Intelligent cleaning operation method and system for railway sideboards in harbor district based on AI technology

Abstract

The invention discloses a harbor area train sideboard intelligent cleaning operation method and system based on the AI technology. The method comprises the steps that S1, scene environment parameters, material accumulation parameters and sideboard parameters are obtained; s2, extracting an environment risk coefficient, a material accumulation risk coefficient and a cleaning difficulty coefficient; s3, building a comprehensive cleaning model and calculating a comprehensive cleaning index; s4, setting a cleaning threshold value to evaluate a comprehensive cleaning index, and determining a cleaning mode and cleaning parameters; s5, collecting sideboard surface data before and after cleaning, and adjusting weight distribution of the comprehensive cleaning model according to a neural network learning technology until a minimum cleaning error; according to the method, multi-dimensional parameter values such as environment parameters, sideboard surface parameters and material accumulation parameters are collected in real time through a system, an environment risk coefficient, a material accumulation risk coefficient and a cleaning difficulty coefficient are extracted, accurate quantitative evaluation is achieved, the pertinence and accuracy of cleaning operation are improved, a comprehensive cleaning model is updated in a self-adaptive mode based on the AI learning technology, and the cleaning efficiency is improved. The system adaptability is improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent cleaning technology, and specifically to an AI-based intelligent cleaning method and system for the sides of port trains. Background Art

[0002] With the rapid development of global trade and the continued growth of port logistics throughput, the cleaning efficiency of train loading bays, core facilities for bulk cargo transshipment, directly impacts the safety of loading and unloading operations and the service life of transportation equipment. Traditional cleaning methods rely on manual labor and are unable to cope with the complex and changing characteristics of accumulated materials, such as adhesion, particle size differences, and dynamic environmental factors such as temperature, humidity, and dust concentration. This leads to unstable cleaning results and waste of resources.

[0003] Currently, smart cleaning technologies primarily focus on automated cleaning using single sensor monitoring or fixed program control. For example, some systems use lidar or visual sensors to identify accumulated material or dirt, and then use preset programs to drive cleaning equipment for cleaning. Other systems utilize historical data storage and simple model analysis to optimize cleaning parameters. Consequently, the application of existing smart cleaning technologies to train side cleaning equipment still faces significant limitations. These limitations include a single perception dimension of the cleaning environment and insufficient comprehensive evaluation capabilities for multimodal data, making it difficult to comprehensively assess the risk of accumulated material, environmental interference, and cleaning difficulty. Cleaning decisions rely on empirical thresholds, failing to accurately quantify the combined impact of the environment and accumulated material. Cleaning models are simple and static, unable to dynamically adjust parameters based on real-time operating conditions. This can lead to over- or under-cleaning in complex operating conditions. Path planning algorithms fail to incorporate the risk distribution of accumulated material, making them unsuitable for loading bay cleaning equipment. This makes it difficult to prioritize areas at high risk of accumulated material, resulting in low cleaning efficiency. Furthermore, the lack of a closed-loop feedback mechanism prevents online iterative optimization of model weights, making the system difficult to adapt to long-term dynamic factors such as side material aging and changes in accumulated material properties.

[0004] Therefore, it is particularly necessary to develop an AI-based intelligent cleaning method and system for port train carriages to solve the above problems. Summary of the Invention

[0005] The purpose of the present invention is to overcome the defects of the prior art and provide an AI-based intelligent cleaning method and system for the side of railway carriages in port areas to solve the problems raised in the above-mentioned background technology.

[0006] In order to achieve the above object, the technical solution of the present invention is as follows:

