Self-adaptive garbage cleaning method and system based on neural fuzziness

Abstract

The invention relates to the technical field of garbage cleaning, in particular to a self-adaptive garbage cleaning method and system based on neural fuzziness, and the method comprises the steps: constructing a multi-mode garbage image data set, and dividing the data set; fusing fuzzy clustering and deep learning features to train a target detection model; performing fuzzification processing on the image to judge the garbage category; establishing a fuzzy rule base to realize real-time decision making; training the model by adopting a hybrid optimization strategy; estimating the volume and mass of the garbage based on a three-dimensional sensing technology; a fuzzy PID controller is used for adjusting the power of the cleaning device; deploying an edge-cloud collaborative architecture update rule base; the system comprises a multi-modal data acquisition unit, a hybrid computing unit, a dynamic power control unit and a cloud management platform. The garbage classification can be accurately identified, the power of the cleaning device is optimized, energy consumption and efficiency collaboration is achieved, the system adaptability and decision real-time performance are improved, and the defects of a traditional garbage cleaning technology are effectively overcome.

Description

Technical Field

[0001] The present invention relates to the technical field of garbage cleaning, and in particular to a neuro-fuzzy-based adaptive garbage cleaning method and system. Background Art

[0002] In today's society, garbage collection plays a key role in urban environmental maintenance, resource recycling, and public health protection. With the continuous expansion of urban scale and the continuous improvement of residents' living standards, the output and types of garbage are increasing sharply, which puts higher demands on garbage collection technology. However, the existing garbage collection technology has the following problems that need to be solved:

[0003] Garbage identification and classification challenges: Traditional garbage cleaning equipment mainly relies on manual identification or simple image recognition technology to distinguish garbage types. Manual identification is inefficient and easily affected by subjective factors and fatigue, and it is difficult to ensure accuracy and consistency in large-scale garbage cleaning scenarios. Simple image recognition technology faces many challenges when processing garbage images in complex backgrounds. In poor lighting conditions (such as at night, on cloudy days or in shadowed areas), image feature extraction is difficult and misjudgment is prone to occur. When garbage is partially blocked, overlapped or has similar shapes, existing algorithms are also difficult to accurately distinguish different types of garbage, resulting in the inability to correctly classify and process the garbage, affecting subsequent recycling or environmentally friendly disposal.

[0004] Imbalance between cleaning efficiency and energy consumption: Current cleaning devices lack intelligence in power regulation and usually operate at fixed power or a limited number of power levels. They cannot achieve accurate power matching when faced with different quantities and types of garbage. When encountering a small amount of light garbage, high-power operation will cause a huge waste of energy. When dealing with a large amount of heavy garbage or tightly packed garbage, insufficient power will greatly reduce the cleaning effect, extend the cleaning time, and reduce the overall cleaning efficiency. This imbalance between efficiency and energy consumption not only increases operating costs, but also does not conform to the concept of sustainable development.

[0005] Lack of real-time adaptive decision-making capabilities: The garbage cleaning environment is complex and changeable. The distribution and types of garbage in different areas (such as commercial areas, residential areas, parks, etc.) vary significantly and may change at any time. Existing cleaning systems are often unable to perceive these changes in real time and make corresponding decision adjustments. For example, when a large amount of garbage suddenly accumulates in a certain area, the traditional system cannot quickly optimize the cleaning path and adjust the working mode of the cleaning device, resulting in delayed cleaning work and affecting environmental quality. In addition, in the face of new types of garbage or special cleaning scenarios, the system lacks self-learning and adaptability, and it is difficult to cope with the ever-changing actual needs.

[0006] Insufficient data processing and utilization: With the development of sensor technology, the amount of data that can be collected during the garbage cleaning process has increased significantly, including garbage images, location information, cleaning equipment operating status data, etc.; however, the existing cleaning system has serious deficiencies in data processing and is unable to fully tap the value of this data; a large amount of data is simply stored or used only for basic status monitoring, and has not been effectively used to optimize cleaning strategies, improve equipment performance, and improve garbage disposal processes; at the same time, there are obstacles to data transmission and sharing, and data between different devices and systems is difficult to use in a coordinated manner, forming information islands and limiting the improvement of overall cleaning efficiency.

[0007] Therefore, in order to solve the above problems, a neuro-fuzzy based adaptive garbage cleaning method and system are proposed. Summary of the invention

[0008] The object of the present invention is to provide a neuro-fuzzy based adaptive garbage cleaning method and system to solve the problems raised in the above background technology.

[0009] To achieve the above object, the present invention provides the following technical solutions:

[0010] A neuro-fuzzy based adaptive garbage cleaning method comprises the following steps:

[0011] S1. Construct a multimodal garbage image dataset by labeling the garbage types and dividing it into training set, validation set and test set;

[0012] S2. Target detection model training based on multimodal feature fusion, integrating fuzzy clustering features and deep learning features, and dynamically optimizing classification boundaries;

[0013] S3, fuzzy processing is performed on the garbage image, a membership matrix is ​​generated by combining the fuzzy clustering algorithm, and a threshold is set to determine the garbage category;

[0014] S4. Establish a fuzzy rule base and realize real-time decision-making through fuzzy reasoning model;

[0015] S5. Use hybrid optimization strategy to train the model, combining global search and second-order precision tuning to improve model efficiency and convergence speed;

[0016] S6, real-time estimation of garbage volume and mass based on 3D sensing technology, matching physical parameters with density database;

[0017] S7, dynamically adjust the power of the cleaning device through the fuzzy PID controller to achieve coordinated optimization of energy consumption and efficiency;

[0018] S8. Deploy edge-cloud collaborative architecture to support real-time decision-making and incremental updates of rule bases.

[0019] As a preferred solution, in step S2, the multimodal feature fusion specifically includes:

[0020] The membership matrix generated by FCM fuzzy clustering Feature map output by YOLOv8 Channel cascading is performed, where The membership matrix generated for FCM fuzzy clustering is, is the number of clusters, is the sample size, It is the feature map output by YOLOv8. is the height of the feature map, is the width of the feature map, is the number of channels of the feature map, and the feature weight is calculated through the dynamic weight allocation module. The formula is: ,in, For the The feature weights of clusters, For the The membership degree of a cluster, For the The membership degree of a cluster, is the temperature coefficient, is the number of clusters;

[0021] Weighted fusion features Input to the detection head for classification, where is the weighted fusion feature, For the corresponding The feature map of the clusters.

