strategies and predict effects. The establishment of the time-varying inverse model enables the system to get rid of its dependence on fixed threshold control rules, and instead learn the implicit mapping relationship between complex and time-varying material characteristics and the optimal cleaning strategy through data-driven methods. Its "time-varying" characteristic ensures that the system can maintain high adaptability and robustness in the long term, effectively solving the core technical problem that traditional control methods are ineffective due to the spatial heterogeneity and temporal variability of materials.

[0050] In some embodiments, the establishment of a comprehensive mapping relationship model based on a first mapping relationship between the multimodal dynamic state feature vector and the control parameter vector, and a second mapping relationship between the multimodal dynamic state feature vector and the residual evaluation parameter, through an adaptive update mechanism, includes: Construct a neural network model with two output branches, wherein the first output branch is used to establish the first mapping relationship to output the predicted value of the control parameter, and the second output branch is used to establish the second mapping relationship to output the predicted value of the residual evaluation parameter; The multimodal dynamic state feature vector is input into the neural network model to obtain the corresponding predicted values ​​of the control parameters and the predicted values ​​of the residual evaluation parameters; A composite loss function is constructed based on the first error between the control parameter vector and the predicted value of the control parameter, and the second error between the residual evaluation parameter and the predicted value of the residual evaluation parameter. An online learning algorithm is used to iteratively optimize and update the parameters of the neural network model based on the composite loss function; The neural network model that has been iteratively updated is used as the mapping model for the synthesis in the current rolling solution cycle.

[0051] In this embodiment, constructing a neural network model with dual output branches is the core hardware architecture for simultaneously learning the first and second mapping relationships. This neural network is defined as a computational graph in software, and its physical implementation relies on the system's central processing unit or a dedicated computing chip. The network includes a shared input layer and several shared hidden layers for receiving and processing the multimodal dynamic state feature vectors of the input. For example, the shared hidden layers can consist of two fully connected layers, each containing dozens of neurons and employing the ReLU activation function to extract high-order abstract features from the input state.

[0052] In this embodiment, at the end of the forward propagation path of the network, the structure branches into two independent, parallel output branches. This is the key design of this embodiment. The first output branch is a fully connected layer with the number of neurons equal to the dimension of the control parameter vector. The activation value of this branch is defined as the control parameter prediction value, and its function is to establish a mapping from state features to control parameters. The second output branch is another fully connected layer with the number of neurons equal to the dimension of the residual evaluation parameters. The activation value of this branch is defined as the residual evaluation parameter prediction value, and its function is to establish a mapping from state features to cleanup effect. This dual-branch structure enables a model to learn both the policy (what to do) and the prediction (what will happen if it is done) simultaneously.

[0053] In this embodiment, the multimodal dynamic state feature vector is input into the neural network model via forward propagation. The system inputs the current multimodal dynamic state feature vector generated in step S3 as a tensor into the input layer of the constructed neural network model. The data propagates layer by layer along the computation graph of the network. After nonlinear transformation through the shared hidden layer, it flows into the first output branch and the second output branch for final linear weighting and transformation. The network ultimately outputs two results simultaneously: the predicted value of the control parameters generated by the first branch and the predicted value of the residual evaluation parameters generated by the second branch.

[0054] In this embodiment, constructing a composite loss function provides a clear, multi-objective learning direction for model parameter optimization. The system first calculates the first error, which is the difference between the actual control parameter vector and the predicted control parameter values ​​output by the neural network. This error is typically calculated using the mean squared error function, and its value reflects the model's current learning bias towards the control strategy. Simultaneously, the system calculates the second error, which is the difference between the actual residual evaluation parameters and the predicted residual evaluation parameters output by the neural network. This error is also calculated using the mean squared error function, and its value reflects the model's current prediction bias towards the cleaning effect.

[0055] The composite loss function L is constructed as a weighted sum of the first error and the second error, and its form is: Where λ1 and λ2 are preset positive weighting coefficients used to reconcile the importance of the two learning objectives; MSE represents the mean square error calculation; U represents the actual control parameter vector; R represents the predicted value of the control parameter; R represents the actual residual assessment parameter. These represent the predicted values ​​of residual assessment parameters. For example, λ1 can be set to 1.0 and λ2 to 0.8, which means that in parameter updates, the deviation of the corrective control strategy is given a slightly higher priority than the predicted deviation of the corrective effect.

