Collision identification method and device of window cleaning robot

Abstract

The invention relates to the technical field of collision recognition, and discloses a collision recognition method and device for a window cleaning robot. The method comprises the following steps: acquiring displacement data of a radial displacement sensor mounted at a side wheel of the window cleaning robot and state data of a photoelectric drop detector mounted at the bottom of the robot to obtain original data of a bimodal sensor; generating a collision judgment result based on the original data of the bimodal sensor; obstacle avoidance path parameters suitable for the current collision condition are calculated, and a path point sequence is generated; the path point sequence is transmitted into a motion control system of the window cleaning robot in real time, the running speed and direction of the robot are adjusted in a self-adaptive mode according to path tracking errors and the window face adsorption state, and intelligent obstacle avoiding and continuous cleaning operation execution are achieved. According to the invention, the limitation of a traditional single sensing mode is overcome, high sensitivity and accuracy of collision detection are maintained under various window materials and environmental conditions, and the cleaning coverage rate and the operation continuity are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of collision recognition, and in particular to a collision recognition method and device for a window cleaning robot. Background Art

[0002] Traditional window-cleaning robots primarily rely on a single sensing method for collision detection, such as a simple contact switch or a single displacement sensor. In practice, this single-mode sensing approach suffers from insufficient sensitivity, susceptibility to external interference, and low accuracy in identifying the window edge area. Especially when the robot operates on glass surfaces of varying materials or in complex window frame structures, a single sensor struggles to accurately distinguish between collisions with the window frame and those with obstacles, resulting in a high rate of false positives and reduced work efficiency.

[0003] Conventional window-cleaning robot collision recognition algorithms typically lack the ability to comprehensively analyze the spatiotemporal characteristics of collision signals, failing to effectively utilize the signal's continuity across both time and space. This results in robots being susceptible to external factors such as window vibration, wind disturbance, and surface friction in complex environments, leading to false collision detections or missed collisions. Furthermore, existing collision recognition methods often rely on prior information about window structure, lacking adaptability to diverse window environments and making them difficult to meet diverse cleaning needs. Summary of the Invention

[0004] The main purpose of the present invention is to provide a collision recognition method and device for a window cleaning robot. The present invention overcomes the limitations of the traditional single sensing method, enables collision detection to maintain high sensitivity and accuracy under various window materials and environmental conditions, and improves cleaning coverage and operation continuity.

[0005] To achieve the above object, the present invention provides a collision recognition method for a window cleaning robot, comprising the following steps: Obtain the displacement data of the radial displacement sensor installed on the side wheel of the window cleaning robot and the status data of the photoelectric drop detector installed on the bottom of the robot to obtain the raw data of the dual-modal sensor; Performing similarity matching on the raw data of the dual-modal sensor with a preset window frame collision feature template library and an obstacle collision feature template library to obtain a collision determination result; Analyzing the force distribution and trigger timing differences of multiple side wheels according to the collision determination results and calculating obstacle avoidance path parameters suitable for the current collision situation to generate a path point sequence; The path point sequence is transmitted to the motion control system of the window cleaning robot in real time, and the robot's running speed and direction are adaptively adjusted according to the path tracking error and the window surface adsorption state, so as to realize intelligent obstacle avoidance and continuous execution of the cleaning operation.

[0006] The present invention also provides a collision recognition device for a window cleaning robot, comprising: An acquisition module is used to acquire the displacement data of the radial displacement sensor installed on the side wheel of the window cleaning robot and the status data of the photoelectric drop detector installed at the bottom of the robot to obtain the raw data of the dual-modal sensor; A similarity matching module is used to perform similarity matching on the raw data of the dual-modal sensor with a preset window frame collision feature template library and an obstacle collision feature template library to obtain a collision determination result; A calculation module, configured to analyze the force distribution and trigger timing differences of the multiple side wheels according to the collision determination results, calculate obstacle avoidance path parameters suitable for the current collision situation, and generate a path point sequence; The adaptive adjustment module is used to transmit the path point sequence to the motion control system of the window cleaning robot in real time, and adaptively adjust the robot's running speed and direction according to the path tracking error and the window surface adsorption state, so as to realize intelligent obstacle avoidance and continuous execution of cleaning operations.