[0022] As a preferred solution, the fuzzy clustering algorithm in step S3 is FCM, and its membership update formula is: ,in, For the The samples belong to The membership degree of a cluster, For the samples, For the Cluster centers, Representative Cluster centers, is the fuzzy index;

[0023] set up , membership threshold The confusion matrix verifies that the classification accuracy is >92%.

[0024] As a preferred solution, the hybrid optimization strategy in step S5 includes:

[0025] Grid search phase: In the parameter space Traverse the combination internally and filter out the errors below the threshold Candidate parameters of and is the parameter to be optimized, is the error threshold;

[0026] LM optimization stage: perform second-order optimization based on candidate parameters, and the update formula is: ,in, is the parameter update amount, is the Jacobian matrix of the error versus parameter, is the Jacobian matrix The transposed matrix of is the damping factor, is the identity matrix, is the error vector, the initial value of the damping factor , the dynamic adjustment range is .

[0027] As a preferred solution, in step S6, the three-dimensional volume estimation formula is: ,in, is the three-dimensional volume of the garbage; is the number of pixels; For pixels in The physical size of the direction, , For pixels in The physical size of the direction, , is the physical width of the sensor, is the width of the image, is the physical height of the sensor, is the height of the image; For the The depth value of pixels; the quality is calculated as , For the quality of garbage, is the density of garbage, Match from a preset database, which contains 20 categories of garbage density values.

[0028] As a preferred solution, in step S7, the rule base of the fuzzy PID controller includes:

[0029] Rule 1: For "Taisho" and For "positive", THEN ,in, is the error at the current moment; is the rate of change of error; is the power adjustment amount; is the proportionality coefficient; is the differential coefficient;

[0030] Rule 2: is "Medium" and is "zero", THEN ;

[0031] Rule 3: is a "small negative" and For "negative slowness", ,in, is the integration coefficient; is the integral of the error;

[0032] The membership function adopts Gaussian type and is defined as ,in, Error The degree of membership belongs to "big", is the center of the Gaussian membership function, is the standard deviation of the Gaussian membership function, , , response time <200ms.

[0033] As a preferred solution, the edge-cloud collaborative architecture in step S8 includes:

[0034] The lightweight TSK fuzzy inference model is deployed on the edge, and its rule output is: ,in, For the The output of the rule, is the input vector; For the The constant term of the rule; For the Rule No. The coefficients of the input variables are updated through incremental learning. is the number of input variables;

[0035] The cloud performs density database expansion every month, adding new garbage category density values and synchronized to the edge device via OTA.

[0036] A neuro-fuzzy based adaptive garbage cleaning system, comprising:

[0037] Multimodal data acquisition unit: integrates RGB-D camera and LiDAR to collect image and depth information;

[0038] Hybrid computing unit: deploys a neural fuzzy-based adaptive garbage cleaning method, including an FPGA-accelerated fuzzy reasoning module. The FPGA-accelerated fuzzy reasoning module adopts a parallel architecture to perform hardware acceleration on the fuzzy reasoning process, with a single-frame processing delay of <50ms and power consumption of <5W;

[0039] Dynamic power control unit: supports 50W-150W power adjustment, and the suction nonlinear mapping formula is: ,in, For suction, is power, is the coefficient, is the index, , , maximum suction force ≥ 200N;

[0040] Cloud management platform: provides rule base version control and incremental learning interface.

[0041] It can be seen from the technical solution provided by the present invention that the neural fuzzy-based adaptive garbage cleaning method and system provided by the present invention has the following beneficial effects:

[0042] Improve the accuracy of garbage identification and classification: By constructing a multimodal garbage image dataset, integrating fuzzy clustering features and deep learning features to train the target detection model, and dynamically optimizing the classification boundaries, different types of garbage can be identified more accurately; the membership matrix generated by FCM fuzzy clustering is used to cascade channels with the feature map output by YOLOv8, and the feature weights are calculated through the dynamic weight allocation module, so that the model can fully mine the fuzzy information and deep learning features in the data to improve the classification accuracy; the garbage category is determined by setting the membership threshold of the fuzzy clustering algorithm, and the classification accuracy is greater than 92% verified by the confusion matrix, which effectively improves the reliability of garbage identification;

[0043] Optimize the power regulation of the cleaning device and reduce energy consumption: Use a fuzzy PID controller to dynamically adjust the power of the cleaning device. According to the real-time situation of the garbage (such as volume, mass, etc.) and the working status of the cleaning device, adjust the power in real time to achieve coordinated optimization of energy consumption and efficiency; by establishing a fuzzy rule base, adjust the power according to factors such as error and error change rate, so that the cleaning device can not only ensure the cleaning effect but also avoid unnecessary energy consumption when facing different garbage scenes; the nonlinear mapping formula of suction force realizes the reasonable matching of power and suction force. Within the power adjustment range of 50W-150W, the maximum suction force is ≥200N, ensuring efficient cleaning under different garbage loads;

[0044] Realize real-time decision-making and adaptive system updates: Deploy edge-cloud collaborative architecture, deploy lightweight TSK fuzzy inference model on the edge, support real-time decision-making, and quickly respond to various situations in the garbage cleaning process; the cloud performs density database expansion every month and synchronizes it to edge devices via OTA to achieve incremental updates of the rule base, so that the system can continuously adapt to new garbage categories and scene changes; with the emergence of new types of garbage, the cloud updates the density database and synchronizes it to edge devices to ensure the accuracy of garbage quality estimation, thereby optimizing the cleaning strategy and improving the overall adaptability and intelligence of the system;

[0045] Improve system computing efficiency and stability: The FPGA-accelerated fuzzy reasoning module in the hybrid computing unit adopts a parallel architecture to perform hardware acceleration on the fuzzy reasoning process. The single-frame processing delay is <50ms and the power consumption is <5W, which greatly improves the computing efficiency of the system and ensures stability and rapid response capabilities when processing large amounts of garbage image data in real time. This hardware acceleration method not only improves the speed of garbage identification and decision-making, but also reduces the energy consumption of the system, enhancing the feasibility and reliability of the system in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A schematic diagram of the steps of a neuro-fuzzy-based adaptive garbage cleaning method of the present invention;

[0047] Figure 2 This is a schematic diagram of the structure of a neuro-fuzzy based adaptive garbage cleaning system of the present invention. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0049] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0050] like Figure 1-2 As shown, an embodiment of the present invention provides a neural fuzzy-based adaptive garbage cleaning method, comprising the following steps:

[0051] S1. Construct a multimodal garbage image dataset by labeling the garbage types and dividing it into training set, validation set and test set;