[0056] In the embodiments of this application, the key mechanism for realizing the time-varying characteristics of the model is to use an online learning algorithm for iterative optimization and updating based on the composite loss function. The online learning algorithm specifically refers to an optimization algorithm that can immediately update the model parameters using newly arrived single or small batches of data, such as stochastic gradient descent or its variants (such as the Adam algorithm). After obtaining a new "state-control-effect" data triple each time, the system executes a learning loop. This loop first performs forward propagation to obtain the predicted value and calculates the composite loss L. Subsequently, the algorithm calculates the gradient of the composite loss L with respect to each trainable parameter (weight and bias) in the neural network model through backpropagation. Finally, the algorithm makes a one-time small adjustment to all these model parameters based on the calculated gradient direction and learning rate parameter so that the composite loss L tends to decrease. This optimization process is iterative and continuous, so that the internal mapping relationship of the neural network model can continuously evolve and adjust with the new experience accumulated during the operation.

[0057] In this embodiment, using the iteratively updated neural network model as the comprehensive mapping model for the current rolling solution cycle is a model deployment action. After one or more online learning iterations, the system saves the latest parameter set of the neural network model. This parameterized model, carrying the latest learning experience, is immediately placed into the prediction module of the model predictive controller, replacing the old model version. At this point, the model officially becomes the comprehensive mapping model for the current and next rolling solution cycle, providing immediate and up-to-date computational basis for predicting the cleaning effect under different control parameters in step S5.

[0058] S5. Based on the time-varying inverse model, a model predictive control algorithm is adopted, with the goal of achieving the best overall cleaning effect within a finite time domain in the future, and the optimal control parameter vector for the next target point to be cleaned is solved in a rolling manner.

[0059] In this embodiment, step S5 is a step of making forward-looking optimization decisions using a prediction model obtained through online learning.

[0060] In this embodiment, the future finite time domain is a forward prediction time range or step range set in the model predictive control algorithm; the optimal overall cleanup effect is the best state of comprehensive performance indicators that is expected to be achieved within the set future finite time domain; the optimal control parameter vector is a set of control command parameters that are considered to achieve the best expected cleanup effect at the next target point to be cleaned, obtained through optimization.

[0061] In some embodiments, the step of using a model predictive control algorithm based on the time-varying inverse model, with the goal of achieving the optimal overall cleanup effect within a future finite time domain, and continuously solving for the optimal control parameter vector for the next target point to be cleaned, includes: Based on the multimodal dynamic state feature vector of the current target point and the current time-varying inverse model, predict the expected cleaning effect when applying different control parameter vectors to the next target point to be cleaned; With the goal of achieving the optimal overall cleanup effect within the future finite time domain, the different control parameter vectors are evaluated and optimized within the framework of the model predictive control algorithm, and the optimal control parameter vector is determined from the solution results.

[0062] In this embodiment, step S5 begins with the prediction calculation of the expected cleaning effect of the next target point to be cleaned. The system first obtains the multimodal dynamic state feature vector generated by step S3, which represents the state of the currently cleaned target point. Due to the continuity of materials at adjacent points in the carriage, the system uses the current point's state feature vector as an estimate of the state of the next target point to be cleaned. The system simultaneously loads the time-varying inverse model updated and provided by step S4. The prediction process is completed by pairing different candidate control parameter vectors with this estimated state feature vector and inputting them into the time-varying inverse model. Specifically, the system internally generates or samples multiple candidate control parameter vectors within the feasible region, for example, generating dozens of different combinations of pressure, rotation speed, and moving speed through Latin hypercube sampling.

[0063] In this embodiment, for each candidate control parameter vector, the system inputs it along with the estimated next target point state feature vector into a time-varying inverse model. The second output branch of the time-varying inverse model is invoked, and its output is the predicted residual evaluation parameter when the candidate control parameter is applied under the assumed state. This output is defined as the expected cleanup effect. For example, for a candidate control parameter vector, the model may output a lower predicted residual rate and a smaller predicted maximum residual area, indicating a better expected cleanup effect.