[0007] An AI-based intelligent cleaning method for railcar sides of port trains comprises the following steps: S1, obtaining environmental parameters, material accumulation parameters and railcar side parameters of a cleaning scene, identifying material accumulation types and train railcar side types, wherein environmental parameters include temperature, humidity and dust concentration, material accumulation parameters include particle size and adhesion of the material, and railcar side parameters include three-dimensional profile data, inclination angle, surface contact pressure and real-time surface roughness of the railcar side; S2, pre-processing environmental parameters, material accumulation parameters and railcar side parameters, performing multi-factor grouping fusion and weight calculation, respectively building an environmental risk model, a material accumulation risk model and a cleaning difficulty model, generating an environmental risk coefficient, a material accumulation risk coefficient and a cleaning difficulty coefficient; S3, building a comprehensive cleaning model, importing the environmental risk coefficient, the material accumulation risk coefficient, the cleaning difficulty coefficient and the material accumulation particle size and assigning weights, and calculating a comprehensive cleaning index; S4, setting a cleaning threshold to evaluate the comprehensive cleaning index, and determining the cleaning mode and cleaning parameters; S5, collecting railcar side surface data before and after cleaning, calculating the cleaning error, setting a learning rate, and adjusting the weight distribution of each coefficient in the comprehensive cleaning model according to the error back propagation mechanism of the neural network until the minimum cleaning error is achieved.

[0008] Furthermore, the environmental risk model outputs the environmental risk coefficient , considering the influence of ambient temperature, humidity and dust concentration:

[0009]

[0010] in, 、 、 ; 、 、 、 、 、 are the environmental weight coefficients of each item, + + + + + =1; T is the temperature value collected in real time, is the mean temperature, is the standard deviation of temperature; is the humidity value collected in real time. is the mean humidity, is the standard deviation of humidity; is the real-time dust concentration value, is the mean dust concentration, is the standard deviation of dust concentration.

[0011] Furthermore, the accumulation risk model outputs the accumulation risk coefficient , taking into account the influence of the deformation of the side of the car, the pressure generated by the accumulated material on the surface of the side of the car, and the adhesion between the accumulated material and the side of the car:

[0012]

[0013] in, 、 and are the weight coefficients of each item, and ; , S is the real-time collected 3D contour data of the vehicle side, It is the standard profile data of the side of the vehicle in the no-load state. is the maximum deformation of the side material or side structure; P is the real-time surface contact pressure of the side at the node, is the average contact pressure of the vehicle side surface collected by the pressure sensor array, is the standard contact pressure threshold, is the standard deviation of surface contact pressure; The adhesion between the accumulated material and the surface of the vehicle side. is the baseline adhesion.

[0014] Furthermore, the cleaning difficulty model outputs the cleaning difficulty coefficient , considering the impact of environmental influence, material accumulation and vehicle side surface roughness on cleaning operations:

[0015]

[0016] in, is the real-time surface roughness of the car side, is the baseline roughness, is the maximum roughness of the car side surface, is the inclination angle.

[0017] Furthermore, the comprehensive cleaning model outputs a comprehensive cleaning index :

[0018]

[0019] in, 、 、 and The cleaning weight coefficient of each item, and + + + =1; The maximum particle size in the particle size distribution of the accumulated material with a concentration of 50% is It is the maximum particle size within the particle size distribution of the accumulated material with a concentration of 90%.

[0020] Furthermore, the weight of the cleaning model is adaptively adjusted based on the cleaning error, which includes the following steps:

[0021] 1) Define the loss function: ,in, The cleaning error is the difference between the thickness of the residual material after cleaning and the expected thickness of the material after cleaning;

[0022] 2) Calculate the loss function for weights 、 、 and gradient;

[0023] 3) Set the learning rate and update the temporary weights in the opposite direction of the gradient using the gradient descent method. The weights are dynamically gradient-decayed with the number of iterations to obtain the temporary weights with the minimum error value.

[0024] 4) Normalize the temporary weights, obtain the final weight value and update the comprehensive cleaning model.

[0025] Furthermore, step S4 also includes priority cleaning path planning based on the improved Dijkstra algorithm, in which key nodes are allocated according to the side contour, the accumulation risk coefficient of each key node is calculated, and the distance between adjacent key nodes can be planned as a sub-path with edge weights. Indicates that the edge weight The calculation formula is:

[0026]

[0027] in, is the Euclidean distance of the subpath between key nodes, is the accumulation risk coefficient of accessing key nodes, is the accumulation risk threshold, is the superthreshold enhancement coefficient;

[0028] The accumulation risk coefficient at the cleaning starting point is initialized to 0, based on the edge weight ,The cleaning path is planned from the cleaning starting point according to the path target and node coverage requirements.