[0052] S2. Target detection model training based on multimodal feature fusion, integrating fuzzy clustering features and deep learning features, and dynamically optimizing classification boundaries;

[0053] S3, fuzzy processing is performed on the garbage image, a membership matrix is ​​generated by combining the fuzzy clustering algorithm, and a threshold is set to determine the garbage category;

[0054] S4. Establish a fuzzy rule base and realize real-time decision-making through fuzzy reasoning model;

[0055] S5. Use hybrid optimization strategy to train the model, combining global search and second-order precision tuning to improve model efficiency and convergence speed;

[0056] S6, real-time estimation of garbage volume and mass based on 3D sensing technology, matching physical parameters with density database;

[0057] S7, dynamically adjust the power of the cleaning device through the fuzzy PID controller to achieve coordinated optimization of energy consumption and efficiency;

[0058] S8. Deploy edge-cloud collaborative architecture to support real-time decision-making and incremental updates of rule bases.

[0059] In this embodiment, step S1 is to construct a multimodal garbage image dataset by marking the types of garbage and dividing it into a training set, a validation set, and a test set. The following is a detailed description of the steps:

[0060] Step S1-1: Selection and deployment of multimodal data acquisition equipment:

[0061] Select high-performance RGB-D cameras to ensure that they have high-resolution imaging capabilities and accurate depth information acquisition functions, and can clearly capture the appearance characteristics and spatial location of garbage; at the same time, when used with laser radar, its point cloud resolution must meet the requirements of accurately identifying the shapes and distribution of different garbage; these devices should be reasonably installed on mobile collection platforms (such as vehicles, robots, etc.) to ensure that when collecting data in different scenes (such as urban streets, community garbage dumps, parks, etc.), the device's field of view can fully cover the garbage area;

[0062] Step S1-2: Raw data collection:

[0063] The collection platform is controlled to move along a preset path in the selected garbage scene. During this time, the RGB-D camera continuously captures garbage images at a set frame rate and simultaneously obtains corresponding depth image data. The lidar scans the surrounding environment in real time to generate point cloud data of the garbage. During the collection process, the impact of factors such as time and weather on the appearance of garbage is fully considered. For example, data is collected under different lighting conditions in sunny, cloudy, rainy days, and in the morning and evening to ensure data diversity.

[0064] Step S1-3: Preliminary screening of raw data:

[0065] After the acquisition is completed, the large amount of raw data obtained is initially screened; images and point cloud data that are incomplete or of poor quality due to equipment failure, occlusion, etc. are eliminated; for example, RGB images that are blurred, partially missing, or have serious noise interference, and depth images with abnormal depth information are directly removed from the data set to reduce the amount of invalid data for subsequent processing;

[0066] Step S1-4: Data preprocessing:

[0067] Image preprocessing: De-noise the filtered RGB images, use algorithms such as Gaussian filtering to remove random noise in the images, and smooth the image details; then perform size normalization to adjust all images to the same size, such as [specific size value], to meet the requirements of subsequent model processing; for depth images, perform denoising and smoothing operations, and align them with RGB images to ensure that the two accurately correspond in spatial position;

[0068] Point cloud data preprocessing: filter the point cloud data generated by the lidar to remove outliers and noise points and improve the quality of the point cloud data; use the downsampling algorithm to reduce the density of the point cloud data without losing key information, reduce the amount of data, and facilitate subsequent processing; at the same time, fuse the point cloud data with the RGB-D image data to establish a unified coordinate system so that different modal data can be used in a coordinated manner;

[0069] Step S1-5: Labeling of garbage types:

[0070] Organize professional labelers to use specialized image labeling tools to label the types of garbage in the pre-processed multimodal data. Labelers need to carefully observe the appearance features of the garbage in the RGB image, and combine the spatial information provided by the depth image and point cloud data to accurately mark the location and category of each type of garbage in the image, such as plastic bottles, cans, waste paper, fruit peels and other common garbage categories. During the labeling process, formulate detailed labeling specifications and review mechanisms to ensure the accuracy and consistency of labeling. For complex scenes or garbage that is difficult to determine the category, labeling is carried out through multi-person discussion or expert guidance.

[0071] Step S1-6: Dataset division:

[0072] A stratified sampling method was used to divide the labeled multimodal garbage image dataset into 70% as the training set, 15% as the validation set, and 15% as the test set. Each subset was ensured to contain data samples of various garbage categories and different scenes and lighting conditions, so as to ensure that the model can fully learn the diverse characteristics of the data during the training, validation, and testing processes, and avoid model performance deviations due to uneven data distribution. After the division was completed, the data of the training set, validation set, and test set and the corresponding annotation information were saved separately to prepare for subsequent model training and evaluation.

[0073] In this embodiment, step S2 is the target detection model training based on multimodal feature fusion, which aims to integrate fuzzy clustering features and deep learning features, dynamically optimize classification boundaries, and thus improve the accuracy of garbage identification and model performance; the detailed steps are as follows:

[0074] Step S2-1: FCM fuzzy clustering generates membership matrix:

[0075] Determine the number of clusters :Determine the appropriate number of clusters based on the complexity of garbage types and image diversity in the multimodal garbage image dataset For example, if the dataset contains garbage of different shapes and materials, you may need to set a larger values ​​to better distinguish the characteristics of different types of garbage; if the garbage type is relatively single, the smaller The value can meet the demand;

[0076] Run the FCM algorithm: input the multimodal junk image data preprocessed by step S1 into the FCM fuzzy clustering algorithm; during the algorithm operation, it will update the membership of each sample (pixel point in the image or feature point after feature extraction, etc.) to each cluster through continuous iteration based on the similarity measurement between data points (such as Euclidean distance, etc.); in each iteration, the algorithm will adjust the cluster center and membership at the same time, so that the objective function (such as the sum of squared errors) gradually decreases; when the change in the objective function is less than the preset threshold (such as 0.001) or the objective function value is basically stable after multiple consecutive iterations, the algorithm converges and generates a stable membership matrix ,in, The membership matrix generated for FCM fuzzy clustering is, is the number of clusters, is the number of samples; this matrix reflects the fuzzy affiliation between each sample and each cluster, that is, each sample belongs to different clusters to different degrees;

[0077] Step S2-2: YOLOv8 output feature map:

[0078] Model loading and configuration: Select the trained YOLOv8 model and load it to a device with computing power (such as a GPU). At the same time, according to the performance and resource conditions of the device, reasonably configure the model's operating parameters, such as setting the batch size and adjusting memory allocation, to ensure that the model can run efficiently and stably.