[0064] In the embodiments of this application, the evaluation and optimization solution aimed at achieving the optimal overall cleanup effect within a finite future time domain is the core of the model predictive control algorithm. This application simplifies the finite future time domain to the next target point that needs to be decided, i.e., the prediction time domain is one step, and the optimal overall cleanup effect is specifically quantified and defined through a cost function.

[0065] The cost function J maps the expected cleanup effect (i.e., the predicted value of the residual evaluation parameters) predicted by the time-varying inverse model to a scalar cost. A smaller cost indicates a better effect. The construction of the cost function includes requirements for the cleanup effect and constraints on the control actions themselves. For example, a typical cost function form is... ,in This function predicts the residual rate. P and ω are the pressure and speed components in the candidate control parameter vector, and γ is a positive weighting coefficient. The purpose of this function is to pursue high cleanliness (minimize...). At the same time, it punishes excessive control energy consumption (minimizes) This achieves a balance between cleaning effectiveness and energy consumption.

[0066] Within the framework of model predictive control algorithms, the optimization process involves searching for the candidate control parameter vector that minimizes the aforementioned cost function J within the feasible region of the control parameters. The feasible region of the control parameters is pre-set by the physical limits and process safety requirements of the cleaning device. For example, the brush head pressure is limited to a minimum and a maximum value, and the rotation speed also has its upper and lower limits.

[0067] The system employs numerical optimization algorithms to search within this multidimensional bounded feasible region. One implementation is the particle swarm optimization algorithm, which initializes a swarm of "particles" (i.e., candidate solutions), each representing a control parameter vector. Particles move and update within the solution space based on their individual and swarm history of optimal solutions, eventually converging to the solution that approximately minimizes the cost function J. Another implementation is gradient descent. If the model is differentiable, the gradient of the cost J with respect to the control parameters can be calculated, and the optimal solution can be iteratively searched along the gradient descent direction. The result of the optimization process is a specific control parameter vector, which is determined as the optimal control parameter vector under the current prediction model and optimization objective.

[0068] The technical effect of this step is to generate a forward-looking, globally optimized control strategy. It uses a time-varying inverse model as a prediction tool to simulate the possible results of different control actions in the future, and actively selects the best overall solution through optimization algorithms, rather than simple reactive control based on the current state. This effectively solves the suboptimal control problem that is prone to occur when fixed parameter control or feedback control without prediction is faced with complex and changing materials, and achieves a balance between cleaning effect and resource consumption.

[0069] In some embodiments, predicting the expected cleaning effect when applying different control parameter vectors to the next target point based on the multimodal dynamic state feature vector of the current target point and the current time-varying inverse model includes: Based on the multimodal dynamic state feature vector of the current target point, determine the estimated state feature vector of the next target point to be cleared; The estimated state feature vector and multiple candidate control parameter vectors are respectively input into the current time-varying inverse model, and the time-varying inverse model outputs the expected cleaning effect corresponding to each candidate control parameter vector; The expected cleanup effect corresponding to each candidate control parameter vector is used as the evaluation basis for the optimization solution in the model predictive control algorithm.

[0070] In this embodiment, determining the estimated state feature vector of the next target point to be cleaned is the first step in making forward-looking predictions. This step requires making reasonable inferences about the unknown state of the next target point based on the known state of the current target point. One basic implementation is to assume that the state of the materials in the carriage has spatial continuity, and therefore directly use the multimodal dynamic state feature vector of the current target point as the estimated state feature vector of the next target point to be cleaned. Another more adaptive implementation is to introduce a simple state transition function, which fine-tunes the current state vector based on the state change trends of historical continuous points to generate an estimated vector. For example, the system can record the state vectors of the most recently cleaned points, calculate their change gradients, and extrapolate these gradients to the current state vector to obtain an estimated state feature vector that better reflects the spatial change trend.