[0007] In summary, the technical solution provided by the present invention forms complementary sensing capabilities through the coordinated work of the edge wheel radial displacement sensor and the photoelectric drop detector, overcomes the limitations of the traditional single sensing method, and enables collision detection to maintain high sensitivity and accuracy under various window materials and environmental conditions. By adopting the spatiotemporal feature fusion matrix technology, the continuity of the collision signal in the time and space dimensions is used as a priori constraint, effectively suppressing environmental vibration and noise interference, and significantly improving the recognition stability in harsh environments. The template matching method based on the dynamic time warping algorithm can accurately distinguish different types of collisions such as horizontal window frames, vertical window frames, corner window frames, hard obstacles, soft obstacles, etc., providing an accurate basis for the subsequent obstacle avoidance strategy selection. Through the comprehensive analysis of the force distribution of the edge wheels and the difference in trigger timing, the collision position, angle and force are accurately calculated to form a comprehensive collision state vector. The boundary following or obstacle avoidance strategy is automatically selected according to the collision type, and a smooth obstacle avoidance path is constructed through the Bezier curve, realizing specialized processing for different collision situations and avoiding the limitations of the traditional fixed obstacle avoidance mode. It adopts an adaptive control method based on fuzzy logic, dynamically adjusts operating parameters according to path tracking error and window adsorption status, and has the ability to recover from working status, ensuring that the robot can seamlessly connect to the original cleaning task after completing obstacle avoidance, thereby improving cleaning coverage and operation continuity. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 1 is a schematic diagram of the steps of a collision identification method for a window cleaning robot according to an embodiment of the present invention; Figure 2 This is a structural block diagram of a collision recognition device of a window cleaning robot in one embodiment of the present invention.

[0009] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0010] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to 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.

[0011] Reference Figure 1 This embodiment provides a collision recognition method for a window cleaning robot, comprising the following steps: S1, obtain the displacement data of the radial displacement sensor installed on the side wheel of the window cleaning robot and the status data of the photoelectric drop detector installed on the bottom of the robot to obtain the raw data of the dual-modal sensor; High-precision radial displacement sensors are installed on the inner wall of the window-cleaning robot's four side wheel sockets. The sensing rod of each sensor maintains constant physical contact with the inner side of its corresponding side wheel. When the side wheel is subjected to external forces (such as striking the window frame or contacting an obstacle), the force causes a tiny radial mechanical displacement. This displacement is instantly transmitted to the core component of the displacement sensor via the sensing rod and quickly converted into a continuous, tiny electrical signal. Simultaneously, infrared photoelectric drop detectors capable of emitting light at a preset wavelength are installed at the four corners of the robot's base. These detectors consist of a transmitter and a receiver. The transmitter emits infrared light, which is reflected by the window surface, and the receiver detects changes in the reflected light intensity in real time. If the robot encounters high-risk operating conditions such as hanging in the air, deviating from the edge, or partially leaving the window surface, the reflected light intensity will significantly decrease, and the detector will output a corresponding status signal. These two signals are simultaneously input into a high-precision ADC module for digitization. The displacement sensor electrical signal is encoded as numerical displacement data, while the photoelectric detector status signal is converted into a clear digital value, reflecting the current safe operating condition. Displacement data and state data are time-series aligned, organizing them into a multi-channel data stream arranged in a time series. This allows data from different sensors within the same time segment to collaboratively reflect the robot's motion state and environmental changes. Filtering algorithms, such as a Butterworth low-pass filter, are used to suppress noise in the time series data stream. By properly setting the cutoff frequency, high-frequency interference signals are effectively filtered out, preserving the intrinsic features associated with collision or fall events. Based on the filtered signal, dynamic baseline correction is performed to compensate for systematic errors such as zero drift and temperature drift generated by the sensor in real time. The sensor output baseline is continuously corrected, making the features obtained in the subsequent recognition process more representative and stable. This results in raw dual-modal sensor data.

[0012] S2, performing similarity matching on the raw data of the dual-modal sensor with the preset window frame collision feature template library and the obstacle collision feature template library to obtain a collision determination result; Specifically, the collected radial displacement sensor data from the side wheels is spatially distributed and structured into displacement vectors. Within each sampling period, the displacement state of each side wheel is recorded as a vector component. Simultaneously, the digitized state data from the four photoelectric drop detectors is combined into a state vector, synchronously reflecting the attachment or suspension status of each position on the robot's bottom. The displacement and state vectors within consecutive sampling periods are segmented according to a preset time window length and then fused and aligned to form an initial spatiotemporal feature matrix. This matrix reflects the dynamics of the robot's motion and its environmental response in both time and space. Based on this initial matrix, a time-domain differential displacement vector, reflecting the rate of change of the side wheel displacement, and a time-domain differential state vector, reflecting sudden changes in the drop detector state, are obtained by performing a differential operation on the data between adjacent sampling points. The original spatiotemporal feature matrix is ​​then concatenated and fused with the two differential vectors in the feature dimension to produce an enhanced spatiotemporal feature matrix. Principal component analysis is performed on the enhanced spatiotemporal feature matrix. Principal components are extracted and ranked through linear transformation. Principal components whose cumulative explained variance reaches a preset threshold (e.g., 95%) are retained. Redundant noise features are removed, and the enhanced spatiotemporal feature matrix is ​​projected onto a reduced feature subspace to form a representative feature fusion matrix. Similarity matching is performed on the feature fusion matrix against a pre-defined library of window frame and obstacle collision feature templates. Flexible time series alignment algorithms, such as dynamic time warping, are employed to achieve flexible matching of complex collision signals in terms of temporal and amplitude variations. By comparing the matching similarity results, the collision category closest to the currently acquired signal is output, resulting in the collision determination result.