[0079] Image input and feature extraction: Multimodal garbage images are input into the YOLOv8 model one by one in order. The model first performs a convolution operation on the input image, sliding different convolution kernels on the image to extract local features in the image. Then, a pooling operation is performed to downsample the feature map obtained by convolution, reducing the amount of data while retaining key features. After multiple layers of convolution and pooling operations, a feature map containing rich semantic and spatial information is output at a specific network layer of the model. ,in, It is the feature map output by YOLOv8. is the height of the feature map, is the width of the feature map, is the number of channels of the feature map; these feature maps describe the characteristics of garbage from different scales and angles. For example, large-size feature maps can capture the overall shape and location information of garbage, while small-size feature maps focus on the detailed texture and other features of garbage;

[0080] Step S2-3: Channel cascade operation:

[0081] Data dimensionality adaptation: Carefully check the membership matrix And the feature map output by YOLOv8 If the two dimensions do not match, adjust according to the actual situation; for example, if The dimension does not meet the cascade requirements, and it may be necessary to expand its dimension and add an axis with a dimension of 1 on the appropriate dimension; if the feature map Some dimensions of are too large or too small, and may need to be cropped or padded to ensure that the two can be correctly spliced ​​in the channel dimension;

[0082] Cascade fusion: The membership matrix of the adapted dimension With feature map Cascading is performed in the channel dimension; the new features obtained after cascading contain the sample classification fuzzy information provided by fuzzy clustering and the image feature information extracted by the deep learning model; for example, for a garbage image, the fuzzy clustering information can reflect the fuzzy association between different areas in the image and different garbage categories, while the deep learning feature map provides the intuitive visual features of the garbage in the image. The combination of the two can more comprehensively describe the characteristics of the garbage and enhance the ability to express garbage characteristics;

[0083] Step S2-4: Dynamic weight allocation module calculates feature weights:

[0084] Set temperature coefficient :Determine the temperature coefficient based on experience and multiple experimental results The value of It plays a key role in regulating the smoothness of weight distribution; if If the value is large, the difference between the weights of different cluster features will be smaller, and the weight distribution will be more even, which means that the importance of the features of each cluster in the fusion feature is relatively balanced; if The smaller the value, the more obvious the weight difference will be. The feature weight corresponding to the cluster with high membership will be larger, and its feature will play a stronger leading role in the fusion feature. For example, in some experiments, it was found that when processing garbage images with complex backgrounds, the larger The value can prevent the model from over-relying on a certain type of clustering feature and improve the robustness of the model. In the scenario where the garbage categories are clearly distinguished, a smaller The value can better highlight the key clustering features and improve the classification accuracy;

[0085] Calculate feature weights: Based on the formula , calculate the corresponding feature weight for each cluster ,in, For the The feature weights of clusters, For the The membership degree of a cluster, For the The membership degree of a cluster, is the temperature coefficient, is the number of clusters; in the calculation process, the membership degree and Respectively represent and The membership degree of a cluster, is the number of clusters; through this formula, the higher the membership of a cluster, the greater the corresponding feature weight, which means that the features contained in the cluster account for a larger proportion in the fusion features and have a more significant impact on subsequent classification;

[0086] Step S2-5: Calculate the weighted fusion features and input them into the detection head for classification:

[0087] Weighted sum: based on the feature weights calculated previously , perform weighted sum operation on the cascaded features to obtain the weighted fusion features ;in, is the weighted fusion feature, For the corresponding The feature graph of each cluster; through weighted summation, the fusion feature can comprehensively consider the importance of different cluster features and effectively integrate the advantageous features of each cluster;

[0088] Detection head classification: weighted fusion features The data is input into the detection head of the target detection model; the detection head usually contains a classifier (such as a Softmax classifier), which calculates the probability that the garbage in the image belongs to each category based on the numerical distribution of the fused features; for example, the classifier performs a series of operations on the fused features to obtain the probability values ​​of the garbage belonging to different categories such as plastic bottles, cans, and waste paper, and finally selects the category with the largest probability value as the predicted category label of the garbage, completing the target detection model training and classification process based on multimodal feature fusion, and realizing accurate identification of garbage categories.

[0089] In this embodiment, step S3 is to perform fuzzy processing on the garbage image, generate a membership matrix by combining the fuzzy clustering algorithm, and set a threshold to determine the garbage category; the detailed steps are as follows:

[0090] Step S3-1: Select fuzzy clustering algorithm: select FCM (Fuzzy C-means Clustering) algorithm as the core algorithm of fuzzy clustering; FCM algorithm realizes data clustering by optimizing the objective function. It can divide data points into different clusters. At the same time, each data point has a membership degree to each cluster. It can handle the fuzziness and uncertainty of data well and is suitable for the classification of junk images.

[0091] Step S3-2: Initialize parameters: Determine fuzzy index The value of ; Fuzzy index The fuzziness that affects the clustering results. The larger it is, the more ambiguous the clustering results are, and the smaller the difference in the degree of membership of each data point to different clusters is; The smaller the value, the closer the clustering result is to hard clustering, and the more likely the data point's membership to a certain cluster is to be 0 or 1. , number of clusters It needs to be determined based on the approximate number of garbage categories in the garbage image dataset and the distribution characteristics of the data. For example, if the dataset mainly contains three types of garbage: plastic, metal, and paper, you can initially set , which can be adjusted later according to the experimental results; in addition, the cluster centers need to be randomly initialized ,The initial value of the cluster center will affect the convergence speed of the algorithm and the final clustering result;

[0092] Step S3-3: Calculate the membership matrix: For each sample in the junk image ( , is the number of samples), according to the membership update formula of the FCM algorithm ,in, For the The samples belong to The membership degree of a cluster, For the samples, For the Cluster centers, Representative Cluster centers, is the fuzzy index, and its The membership of the cluster ; During the calculation process, the membership matrix is ​​updated through continuous iteration and cluster centers , until the change of the membership matrix is ​​less than a preset threshold (such as 0.001), indicating that the algorithm converges and a stable membership matrix is ​​obtained;

[0093] Step S3-4: Setting a threshold to determine the garbage category: Setting a membership threshold It is judged as "large garbage" when it is found; traverse the membership matrix, for each sample, if its membership to a cluster is greater than or equal to 0.7, the garbage corresponding to the sample is judged as "large garbage"; otherwise, further analysis or judgment into other categories is performed according to the actual situation;

[0094] Step S3-5: Verify classification accuracy: Use confusion matrix to verify classification accuracy; compare the garbage category predicted by the model with the actually labeled garbage category to construct a confusion matrix; for example, for the two categories of "large garbage" and "non-large garbage", the confusion matrix will record the number of samples that are actually "large garbage" and correctly predicted as "large garbage", the number of samples that are actually "large garbage" but incorrectly predicted as "non-large garbage", the number of samples that are actually "non-large garbage" and correctly predicted as "non-large garbage", and the number of samples that are actually "non-large garbage" but incorrectly predicted as "large garbage"; by calculating the various indicators in the confusion matrix, the classification accuracy is obtained; the present invention requires that the classification accuracy is greater than 92%. If this accuracy is not achieved, it is necessary to adjust the algorithm parameters (such as the number of clusters C, the fuzzy index m, etc.) or re-preprocess the data, and then perform clustering and verification again until the accuracy requirements are met.