[0071] In this embodiment, inputting the estimated state feature vector and multiple candidate control parameter vectors into the time-varying inverse model is a process of batch simulation. First, it is necessary to generate a set of multiple candidate control parameter vectors covering the feasible domain of the control parameters. This can be accomplished by designing a spatial sampling algorithm, such as the Latin hypercube sampling method. This method can efficiently generate a series of uniformly distributed and non-repeating parameter combinations in a multidimensional parameter space. Each sampled candidate control parameter vector represents a possible control strategy to be applied to the cleaning device. Subsequently, the system pairs the same estimated state feature vector with each candidate control parameter vector to form multiple sets of "state-control" input pairs. These input pairs are sequentially fed into the latest time-varying inverse model. After receiving each set of inputs, the second output branch of the time-varying inverse model is activated and calculated, finally outputting the corresponding predicted value of the residual evaluation parameter. This value is defined as the expected cleaning effect under the candidate control parameter vector.

[0072] In this embodiment, using the expected cleanup effect corresponding to each candidate control parameter vector as the evaluation basis for optimization is the link between prediction and decision-making. The optimization solver in the model predictive control algorithm needs a clear evaluation standard to compare the merits of different candidate schemes. The system establishes a record for each candidate control parameter vector. This record not only contains the vector itself but also associates it with the expected cleanup effect output by the time-varying inverse model. This expected cleanup effect is a vector that includes multiple quantitative indicators such as predicted residual rate and predicted maximum residual area. Based on the pre-set overall optimal cleanup effect target, the optimization solver comprehensively calculates these multi-dimensional prediction effect indicators into a scalar cost function value. Specifically, the solver calls a cost function calculation module. The input of this module is the expected cleanup effect vector and the corresponding candidate control parameter vector itself, and the output is a scalar cost. Each candidate control parameter vector obtains a specific and comparable cost evaluation value through the prediction of the time-varying inverse model and the calculation of the cost function. This evaluation value is the direct basis for the optimization solver to evaluate and optimize.

[0073] The technical effect of this step is to provide a key and efficient predictive evaluation capability for model predictive control. It uses time-varying inverse models to quickly predict the future results of various control strategies through batch simulations, and finally quantifies the complex cleanup effect prediction into comparable optimization target values. This makes the optimization solution process no longer a blind search, but a directional and evidence-based intelligent optimization guided by rich predictive information.

[0074] S6. Output the optimal control parameter vector to the cleaning device to drive the cleaning device to perform adaptive cleaning on the next target point to be cleaned.

[0075] In the embodiments of this application, step S6 is the final step in realizing the transformation of intelligent decision-making into physical execution and completing the adaptive control closed loop.

[0076] In this embodiment, the cleaning device controller is a programmable logic controller or motion control card deployed locally on the cleaning device to directly drive the movement of each execution unit; the execution mechanism is a combination of physical components that receive controller instructions and generate mechanical movement.

[0077] In some embodiments, outputting the optimal control parameter vector to the cleaning device to drive the cleaning device to perform adaptive cleaning on the next target point to be cleaned includes: The optimal control parameter vector is sent to the controller of the cleaning device; The controller drives the actuator of the cleaning device to perform a cleaning operation on the next target point to be cleaned, based on the optimal control parameter vector. After the cleaning operation is completed, the next target point to be cleaned is taken as the new current target point, and the process returns to step S1 to start a new round of adaptive cleaning control loop.

[0078] In this embodiment, sending the optimal control parameter vector to the controller of the cleaning device is a communication process that completes the handover of control commands. The central decision-making system packages the optimal control parameter vector obtained in step S5 into a data frame of a specific format through an industrial communication network. For example, it sends out an array containing brush head pressure setpoints, rotation speed setpoints, and translation speed setpoints via Ethernet or fieldbus protocol. The cleaning device controller continuously monitors the network through its communication interface, and after receiving the data frame, it parses and verifies it to confirm that it contains complete and valid control parameters.

[0079] In this embodiment, the controller driving the actuator to perform the cleaning operation based on the optimal control parameter vector is the specific physical control process. The control program inside the controller generates corresponding low-level control signals based on the received parameter settings. For the brush head pressure setting, the controller drives the hydraulic or pneumatic actuator through an analog output module or a dedicated pressure control valve, so that the cleaning tool applies precise pressure to the material surface. For the rotation speed setting, the controller drives the servo motor or frequency converter through a pulse sequence or analog signal, so that the cleaning tool reaches the set rotation speed. For the translation speed setting, the controller plans the path through the motion control module and drives the linear module or articulated arm, so that the cleaning tool moves to the next target point to be cleaned at the set speed and covers the area. All these sub-actions are performed synchronously under the coordination of the controller, together constituting a complete, parameterized cleaning operation.