[0029] An intelligent cleaning operation system includes a multimodal data acquisition terminal, an intelligent processing terminal, a cloud database, an intelligent execution control terminal and an automatic cleaning device; wherein,

[0030] The cloud database deploys a distributed relational database cluster to create a historical database, a model parameter database, and a benchmark database.

[0031] The multimodal data acquisition terminal uploads the real-time collected parameter signals to the historical database for storage;

[0032] The intelligent processing terminal retrieves the corresponding parameter values ​​from the historical database, model parameter database, and benchmark database, performs pre-processing analysis on each parameter data, extracts the environmental risk coefficient Er, the accumulation risk coefficient Ar, and the cleaning difficulty coefficient Dc, and imports them into the comprehensive cleaning model to generate a comprehensive cleaning index. The intelligent processing terminal uploads each coefficient and the comprehensive cleaning index to the cloud database for storage and sends them to the intelligent execution control terminal at the same time.

[0033] The intelligent execution control terminal is electrically connected to the automatic cleaning device, which dynamically adjusts the cleaning mode and cleaning parameters according to the comprehensive cleaning index, plans the cleaning path, and sends control instructions and scheduling instructions to the automatic cleaning device. The automatic cleaning device performs the cleaning operation according to the instructions.

[0034] Furthermore, the multimodal data acquisition terminal includes a three-dimensional laser scanner, a pressure sensor array, an inclination sensor, a roughness sensor, a particle size detection sensor, an adhesion sensor, a temperature and humidity sensor, and a dust concentration sensor.

[0035] Compared with the existing technology, the AI-based intelligent cleaning method and system for port train sides of the present invention has the following beneficial effects:

[0036] 1. This cleaning operation method integrates multi-dimensional influencing factors. The system can collect multi-dimensional parameter values ​​such as environmental parameters, vehicle side surface parameters, and material accumulation parameters in real time. The collected multi-dimensional parameters are grouped and noise reduction filtered for pre-processing, and weighted fusion is assigned to extract the environmental risk factor, material accumulation risk factor, and cleaning difficulty factor respectively. This achieves a precise quantitative assessment of the cleaning operation scenario, avoiding the strong subjectivity of traditional manual judgment and the insufficient perception dimension of automated equipment, and significantly improving the pertinence and accuracy of cleaning operations.

[0037] 2. By collecting the parameters of the side of the vehicle before and after cleaning, the cleaning error is determined, indicating that the actual cleaning effect deviates from the expected cleaning effect. Based on AI neural network learning technology, the comprehensive cleaning model is optimized online through adaptive weight update method. Combined with the support of benchmark parameter data, the system can autonomously adapt to different side models, material types and environmental changes.

[0038] 3. The system's intelligent execution control terminal dynamically adjusts the negative pressure and cleaning frequency of the cleaning device based on the comprehensive cleaning index, achieving accurate and effective cleaning and targeted allocation of cleaning capacity. Based on the improved Dijkstra algorithm, it plans priority cleaning paths according to the accumulation risk coefficient, giving priority to high-risk accumulation areas, achieving "on-demand cleaning" and efficient operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1This is a flow chart of the intelligent cleaning method disclosed in the present invention;

[0040] Figure 2 This is a schematic diagram of the composition of the intelligent cleaning operation system disclosed in the present invention. DETAILED DESCRIPTION

[0041] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only the best embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0042] References herein to "embodiments" mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present invention. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute a separate or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0043] Trains transport a variety of materials, and the transport wagons need to be effectively cleaned to avoid over-cleaning or under-cleaning due to mismatched cleanliness. Existing intelligent cleaning technologies have a single perception mode and cannot perceive complex environments. They lack comprehensive evaluation capabilities for accumulated materials and are therefore not suitable for cleaning accumulated materials in complex environments on train sides. For example, for coal deposits with strong adhesion in humid environments, the adhered materials cannot be effectively removed. For sandy materials that are prone to dust in dry and high-temperature environments, insufficient suction leads to serious dust. Therefore, this example embodiment provides an AI-based intelligent cleaning method for port train sides, which comprehensively evaluates the cleaning index by integrating multimodal perception data, calls corresponding cleaning modes for targeted cleaning, and improves cleaning quality. Specifically, Figure 1 As shown, it includes the following steps:

[0044] S1. Acquire environmental parameters, material accumulation parameters, and vehicle side parameters of the cleaning scene through multi-source sensing elements. Environmental parameters include temperature, humidity, and dust concentration. Material accumulation parameters include particle size and adhesion. Vehicle side parameters include 3D profile data, inclination angle, surface contact pressure, and real-time surface roughness of the vehicle side.