[0095] In this embodiment, step S4 is to establish a fuzzy rule base and implement real-time decision-making through a fuzzy reasoning model to better control the garbage cleaning process; the detailed description of this step is as follows:

[0096] Step S4-1: Determine input and output variables:

[0097] Input variable selection: Comprehensively consider the key influencing factors in the actual garbage cleaning scene and determine the input variables; these variables are usually related to the state of the garbage and the operation of the cleaning device; for example, the current garbage detection results (including garbage type, quantity, distribution and other information), which can reflect the complexity of the garbage cleaning task; the current power of the cleaning device, which is directly related to the cleaning efficiency and energy consumption; and the environmental parameters of the cleaning area (such as ground material, whether there are obstacles, etc.), these factors will affect the difficulty of cleaning;

[0098] Output variable determination: Determine the output variables according to the control requirements of the garbage cleaning task; the main output variables are the control instructions of the cleaning device, such as power adjustment, movement direction and speed adjustment of the cleaning head, etc.; these output variables will directly act on the cleaning device to achieve precise control of the garbage cleaning process;

[0099] Step S4-2: Define fuzzy linguistic variables:

[0100] Fuzzification of input variables: For each input variable, different fuzzy language values ​​are divided according to its value range and actual meaning. For example, for the input variable of garbage quantity, fuzzy language values ​​such as "very little", "little", "medium", "many", and "very much" can be defined; for the current power of the cleaning device, "low", "lower", "moderate", "higher", and "high" can be defined; at the same time, the corresponding membership function is determined for each fuzzy language value. Commonly used membership functions include Gaussian, triangle, trapezoid, etc.; taking the Gaussian membership function as an example, if the membership function of the garbage quantity "many" is defined as , the center value needs to be determined and standard deviation , so that the function can accurately describe the degree to which the amount of garbage is "too much";

[0101] Fuzzification of output variables: Fuzzification is also performed on output variables; for power regulation, fuzzy language values ​​such as "substantially reduced", "slightly reduced", "remain unchanged", "slightly increased", and "substantially increased" can be defined, and their respective membership functions can be determined; in this way, the precise input and output values ​​are converted into fuzzy language variables, which is more in line with the fuzzy characteristics in actual decision-making;

[0102] Step S4-3: Formulate fuzzy rules:

[0103] Rule generation basis: Fuzzy rules are formulated based on expert experience, actual operation data and in-depth understanding of the garbage cleaning process. For example, if the amount of garbage is detected to be "high" and the current power of the cleaning device is "low", based on experience, in order to efficiently clean the garbage, the power of the cleaning device needs to be increased, thus generating a fuzzy rule: "IF the amount of garbage is 'high' and the current power of the cleaning device is 'low', THEN the power adjustment amount is 'substantially increased'";

[0104] Comprehensiveness of rules: Cover all possible combinations of input variables as much as possible, formulate a series of complete fuzzy rules, and form a fuzzy rule base; the rules in the rule base should be coordinated with each other to avoid contradictions or conflicts, and ensure that reasonable decision-making suggestions can be given in different garbage cleaning scenarios;

[0105] Step S4-4: Constructing a fuzzy reasoning model:

[0106] Select the reasoning method: Common fuzzy reasoning methods include Mamdani reasoning method and Sugeno reasoning method. The present invention can select a suitable reasoning method according to actual needs. Taking Mamdani reasoning method as an example, it performs reasoning based on fuzzy relationship synthesis operation.

[0107] Reasoning process implementation: When there is new input data, first determine the membership of the input data to each fuzzy language value based on the membership function of the input variable; then, find the matching fuzzy rules based on the fuzzy rule base; for each matching rule, determine the activation strength of the rule based on the input membership (usually take the minimum value of each input membership); finally, synthesize the outputs of all activated rules to obtain the final fuzzy output result;

[0108] Step S4-5: Defuzzification:

[0109] Method selection: Since the result of fuzzy reasoning is a fuzzy set, a defuzzification operation is required to convert the fuzzy output into an accurate control quantity; commonly used defuzzification methods include the centroid method and the maximum membership method. If the centroid method is used, the calculation formula is ,in, is an element in the domain of output variables, is its corresponding membership degree;

[0110] Output precise control instructions: Through defuzzification calculation, accurate output values ​​are obtained, such as specific power adjustment values, cleaning head movement speed adjustment values, etc. These precise control instructions will be transmitted to the actuator of the cleaning device to achieve real-time control of the garbage cleaning device to adapt to different garbage cleaning scenarios.

[0111] In this embodiment, step S5 adopts a hybrid optimization strategy to train the model, combining global search and second-order precision tuning to improve model efficiency and convergence speed. The specific steps are as follows:

[0112] Step S5-1: Grid search phase:

[0113] Determine the parameter space: clarify the value range of the parameter to be optimized and construct the parameter space ;in, and They are important parameters that are closely related to model performance, and their values ​​will have a significant impact on the training effect of the model;

[0114] Traversing parameter combinations: In a given parameter space, and All possible combinations of exist The value is taken in a certain step within the range. exist The value is taken in the range according to the corresponding step length, and then for each group Combined for subsequent model training and evaluation;

[0115] Screening candidate parameters: For each set of parameter combinations, use the training set data to train the model, and use the validation set data to evaluate the performance of the model; use the error as the evaluation indicator, and set the error below the threshold The parameter combinations are screened out as candidate parameters for subsequent further optimization; these candidate parameters represent the parameter settings that can enable the model to meet certain performance requirements within the current parameter space search range;

[0116] Step S5-2: LM optimization stage:

[0117] Initialization based on candidate parameters: The candidate parameters selected in the grid search phase are used as initial values ​​to start the LM (Levenberg-Marquardt) optimization process; LM optimization is an optimization algorithm for nonlinear least squares problems, especially suitable for fine-tuning model parameters;

[0118] Calculate the correlation matrix and vector: During the optimization process, it is necessary to calculate the Jacobian matrix of the error versus parameter ; The Jacobian matrix describes the rate of change of the error function to each parameter, which reflects the sensitivity of the model error to the change of parameters; at the same time, the damping factor is determined , whose initial value is set to , and the dynamic adjustment range is ; In addition, the error vector , which represents the difference between the model prediction and the true value;

[0119] Parameter update iteration: based on the update formula of LM optimization , calculate the update amount of the parameters ,in, is the parameter update amount, is the Jacobian matrix of the error versus parameter, is the Jacobian matrix The transposed matrix of is the damping factor, is the identity matrix, is the error vector; by continuously iteratively calculating the update amount and updating the parameters, the model parameters are gradually adjusted to gradually reduce the model error until the preset convergence conditions are met (such as the parameter update amount is less than a certain threshold, or the model error no longer decreases significantly in multiple consecutive iterations); in each iteration process, the damping factor is dynamically adjusted according to the performance of the model ; When the model converges quickly, reduce To speed up the convergence; when the model is unstable or oscillating, increase To ensure the stability of the algorithm; after multiple rounds of iterative optimization, a set of parameters tuned with second-order precision are obtained. These parameters can effectively improve the efficiency and convergence speed of the model, making the model perform better in garbage cleaning tasks.

[0120] In this embodiment, step S6 estimates the volume and mass of garbage in real time based on three-dimensional sensing technology, and matches physical parameters in combination with the density database to provide more accurate data support for garbage cleaning. The specific steps are as follows:

[0121] Step S6-1: Obtain relevant parameters and data:

[0122] Determine the physical pixel size: Based on the physical width of the sensor and the width of the image , calculate the pixel in Physical size of direction ; Similarly, according to the physical height of the sensor and the height of the image , and the pixel is Physical size of direction ,in, is the physical height of the sensor, is the height of the image; these physical size parameters are the basis for subsequent accurate estimation of the garbage volume;

[0123] Collect depth value data: Use 3D sensing equipment (such as RGB-D cameras, LiDAR, etc.) to obtain the depth value of each pixel in the garbage image ( , is the number of pixels); the depth value reflects the distance between the garbage and the sensor. Accurate depth value collection is crucial for accurately estimating the volume of garbage;

[0124] Step S6-2: Calculate the three-dimensional volume of garbage:

[0125] Execution volume estimation formula: According to the formula , the small volume elements corresponding to each pixel are accumulated and calculated; in the calculation process, each pixel is , The physical size of the direction is multiplied by the corresponding depth value to obtain the tiny volume represented by each pixel, and then the tiny volumes of all pixels are added together to obtain the three-dimensional volume of the garbage. ; This calculation process takes into account the actual occupation of garbage in three-dimensional space and can estimate the volume of garbage more accurately;

[0126] Step S6-3: Match garbage density and calculate mass:

[0127] Query density database: From the preset database containing 20 types of garbage density values, match the corresponding garbage density according to the previously detected garbage category ; For example, if the garbage is detected to be a plastic bottle, the density value corresponding to the plastic bottle is searched in the database;

[0128] Calculating waste mass: Using the formula , the garbage density obtained by matching The calculated volume of garbage Multiply them together to get the mass of the garbage In this way, combined with the volume and density information of the garbage, an accurate estimation of the garbage quality can be achieved, providing important reference data for the power adjustment and transportation planning of subsequent garbage cleaning equipment.

[0129] In this embodiment, step S7 aims to dynamically adjust the power of the cleaning device through the fuzzy PID controller to achieve coordinated optimization of energy consumption and efficiency. The detailed steps are as follows:

[0130] Step S7-1: Determine the input and output of the fuzzy PID controller:

[0131] Clarify the input variables: select the error at the current moment and the rate of change of error As the input variable of the fuzzy PID controller; among them, the error Calculated by comparing the actual working status of the current cleaning device (such as the difference between the amount of garbage cleaned, the coverage of the cleaning area, etc.) and the expected target; the rate of change of error It is obtained by calculating the change of error at adjacent moments, reflecting the change trend of error;

[0132] Determine the output variable: adjust the power As the output variable of the fuzzy PID controller; this variable directly acts on the cleaning device and is used to adjust the power of the cleaning device in real time to adapt to different garbage cleaning scenarios;

[0133] Step S7-2: Define fuzzy linguistic variables and membership functions:

[0134] Partitioning fuzzy language values: for input variables and Partition fuzzy linguistic values; for example, Divide into "large positive", "medium", "small negative", etc. Divided into "positive fast", "zero", "negative slow", etc.; these fuzzy language values ​​can more intuitively describe the state of the error and its rate of change;

[0135] Determine the membership function: Use Gaussian membership function to quantify the fuzzy language value; Taking the membership function of "large" as an example, it is defined as ,in, Error The degree of membership belongs to "big", is the center of the Gaussian membership function, is the standard deviation of the Gaussian membership function, , is the center of the Gaussian membership function, , is the standard deviation, and the response time is <200ms; in this way, the precise input value is converted into fuzzy language variables, which better reflects the fuzziness in actual control;

[0136] Step S7-3: Constructing fuzzy rule base:

[0137] Formulate fuzzy rules: Formulate fuzzy rules based on expert experience and actual control needs; for example:

[0138] Rule 1: If For "Taisho" and For "positive", THEN ,in, is the error at the current moment; is the rate of change of error; is the power adjustment amount; is the proportionality coefficient; is the differential coefficient. This rule is applicable to situations where the error is large and increases rapidly. The power is adjusted quickly through proportional and differential control to reduce the error.

[0139] Rule 2: If is "Medium" and is "zero", THEN , when the error is at a moderate level and the rate of change is zero, the power is adjusted only by proportional control;

[0140] Rule 3: If is a "small negative" and For "negative slowness", ,in, is the integration coefficient; is the integral of the error. When the error is small and decreases slowly, the integral control is used to eliminate the steady-state error and further optimize the power adjustment.