[0080] In this embodiment, the logical condition for triggering a new round of adaptive loop is to set the next target point to be cleaned as the new current target point after the cleaning operation is completed and return to step S1. The controller or the upper-level monitoring system confirms that the cleaning operation for the current "next target point to be cleaned" has been completed as planned by judging through position sensors or program stepping. The system updates the status flag in memory and re-marks the point that has just completed the cleaning operation as the current target point. This flag update is an automatic trigger condition. Once this condition is met, the system process immediately and automatically jumps back to the starting point of step S1. This means that the system will again synchronously trigger the machine vision sensor and the force sensor for this new current target point to collect multimodal data before cleaning, thereby starting a brand-new "perception-modeling-decision-execution-evaluation" adaptive control loop.

[0081] The technical effect of this step is to accurately translate the intelligent and customized decision-making schemes formed in the previous steps into actual physical actions, and to ensure the continuous and automatic operation of the entire adaptive process. It constitutes a bridge from virtual decision-making to real-world application, allowing the optimal strategy calculated for the spatial heterogeneity of material characteristics to be verified in the field. Its automatic return closed-loop design ensures that the system can continuously and proactively adapt to the different conditions of each new point in the carriage, realizing adaptive intelligent cleaning without human intervention throughout the entire process.

[0082] like Figure 2 The diagram shown is a functional block diagram of an adaptive intelligent cleaning control system for open wagons provided in an embodiment of this application.

[0083] The adaptive intelligent cleaning control system 100 for open wagons described in this application can be installed in an electronic device. Depending on the functions implemented, the adaptive intelligent cleaning control system 100 for open wagons may include a multimodal perception module 101, a quality inspection and decision-making module 102, a feature extraction and fusion module 103, a time-varying inverse model construction module 104, a model prediction control module 105, and an execution control and loop module 106. The module described in this application can also be called a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, and which are stored in the memory of the electronic device.

[0084] In this embodiment, the functions of each module / unit are as follows: The multimodal perception module 101 is used to simultaneously acquire, by means of a machine vision sensor, an image of the material appearance of the target point before cleaning, and by means of a force sensor, interaction force data between the cleaning tool and the material when the cleaning device performs cleaning action on the target point inside the open wagon compartment. The quality inspection and decision module 102 is used to acquire the cleaning effect image of the target point collected by the machine vision sensor after the cleaning action is completed, and calculate the residual rate based on the cleaning effect image; determine whether the residual rate exceeds a preset residual threshold, and decide whether to return to cleaning again or perform subsequent processing. The feature extraction and fusion module 103 is used to extract the material physical property features and contact state features of the target point based on the material appearance image before cleaning and the interaction force data, and fuse them to generate a multimodal dynamic state feature vector of the target point. The time-varying inverse model construction module 104 is used to construct a time-varying inverse model describing the mapping relationship between control input, state and cleaning effect by taking the multimodal dynamic state feature vector as input and the control parameter vector of the cleaning action and the residual evaluation parameters extracted from the cleaning effect image as output. The model predictive control module 105 is used to, based on the time-varying inverse model, employ a model predictive control algorithm, with the goal of achieving the optimal overall cleaning effect within a finite future time domain, and to continuously solve for the optimal control parameter vector for the next target point to be cleaned. The execution control and loop module 106 is used to output the optimal control parameter vector to the controller of the cleaning device to drive the cleaning device to perform adaptive cleaning on the next target point to be cleaned. In the several embodiments provided in this application, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for example, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0085] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0086] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0087] It will be apparent to those skilled in the art that this application is not limited to the details of the exemplary embodiments described above, and that this application can be implemented in other specific forms without departing from the spirit or essential characteristics of this application.

[0088] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application and are not intended to limit it. Although this application has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of this application without departing from the spirit and scope of the technical solutions of this application.