[0045] The multi-source parameter data collected in real time are uploaded to the cloud database for storage. Based on the real-time data, the type data of the accumulated materials and the different models and materials of the train side are identified, and the corresponding benchmark data are called out. The environmental benchmark data include the mean and standard deviation of temperature, humidity and dust concentration, which are calculated by collecting the environmental data of the side in the unloaded state for a long time. The benchmark data of the side includes the roughness benchmark value of the side surface in the ideal state and the standard profile data of the unloaded side. Among them, the roughness benchmark value is the laser roughness measuring instrument selecting three representative points in the undamaged and uncontaminated area of ​​the new side surface for measurement and taking the average value, such as the middle, edge and corner of the side, and the benchmark roughness is calculated. The standard profile data is bound to the information of the side panels of different models and materials and stored in the benchmark database; the benchmark adhesion between different materials and different side panels is obtained through adhesion testing. The material of the adhesion testing device is consistent with the material and roughness of the side panel. Under standard conditions, the contact state between each measuring point of the side panel and the material is simulated. Standard pressure is applied and maintained for a certain period of time to simulate the natural deposition process of the material on the side panel. The adhesion testing device is then pulled vertically at a uniform speed to record the maximum pulling force at the moment of separation. The measurement is repeated 5 times and the average value is taken. The average value is the benchmark adhesion. Finally, the benchmark adhesion is bound to the side panel model, measurement point, and material type information and stored.

[0046] In addition, the surface contact pressure of the side of the vehicle includes the surface contact pressure at the measuring point and the average surface contact pressure. , mean surface contact pressure The pressure sensor array collects the average value of all measurement points and calculates the standard deviation of the surface contact pressure. The pressure sensor array is evenly installed on the inner surface of the train side, and the installation density is not less than 10 / ㎡;

[0047] S2. Retrieve the corresponding parameter benchmark values, pre-process the environmental parameters, material accumulation parameters and vehicle side parameters, and perform multi-factor grouping fusion and weight calculation to build the environmental risk model, material accumulation risk model and cleaning difficulty model respectively, and generate the environmental risk coefficient, material accumulation risk coefficient and cleaning difficulty coefficient; specifically,

[0048] (1) Considering the impact of the environment on cleanliness, the environmental risk factor is introduced Conduct assessments, integrate ambient temperature, humidity, and dust concentration to build an environmental risk model, and calculate the environmental risk coefficient :

[0049]

[0050] in, 、 、 ; 、 、 、 、 、 is the environmental weight coefficient of each environmental factor, Inside, + + + + + =1; T is the temperature value collected in real time, is the mean temperature, is the standard deviation of temperature; is the humidity value collected in real time. is the mean humidity, is the standard deviation of humidity; is the real-time dust concentration value, is the mean concentration, is the standard deviation of the concentration;

[0051] (2) The deformation of the side of the vehicle, the pressure generated by the accumulated material on the side of the vehicle, and the adhesion between the accumulated material and the side of the vehicle are the key factors affecting the deposition of materials and the generation of accumulation. The accumulation risk model is constructed by integrating the parameters of these three factors and the accumulation risk coefficient is calculated. :

[0052]

[0053] in, 、 and is the weight coefficient of each factor, all in Inside, and ; Material thickness , S is the real-time collected 3D contour data of the vehicle side, It is the standard profile data of the vehicle side in the unloaded state. is the maximum deformation of the car side material or structure; P is the real-time surface contact pressure of the car side at the node, is the average contact pressure of the vehicle side surface collected by the pressure sensor array, is the standard contact pressure threshold, is the standard deviation of surface contact pressure; The adhesion between the accumulated material and the surface of the vehicle side. is the baseline adhesion;

[0054] (III) The cleaning difficulty needs to comprehensively consider the surface condition of the side of the vehicle, combined with the roughness and inclination, and integrated with the environmental impact and the risk of material accumulation, to build a cleaning difficulty model that comprehensively evaluates the difficulty of cleaning the material accumulation and output the cleaning difficulty coefficient. :

[0055]

[0056] in, is the real-time surface roughness of the car side, is the baseline roughness, is the maximum roughness of the car side surface, The inclination angle of the car side.