[0141] Improve the rule base: Ensure that the fuzzy rule base covers all possible input combinations so that the controller can make reasonable decisions under different working conditions; the rules should be coordinated with each other to avoid contradictions or conflicts;

[0142] Step S7-4: Fuzzy reasoning and defuzzification:

[0143] Fuzzy reasoning process: When there is new input data and When the membership function is used to determine their membership to each fuzzy language value, the matching fuzzy rules are found according to the fuzzy rule base, and the activation strength of each rule is determined according to the input membership (usually the minimum value of each input membership is taken); finally, the outputs of all activated rules are synthesized to obtain the fuzzy output result;

[0144] Defuzzification operation: Use appropriate defuzzification methods (such as centroid method, etc.) to convert fuzzy output into accurate power adjustment ; For example, the centroid method obtains the exact value by calculating the centroid of the fuzzy output set, and the formula is ;

[0145] Step S7-5: Power regulation and effect monitoring:

[0146] Adjust the cleaning device power: The power adjustment amount obtained by defuzzification Act on the cleaning device and adjust its power in real time; for example, when When it is a positive value, the power of the cleaning device is increased; when When it is a negative value, the power of the cleaning device is reduced;

[0147] Monitoring and feedback: After power adjustment, the working status of the cleaning device is continuously monitored, including garbage cleaning efficiency, energy consumption and other indicators; according to the monitoring results, the parameters of the fuzzy PID controller are continuously adjusted (such as , , ) to further optimize the power regulation effect and achieve coordinated optimization of energy consumption and efficiency; at the same time, ensure that the response time of the entire regulation process is less than 200ms to meet the requirements of real-time control.

[0148] Step S8 deploys edge-cloud collaborative architecture to support real-time decision making and incremental update of rule base. The specific steps are as follows:

[0149] Step S8-1: Deploy lightweight TSK fuzzy inference model at the edge:

[0150] Model selection and configuration: The lightweight TSK (Takagi-Sugeno-Kang) fuzzy inference model is deployed on the edge device. The model has a relatively simple structure and low computational complexity, making it suitable for running on resource-constrained edge devices. The model is configured to determine the number of input variables based on the hardware performance of the edge device (such as computing power, memory size) and the requirements of the garbage cleaning task. , and initialize the model's rule parameters, i.e. The constant term of the rule and Rule No. The coefficients of the input variables ;

[0151] Data input and reasoning: The edge device obtains data such as images and depth information collected by the multimodal data acquisition unit in real time and uses it as an input vector Input into the lightweight TSK fuzzy inference model; the model calculates the output of each rule according to the preset rules ; By performing weighted summation and other operations on all rule outputs, the final reasoning result is obtained to support real-time decision-making, such as controlling the action of the cleaning device;

[0152] Incremental learning updates parameters: During the garbage cleaning process, as new data is continuously acquired, the edge uses an incremental learning algorithm to update the rule parameters of the TSK fuzzy inference model; for example, when encountering a new garbage distribution scenario or a change in garbage category characteristics, the model adjusts the constant term based on the new data. and coefficient , so that the model can adapt to environmental changes and improve the accuracy of decision-making;

[0153] Step S8-2: Cloud execution density database expansion and synchronization:

[0154] Data collection and analysis: The cloud collects garbage data uploaded by various edge devices every month, including garbage images, volume and mass estimation data, and information on garbage samples found during the cleaning process that are difficult to match existing density values. The cloud analyzes this data to identify possible new garbage categories or changes in the density values ​​of existing garbage categories.

[0155] Database expansion: Based on the results of data analysis, new garbage category density values ​​will be added Add to the density database; for example, if a new type of plastic waste is discovered, after determining its density value through experimental measurement or reference to relevant data, the value is entered into the database to enrich the content of the database and provide more accurate data support for subsequent waste mass estimation;

[0156] TA synchronization to edge devices: Use OTA (Over-The-Air) technology to synchronize the updated density database to each edge device; after receiving the update information, the edge device automatically downloads and updates the local density database to ensure that the edge can use the latest density data when estimating the garbage mass based on three-dimensional sensing technology, thereby improving the accuracy of garbage mass estimation.

[0157] See also Figure 2 , a neuro-fuzzy based adaptive garbage cleaning system, comprising:

[0158] Multimodal data acquisition unit: integrates RGB-D camera and LiDAR to collect image and depth information;

[0159] Hybrid computing unit: deploys a neural fuzzy-based adaptive garbage cleaning method, including an FPGA-accelerated fuzzy reasoning module. The FPGA-accelerated fuzzy reasoning module adopts a parallel architecture to perform hardware acceleration on the fuzzy reasoning process, with a single-frame processing delay of <50ms and power consumption of <5W;

[0160] Dynamic power control unit: supports 50W-150W power adjustment, and the suction nonlinear mapping formula is: ,in, For suction, is power, is the coefficient, is the index, , , maximum suction force ≥ 200N;

[0161] Cloud management platform: provides rule base version control and incremental learning interface;

[0162] Furthermore, the garbage cleaning system integrates multimodal data collection, hybrid computing, dynamic power control and cloud management functions, and realizes intelligent garbage cleaning through multi-technology collaboration, improves cleaning efficiency, reduces energy consumption and adapts to complex scenarios. Specifically:

[0163] Multimodal data acquisition unit: integrated RGB-D camera and LiDAR, which can simultaneously acquire image and depth information; RGB-D camera is responsible for taking color images of garbage, providing the appearance characteristics of garbage, such as color, shape, texture and other information; LiDAR generates point cloud data of garbage by emitting laser beams and receiving reflected signals, accurately acquiring the spatial location and depth information of garbage; the two work together to comprehensively collect multimodal data of garbage, providing rich data support for subsequent garbage identification and volume estimation; in different scenarios, such as urban streets and community garbage dumping sites, data can be stably collected, and the impact of different times and weather on the appearance of garbage is taken into account to ensure the diversity and comprehensiveness of the collected data;

[0164] Hybrid computing unit: Deploys an adaptive garbage cleaning method based on neural fuzzy, the core of which is a fuzzy reasoning module with FPGA acceleration. This module adopts a parallel architecture and uses the hardware parallel characteristics of FPGA to perform hardware acceleration on the fuzzy reasoning process. Compared with traditional software computing methods, it greatly improves the computing speed, with a single-frame processing delay of less than 50ms and power consumption of less than 5W. When processing garbage images, it can quickly integrate fuzzy clustering features with deep learning features, dynamically optimize classification boundaries, and accurately identify garbage categories. By establishing a fuzzy rule library, it can achieve real-time decision-making, efficiently process large amounts of data, provide accurate control instructions for garbage cleaning, and improve the operating efficiency of the entire system.