Claims

1. A self-adaptive intelligent cleaning control method for a train gondola, characterized in that, The method comprises: S1, synchronously acquiring a pre-cleaning material appearance image of a target point in a gondola car collected by a machine vision sensor and interaction force data between a cleaning tool and the material collected by a mechanical sensor when a cleaning device performs a cleaning action on the target point in the gondola car; S2, after the cleaning action is completed, acquiring a cleaning effect image of the target point collected by the machine vision sensor, and calculating a residual rate based on the cleaning effect image; judging whether the residual rate exceeds a preset residual threshold; if yes, returning to step S1 to perform re-cleaning on the target point; if no, performing step S3; S3, based on the pre-cleaning material appearance image and the interaction force data, extracting material physical property feature and contact state feature of the target point, and fusing to generate a multi-modal dynamic state feature vector of the target point; S4, taking the multi-modal dynamic state feature vector as input, taking a control parameter vector of the cleaning action and a residual evaluation parameter extracted from the cleaning effect image as output, and constructing a time-varying inverse model describing the control input-state-cleaning effect mapping relationship; S5, based on the time-varying inverse model, using a model predictive control algorithm to take the overall cleaning effect optimal in a future finite time domain as a control target, and rolling to solve an optimal control parameter vector for a next target point to be cleaned; S6, outputting the optimal control parameter vector to the cleaning device to drive the cleaning device to perform adaptive cleaning on the next target point to be cleaned.

2. The adaptive intelligent cleaning control method of a railcar open car as claimed in claim 1, wherein, The interaction force data collected by the mechanical sensor between the cleaning tool and the material comprises: acquiring original interaction force signals continuously collected by the mechanical sensor during the cleaning process; decomposing the original interaction force signals into force amplitude, force change rate and force direction change data; based on the force amplitude, the force change rate and the force direction change data, calculating and generating a mechanical feature vector representing the material adhesion strength and shear resistance characteristics; taking the mechanical feature vector as the interaction force data.

3. The adaptive smart cleaning control method of a railcar open car as claimed in claim 1, wherein, After the cleaning action is completed, after acquiring the cleaning effect image of the target point collected by the machine vision sensor, the method further comprises: processing the cleaning effect image to distinguish and quantify the residual material area in the image; based on the quantification result of the residual material area, calculating and generating a residual evaluation parameter for representing the cleaning completion degree; the residual evaluation parameter is used for constructing the time-varying inverse model.

4. The method of adaptive intelligent cleaning control of a railcar gondola of claim 1, wherein, The method comprises: based on the pre-cleaning material appearance image, extracting a material appearance visual feature vector of the target point; based on the interaction force data, extracting a material mechanical response feature vector of the target point; based on the material appearance visual feature vector and the material mechanical response feature vector, weighted fusion is performed through feature fusion weight to generate the multi-modal dynamic state feature vector.

5. The adaptive smart cleaning control method of a railcar open car as claimed in claim 1, wherein, The multi-modal dynamic state feature vector is taken as input, and the control parameter vector of the cleaning action and the residual evaluation parameter extracted from the cleaning effect image are taken as output, a time-varying inverse model describing the control input-state-cleaning effect mapping relationship is constructed, including: Obtaining the multi-modal dynamic state feature vector, the control parameter vector of the cleaning action and the residual evaluation parameter; Based on the first mapping relationship between the multi-modal dynamic state feature vector and the control parameter vector, and the second mapping relationship between the multi-modal dynamic state feature vector and the residual evaluation parameter, a comprehensive mapping relationship model is established through an adaptive updating mechanism; The comprehensive mapping relationship model is updated to the current time-varying inverse model for the rolling solution.

6. The adaptive smart cleaning control method of a railcar open car as claimed in claim 5, wherein, The first mapping relationship between the multi-modal dynamic state feature vector and the control parameter vector, and the second mapping relationship between the multi-modal dynamic state feature vector and the residual evaluation parameter are established through an adaptive updating mechanism, including: A neural network model with double output branches is constructed, wherein the first output branch is used to establish the first mapping relationship to output the control parameter prediction value, and the second output branch is used to establish the second mapping relationship to output the residual evaluation parameter prediction value; The multi-modal dynamic state feature vector is input into the neural network model to obtain the corresponding control parameter prediction value and residual evaluation parameter prediction value; Based on the first error between the control parameter vector and the control parameter prediction value, and the second error between the residual evaluation parameter and the residual evaluation parameter prediction value, a composite loss function is constructed; An online learning algorithm is used to iteratively optimize and update the parameters of the neural network model based on the composite loss function; The neural network model after iterative update is used as the comprehensive mapping relationship model for the current rolling solution period.