[0057] S3. Build a comprehensive cleaning model, import the environmental risk factor, material accumulation risk factor, cleaning difficulty factor, and material accumulation particle size, assign weights, and calculate the comprehensive cleaning index:

[0058] Comprehensive cleaning model outputs comprehensive cleaning index :

[0059]

[0060] in, 、 、 and is the cleaning weight coefficient of each influencing factor, all in Inside, and + + + =1; The maximum particle size in the particle size distribution of the accumulated material with a concentration of 50% is It is the maximum particle size in the particle size distribution of the accumulated material with a concentration of 90%.

[0061] S4. According to the preset cleaning threshold, compare the size relationship between the comprehensive cleaning index and the preset threshold, and determine the corresponding cleaning mode and cleaning parameters. If it is lower than the preset threshold, use the standard mode, set the vibration frequency to the standard frequency, and the negative pressure to H. , H is the negative pressure factor, a preset constant; when it exceeds the preset threshold, the high-frequency vibration mode is triggered, and the negative pressure is H , the vibration frequency is f+v , f is the basic frequency, and v is the frequency factor.

[0062] According to the accumulation risk coefficient of each key node on the car body contour, the priority cleaning path of the cleaning operation is planned, and the key nodes with high accumulation risk are processed first. The logic of "the higher the risk, the lower the weight, and the priority access" is adopted to introduce the accumulation risk coefficient. Based on the improved Dijkstra algorithm, priority cleaning path planning is performed to obtain the shortest path from the cleaning starting point to other key nodes. The accumulation risk coefficient of the cleaning starting point (the first key node) is initialized to 0, and the accumulation risk coefficients of other key nodes are calculated respectively. The path distance between adjacent key nodes of the plannable path is represented by the edge weight. The edge weight between key nodes where the path cannot be planned is infinite, and the edge weight is , is the Euclidean distance of the subpath between key nodes, To calculate the accumulation risk coefficient of visiting key nodes, the shortest path from the cleaning starting point to all other key nodes is calculated;

[0063] For example, assuming that the side of the vehicle has five key nodes: A(0,0,0), B(2,0,0), C(3,2,0), D(1,3,0), and E(0,2,0), the risk threshold of material accumulation is preset. is 0.5, where Ar of A is 0.3, Ar of B is 0.6, Ar of C is 0.4, Ar of D is 0.7, and Ar of E is 0.4. The path distance between adjacent key nodes is defined by the edge weight W. The calculation formula is:

[0064]

[0065] in, is the Euclidean distance of the subpath between key nodes, is the accumulation risk coefficient of accessing key nodes, is the accumulation risk threshold, is the superthreshold enhancement coefficient;

[0066] The accumulation risk coefficient at the cleaning starting point is initialized to 0. According to the known data, the Euclidean distance between AB is d A-B =2, access B's edge weight W A-B =0.88; the Euclidean distance of AE is d A-E ≈2, access E's edge weight W A-E =1.2; the Euclidean distance of ED is d E-D ≈1.414, the edge weight W from E to D E-D =0.509; the Euclidean distance between BC is d B-C ≈2.236, the edge weight W from B to C B-C =1.342, the Euclidean distance of DC is d D-C ≈2.236, the edge weight W of access C D-C =1.342;

[0067] If the path target is set to a single continuous path and covers all nodes, then each time the unvisited node closest to the cleaning starting point is selected, and the shortest path to its neighboring nodes is updated according to the edge weight. This process is repeated until all nodes are visited. The improved Dijkstra algorithm path planning process is as follows:

[0068] The priority queue of the cleaning starting point A is initialized: {A: 0}, indicating that the weight of the cleaning starting point A is 0;

[0069] Step 1: Take out A, expand neighbor B with a weight of 0.88 and E with a weight of 1.2, and update the queue to {B: 0.8, E: 1.2};