[0165] Dynamic power control unit: supports 50W-150W power adjustment, dynamically adjusts the cleaning device power according to the output of the fuzzy PID controller; through the suction nonlinear mapping formula ( , ), to achieve precise matching of power and suction, with the maximum suction force not less than 200N; during the garbage cleaning process, the power is dynamically adjusted according to the volume, quality, distribution of the garbage and the real-time working status of the cleaning device; when encountering large or heavy garbage, the power is increased to improve the suction force; when the garbage is small or light, the power is reduced to save energy, thereby achieving the coordinated optimization of energy consumption and efficiency;

[0166] Cloud management platform: provides rule base version control and incremental learning interface. On the one hand, it manages the fuzzy rule base version and records the content and time of each update of the rule base for easy traceability and management. On the other hand, it supports incremental learning. When edge devices upload new garbage data or discover new garbage categories, the cloud uses these data for learning and updates the rule base. It also performs density database expansion every month, adds new garbage category density values, and synchronizes them to edge devices via OTA, so that the entire system can continuously adapt to new garbage types and scenarios, and continuously improve the accuracy and adaptability of garbage cleaning.

[0167] Although 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 the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A neuro-fuzzy based adaptive garbage cleaning method, characterized by: The following steps are involved: S1. Construct a multimodal garbage image dataset by labeling the garbage types and dividing it into training set, validation set and test set; S2. Target detection model training based on multimodal feature fusion, integrating fuzzy clustering features and deep learning features, and dynamically optimizing classification boundaries; S3, fuzzy processing is performed on the garbage image, a membership matrix is ​​generated by combining the fuzzy clustering algorithm, and a threshold is set to determine the garbage category; S4. Establish a fuzzy rule base and realize real-time decision-making through fuzzy reasoning model; S5. Use hybrid optimization strategy to train the model, combining global search and second-order precision tuning to improve model efficiency and convergence speed; S6, real-time estimation of garbage volume and mass based on 3D sensing technology, matching physical parameters with density database; S7, dynamically adjust the power of the cleaning device through the fuzzy PID controller to achieve coordinated optimization of energy consumption and efficiency; S8. Deploy edge-cloud collaborative architecture to support real-time decision-making and incremental updates of rule bases.

2. The neural fuzzy-based adaptive garbage cleaning method according to claim 1, characterized in that: In step S2, the multimodal feature fusion specifically includes: The membership matrix generated by FCM fuzzy clustering Feature map output by YOLOv8 Channel cascading is performed, where The membership matrix generated for FCM fuzzy clustering is, is the number of clusters, is the sample size, It is the feature map output by YOLOv8. is the height of the feature map, is the width of the feature map, is the number of channels of the feature map, and the feature weight is calculated through the dynamic weight allocation module. The formula is: ,in, For the The feature weights of clusters, For the The membership degree of a cluster, For the The membership degree of a cluster, is the temperature coefficient, is the number of clusters; Weighted fusion features Input to the detection head for classification, where is the weighted fusion feature, For the corresponding The feature map of the clusters.

3. The neural fuzzy-based adaptive garbage cleaning method according to claim 1, characterized in that: The fuzzy clustering algorithm in step S3 is FCM, and its membership update formula is: ,in, For the The samples belong to The membership degree of a cluster, For the samples, For the Cluster centers, Representative Cluster centers, is the fuzzy index; set up , membership threshold The classification accuracy was >92% as verified by the confusion matrix.

4. The neural fuzzy-based adaptive garbage cleaning method according to claim 1, characterized in that: The hybrid optimization strategy in step S5 includes: Grid search phase: In the parameter space Traverse the combination internally and filter out the errors below the threshold Candidate parameters of and is the parameter to be optimized, is the error threshold; LM optimization stage: perform second-order optimization based on candidate parameters, and the update formula is: ,in, is the parameter update amount, is the Jacobian matrix of the error versus parameter, is the Jacobian matrix The transposed matrix of is the damping factor, is the identity matrix, is the error vector, the initial value of the damping factor , the dynamic adjustment range is .

5. The neural fuzzy-based adaptive garbage cleaning method according to claim 1, characterized in that: In step S6, the three-dimensional volume estimation formula is: ,in, is the three-dimensional volume of the garbage; is the number of pixels; For pixels in The physical size of the direction, , For pixels in The physical size of the direction, , is the physical width of the sensor, is the width of the image, is the physical height of the sensor, is the height of the image; For the The depth value of pixels; the quality is calculated as , For the quality of garbage, is the density of garbage, Match from a preset database, which contains 20 categories of garbage density values.

6. The neural fuzzy-based adaptive garbage cleaning method according to claim 1, characterized in that: In step S7, the rule base of the fuzzy PID controller includes: Rule 1: For "Taisho" and For "positive", THEN ,in, is the error at the current moment; is the rate of change of error; is the power adjustment amount; is the proportionality coefficient; is the differential coefficient; Rule 2: is "Medium" and is "zero", THEN ; Rule 3: is a "small negative" and For "negative slowness", ,in, is the integration coefficient; is the integral of the error; The membership function adopts Gaussian type and is defined as ,in, Error The degree of membership belongs to "big", is the center of the Gaussian membership function, is the standard deviation of the Gaussian membership function, , , response time <200ms.

7. The neural fuzzy-based adaptive garbage cleaning method according to claim 1, characterized in that: The edge-cloud collaborative architecture in step S8 includes: The lightweight TSK fuzzy inference model is deployed on the edge, and its rule output is: ,in, For the The output of the rule, is the input vector; For the The constant term of the rule; For the Rule No. The coefficients of the input variables are updated through incremental learning. is the number of input variables; The cloud performs density database expansion every month, adding new garbage category density values and synchronized to the edge device via OTA.

8. A neuro-fuzzy based adaptive garbage cleaning system, characterized by: include: Multimodal data acquisition unit: integrates RGB-D camera and LiDAR to collect image and depth information; Hybrid computing unit: deploying the method described in claims 1-7, including an FPGA-accelerated fuzzy reasoning module, the FPGA-accelerated fuzzy reasoning module adopts a parallel architecture, performs hardware acceleration on the fuzzy reasoning process, and has a single-frame processing delay of <50ms and a power consumption of <5W; Dynamic power control unit: supports 50W-150W power adjustment, and the suction nonlinear mapping formula is: ,in, For suction, is power, is the coefficient, is the index, , , maximum suction force ≥ 200N; Cloud management platform: provides rule base version control and incremental learning interface.

Images