7. The adaptive smart cleaning control method of a railcar open car as claimed in claim 1, wherein, Based on the time-varying inverse model, a model predictive control algorithm is used to take the overall cleaning effect optimal in the future finite time domain as the control target, and the optimal control parameter vector for the next target point to be cleaned is solved in a rolling manner, including: Based on the multi-modal dynamic state feature vector of the current target point and the current time-varying inverse model, the expected cleaning effect when different control parameter vectors are applied to the next target point to be cleaned is predicted; Taking the overall cleaning effect optimal in the future finite time domain as the target, the different control parameter vectors are evaluated and optimized under the framework of the model predictive control algorithm, and the optimal control parameter vector is determined from the solution result.

8. The adaptive smart cleaning control method of a railcar open car as claimed in claim 7, wherein, Based on the multi-modal dynamic state feature vector of the current target point and the current time-varying inverse model, the expected cleaning effect when different control parameter vectors are applied to the next target point to be cleaned is predicted, including: Based on the multi-modal dynamic state feature vector of the current target point, the estimated state feature vector of the next target point to be cleaned is determined; inputting the estimated state feature vector and a plurality of candidate control parameter vectors into a current time-varying inverse model respectively, outputting an expected cleaning effect corresponding to each candidate control parameter vector by the time-varying inverse model; taking the expected cleaning effect corresponding to each candidate control parameter vector as an evaluation basis for optimization and solution in the model predictive control algorithm.

9. The adaptive smart cleaning control method of a railcar open car as claimed in claim 1, wherein, The outputting of the optimal control parameter vector to the cleaning device to drive the cleaning device to perform adaptive cleaning on the next target point to be cleaned comprises: sending the optimal control parameter vector to a controller of the cleaning device; driving an execution mechanism of the cleaning device to perform a cleaning operation on the next target point to be cleaned according to the optimal control parameter vector; after the cleaning operation is completed, taking the next target point to be cleaned as a new current target point, and returning to step S1 to start a new round of adaptive cleaning control cycle.

10. An adaptive intelligent cleaning control system for a railway gondola car for implementing the adaptive intelligent cleaning control method for a railway gondola car according to any one of claims 1 to 9, characterized in that, The system comprises: a multi-modal perception module configured to synchronously acquire a pre-cleaning material appearance image of a target point in a gondola car collected by a machine vision sensor and interaction force data between a cleaning tool and the material collected by a mechanical sensor when the cleaning device performs a cleaning action on the target point; a quality inspection and decision module configured to acquire a cleaning effect image of the target point collected by the machine vision sensor after the cleaning action is completed, calculate a residual rate based on the cleaning effect image, and determine whether to return to re-cleaning or perform subsequent processing by judging whether the residual rate exceeds a preset residual threshold; a feature extraction and fusion module configured to extract material physical property features and contact state features of the target point based on the pre-cleaning material appearance image and the interaction force data, and fuse to generate a multi-modal dynamic state feature vector of the target point; a time-varying inverse model construction module configured to take the multi-modal dynamic state feature vector as input, and take a control parameter vector of the cleaning action and a residual evaluation parameter extracted from the cleaning effect image as output, to construct a time-varying inverse model describing the mapping relationship between control input, state and cleaning effect; a model predictive control module configured to take the time-varying inverse model as a basis, adopt a model predictive control algorithm, and solve an optimal control parameter vector for a next target point to be cleaned with the optimal overall cleaning effect in a future finite time domain as a control target; an execution control and cycle module configured to output the optimal control parameter vector to a controller of the cleaning device to drive the cleaning device to perform adaptive cleaning on the next target point to be cleaned. an execution control and cycle module configured to output the optimal control parameter vector to a controller of the cleaning device to drive the cleaning device to perform adaptive cleaning on the next target point to be cleaned.

Images