[0070] Step 2: First remove B with the smallest weight and mark it as "processed". The weight of the extended neighbor C is 0.88+1.342=2.222, and the queue is updated to {E: 1.2, C: 2.222};

[0071] Step 3: Take out E, expand the weight of neighbor D to 1.2 + 0.509 = 1.709, and update the queue to {D: 1.709, C: 2.222};

[0072] Step 4: Take out D and mark it as "processed". The weight of the expanded neighbor C is 1.709+1.342=3.054, and the queue is updated to {C: 2.222};

[0073] In the above path planning process, high-risk area B is processed earlier than D, reflecting the logic of "the higher the risk, the lower the weight, the priority access". If the goal is to "cover all high-risk areas", the shortest path and its processing nodes are A-BEDC. After priority clean path planning, nodes B and D are processed first;

[0074] If the goal is set to an efficient path and covers high-risk nodes, the efficient path to access high-risk nodes B and D is: execute path AED and process AB separately. The planning process will not be described here.

[0075] If the goal is set to a single continuous path that only covers high-risk nodes, the optimal path is ABCD. The planning process will not be described here.

[0076] S5. Use a 3D laser scanner to generate 3D point cloud data of the sidewall surface before and after cleaning, and determine the thickness of residual material at each key node on the sidewall surface. , then according to the expected thickness of the accumulated material after cleaning , calculate the cleaning error :

[0077] = ;

[0078] Based on AI neural network deep learning technology, the comprehensive cleaning model is updated through the error back propagation mechanism to reduce cleaning errors. Specifically, the following steps are included:

[0079] 1) Define the loss function of the error: ;

[0080] 2) Calculate the loss function for the cleaning weight coefficient 、 、 and Gradient:

[0081]

[0082] in, ;

[0083] Similarly:

[0084]

[0085] 3) If > , indicating insufficient cleanliness, then It is a positive value, the gradient is positive, and a step size of 0.01 is used, that is, the learning rate is 0.01, and the cleaning weight coefficient decreases along the negative direction of the gradient, that is, the weight value is reduced, and each cleaning weight coefficient is iteratively updated;

[0086]

[0087] in, is the temporary weight, is the learning rate, is the initial weight value or the weight value of the previous gradient;

[0088] Similarly, the gradient updates other temporary weights in the same way, and each weight is dynamically gradient-decreased with the number of iterations. When the value is continuously reduced until it approaches the minimum value of 0, the temporary weight under the minimum error value is obtained; the temporary weight values ​​updated last time are normalized to obtain the final weight value and update the comprehensive cleaning model;

[0089] Understandably, if < , indicating excessive cleaning, then If it is a negative value, the gradient is negative, and a step size of 0.01 is used to make the weight decrease along the positive gradient, that is, to increase the weight value and iteratively update the cleaning weight coefficients.

[0090] This exemplary embodiment also provides an intelligent cleaning operation system, such as Figure 2 As shown, the operation method based on the above embodiment includes:

[0091] The multimodal data acquisition terminal is used to collect environmental parameters, car side parameters and material accumulation parameters in real time. It includes a three-dimensional laser scanner, a pressure sensor array, an inclination sensor, a roughness sensor, a particle size detection sensor, an adhesion sensor, a temperature and humidity sensor, and a dust concentration sensor. Among them, the three-dimensional laser scanner is installed on the top or both sides of the loading building, and collects the three-dimensional contour data S (x, y, z) of the car side by emitting lasers. The pressure sensor array is used to directly collect the contact pressure distribution of the material on the car side during loading. It is evenly installed on the inner surface of the train car side with an installation density of ≥10 / ㎡. There are at least 2 inclination sensors, covering the main tilt area, to collect the average car side inclination angle θ in real time. It is installed at least at the center of the car side panel and the bottom plate. The roughness sensor collects the current car side surface roughness R in real time. a The adhesion force sensor collects the adhesion force F at each key node between the current material and the side of the vehicle in real time. a ,The particle size detection sensor collects the particle size value and distribution of the accumulated material, and the temperature and humidity sensor and dust concentration sensor collect the environmental parameters of the scene in real time;

[0092] The cloud database deploys a distributed relational database cluster to create a historical database, a model parameter database, and a benchmark database. The historical database stores time series data of vehicle side parameters, material accumulation parameters, and environmental parameters uploaded by the multimodal data acquisition terminal, as well as data on different vehicle side models, materials, and material accumulation types. It adopts a time-sliced ​​storage strategy for historical data and implements data lifecycle management. The model parameter database stores the weight coefficients and version iteration records of the comprehensive cleaning model. The benchmark database stores benchmark statistical values ​​of ambient temperature, humidity, and dust concentration, material accumulation benchmark parameters, standard three-dimensional profiles of unloaded vehicle side, and material benchmark parameters.

[0093] The intelligent processing terminal retrieves the real-time vehicle side parameters, material accumulation parameters, environmental parameters and pressure parameters, and calculates the environmental risk coefficient E after noise reduction filter preprocessing and combining with the benchmark parameters. r , material accumulation risk coefficient A r and cleaning difficulty coefficient D c , and pour the above coefficients into the comprehensive cleaning model to calculate the comprehensive cleaning index I c and upload the result data to the cloud database for storage;

[0094] The intelligent execution control terminal is connected to the intelligent processing terminal signal, and the intelligent processing terminal sends the comprehensive cleaning index I c Dynamically adjust the cleaning mode and cleaning parameters of the automatic cleaning device, such as negative pressure and vibration frequency, and process the vehicle side 3D profile and material accumulation risk coefficient A sent by the intelligent processing terminal. rDistribution, by improving the Dijkstra algorithm to plan the priority cleaning path, and give priority to scheduling the automatic cleaning device to the high A r Clean up the material accumulation area.

[0095] Those skilled in the art will understand that all or part of the steps of the above-mentioned embodiment method can be implemented by instructing the relevant hardware through a software program, and the program can be applied to relevant devices and equipment and can be stored in a computer-readable storage medium.

[0096] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An AI-based intelligent cleaning method for railway carriages in port areas, characterized in that: The following steps are involved: S1. Acquire environmental parameters, material accumulation parameters, and side rail parameters of the cleaning scene, and identify the material accumulation type and train side rail type. The environmental parameters include temperature, humidity, and dust concentration. The material accumulation parameters include the particle size and adhesion of the material accumulation. The side rail parameters include the three-dimensional profile data, inclination angle, surface contact pressure, and real-time surface roughness of the side rail. S2. Preprocess the environmental parameters, material accumulation parameters, and side rail parameters, perform multi-factor grouping fusion and weight calculation, and respectively build an environmental risk model, a material accumulation risk model, and a cleaning difficulty model to generate an environmental risk coefficient, a material accumulation risk coefficient, and a cleaning difficulty coefficient. S3. Build a comprehensive cleaning model, import the environmental risk factor, the material accumulation risk factor, the cleaning difficulty factor, and the material accumulation particle size, assign weights, and calculate a comprehensive cleaning index; S4. Set a cleaning threshold to evaluate the comprehensive cleaning index and determine the cleaning mode and cleaning parameters. S5. Collect the surface data of the vehicle side before and after cleaning, calculate the cleaning error, set the learning rate, and adjust the weight distribution of each coefficient in the comprehensive cleaning model according to the error back propagation mechanism of the neural network until the minimum cleaning error is achieved.

2. The AI-based intelligent cleaning method for railway carriages in port areas according to claim 1 is characterized by: The environmental risk model outputs the environmental risk coefficient , considering the influence of the temperature, humidity and dust concentration: ； in, 、 、 ; 、 、 、 、 、 are the environmental weight coefficients of each item, + + + + + =1; T is the temperature value collected in real time, is the mean temperature, is the standard deviation of temperature; is the humidity value collected in real time. is the mean humidity, is the standard deviation of humidity; is the real-time dust concentration value, is the mean dust concentration, is the standard deviation of dust concentration.

3. The AI-based intelligent cleaning method for railway carriages in port areas according to claim 2 is characterized by: The accumulation risk model outputs the accumulation risk coefficient , taking into account the influence of the deformation of the side of the car, the pressure generated by the accumulated material on the surface of the side of the car, and the adhesion between the accumulated material and the side of the car: ； in, 、 and are the weight coefficients of each item, and ; , S is the three-dimensional profile data collected in real time, It is the standard profile data of the side of the vehicle in the no-load state. is the maximum deformation of the side material or side structure; P is the real-time surface contact pressure of the side at the node, is the average value of the surface contact pressure collected by the pressure sensor array, is the standard contact pressure threshold, is the standard deviation of surface contact pressure; is the adhesion between the accumulated material and the surface of the vehicle side, is the baseline adhesion.

4. The AI-based intelligent cleaning method for railway carriages in port areas according to claim 3 is characterized by: The cleaning difficulty model outputs the cleaning difficulty coefficient , considering the impact of environmental influence, material accumulation and vehicle side surface roughness on cleaning operations: ； in, is the real-time surface roughness of the vehicle side, is the baseline roughness, is the maximum roughness of the car side surface, is the inclination angle.

5. The AI-based intelligent cleaning method for railway carriages in port areas according to claim 4 is characterized by: The comprehensive cleaning model outputs the comprehensive cleaning index : ； in, 、 、 and is the cleaning weight coefficient of each item, and + + + =1; The maximum particle size in the particle size distribution of the accumulated material with a concentration of 50% is It is the maximum particle size within the particle size distribution of the accumulated material with a concentration of 90%.

6. The AI-based intelligent cleaning method for railway carriages in port areas according to claim 5 is characterized by: The weight of the cleaning model is adaptively adjusted based on the cleaning error, which includes the following steps: 1) Define the loss function: ,in, The cleaning error is the difference between the thickness of the residual material after cleaning and the expected thickness of the material after cleaning; 2) Calculate the loss function for weights 、 、 and gradient; 3) setting the learning rate, and updating each temporary weight in the opposite direction of the gradient by the gradient descent method, wherein each weight is dynamically gradient-decayed with the number of iterations, and obtaining the temporary weight under the minimum error value; 4) Normalizing each temporary weight to obtain a final weight value and updating the comprehensive cleaning model.

7. The AI-based intelligent cleaning method for railway carriages in port areas according to claim 6 is characterized by: Step S4 also includes cleaning path planning based on the improved Dijkstra algorithm, in which key nodes are allocated according to the side contour, the accumulation risk coefficient of each key node is calculated, and the distance between adjacent key nodes can be planned as a sub-path with edge weights. Indicates that the edge weight The calculation formula is: ； in, is the Euclidean distance of the subpath between the key nodes, is the material accumulation risk coefficient for accessing key nodes, is the accumulation risk threshold, is the superthreshold enhancement coefficient; The accumulation risk coefficient at the cleaning starting point is initialized to 0, based on the edge weight , the cleaning path is planned from the cleaning starting point according to the path target and node coverage requirements.

8. An intelligent cleaning system, characterized by: It includes multimodal data acquisition terminal, intelligent processing terminal, cloud database, intelligent execution control terminal and automatic cleaning device; among them, The cloud database creates a history database, a model parameter database and a benchmark database by deploying a distributed relational database cluster; The multimodal data acquisition terminal uploads the parameter signals collected in real time to the historical database for storage; The intelligent processing terminal retrieves corresponding parameter values ​​from the historical database, the model parameter database, and the benchmark database, performs preprocessing analysis on each parameter data, extracts the environmental risk coefficient Er, the accumulation risk coefficient Ar, and the cleaning difficulty coefficient Dc, and imports them into the comprehensive cleaning model to generate a comprehensive cleaning index; the intelligent processing terminal uploads each coefficient and the comprehensive cleaning index to the cloud database for storage, and simultaneously sends them to the intelligent execution control terminal; The intelligent execution control terminal is electrically connected to the automatic cleaning device. It dynamically adjusts the cleaning mode and cleaning parameters according to the comprehensive cleaning index, plans the cleaning path, and sends control instructions and scheduling instructions to the automatic cleaning device. The automatic cleaning device performs the cleaning operation according to the instructions.

9. The intelligent cleaning system according to claim 8, characterized in that: The multimodal data acquisition terminal includes a three-dimensional laser scanner, a pressure sensor array, an inclination sensor, a roughness sensor, a particle size detection sensor, an adhesion sensor, a temperature and humidity sensor, and a dust concentration sensor.

Images