[0013] In this embodiment, the feature fusion matrix is ​​resampled at different time scales, that is, the feature vectors of the original time resolution are mapped into multiple sets of multi-scale feature representations with different sampling densities or sequence lengths by scaling or sampling methods. The multi-scale feature representations are respectively preliminarily matched with the preset window frame collision feature template library. The window frame collision template library is subdivided into three types of templates: horizontal window frames, vertical window frames, and corner window frames. Each type of template is derived from a large amount of measured data. By comparing the similarity between the multi-scale features and the three types of window frame templates, the set of candidate window frame collision templates that are closest to the current collision event is preliminarily screened out, narrowing the search space for subsequent precise matching. Similarly, the multi-scale features are preliminarily matched with the obstacle collision feature template library. The obstacle templates are divided into two categories: hard obstacles and soft obstacles. The set of candidate obstacle collision templates is obtained by comparing the similarity. For each of the two candidate template sets, a dynamic time warping algorithm is used to precisely align the feature fusion matrix with each candidate template. This algorithm addresses the elastic deformation of collision signals in duration and response speed, thereby calculating the window frame collision and obstacle collision similarity measures for each template. During this process, temporal characteristics of the collision signal, such as collision duration, peak amplitude, and signal decay rate, are extracted from the feature fusion matrix to serve as a basis for weighting the similarity measures. By dynamically weighting the similarity measures using temporal characteristics, the system more accurately reflects the degree of match between the actual collision event and the templates. Corrected match probabilities for window frame collision and obstacle collision are calculated, reflecting the likelihood that the current collision signal is attributable to a window frame collision or an obstacle collision, respectively. Based on these two corrected match probabilities, the one with the highest probability is selected as the final collision decision, and the corresponding type label and confidence score are output.

[0014] S3, analyzing the force distribution and trigger timing differences of multiple side wheels based on the collision judgment results, and calculating obstacle avoidance path parameters suitable for the current collision situation to generate a path point sequence; It should be noted that the collision type is determined based on the collision judgment results, clarifying whether the current collision is a window frame collision, an obstacle collision, or an unknown type. Each collision event corresponds to a different obstacle avoidance priority and action mode. The displacement data of the four side wheels is then weighted and averaged, combined with the actual spatial distribution coordinates of the side wheels. The center coordinates of the collision force are calculated by weighted average of the displacement amplitudes. The time of the first triggering of each side wheel radial displacement sensor is recorded and these triggering moments are arranged in chronological order to form a time series that reflects the order of collision propagation. Using the center coordinates of the collision force and the time series, the actual relative positions of the robot, the window frame, and the obstacle are triangulated to locate the target location of the collision point. This process utilizes the improved spatial positioning accuracy achieved by multi-point information synergy. Based on the relative relationship between the target location of the collision point and the robot's center of mass, the spatial angle between them is calculated, thereby deducing the collision incident direction and converting it into a collision angle value. The maximum value of the radial displacement data of all side wheels is extracted, and the distribution of the collision force across each side wheel is analyzed. The stiffness coefficient is automatically adjusted based on the actual collision type to comprehensively determine the collision force value. The collision state information is integrated into the collision type, collision point target location, collision angle, and collision force, reflecting the spatial, temporal, and mechanical characteristics of the collision event. Based on this collision state information, the built-in obstacle avoidance decision-making algorithm dynamically selects path planning modes such as boundary following and obstacle avoidance, appropriately sets path parameters such as obstacle avoidance distance, turning radius, and target motion direction, and combines smooth path generation techniques such as Bezier curves to form a set of path point sequences.

[0015] In this embodiment, the obstacle avoidance strategy is determined based on the collision type and target location of the collision point in the collision status information. In the case of a window frame collision, the system automatically switches to a boundary-following strategy. This means that when the robot approaches the window edge and collides with it, it uses the obstacle avoidance distance as a parameter and moves parallel to the window frame boundary to avoid the risk, while maintaining a certain safety distance to prevent secondary collisions. If the collision type is determined to be an obstacle collision, the system adopts an obstacle avoidance strategy. This strategy automatically adjusts the obstacle avoidance radius based on the collision force value and determines the direction of detour (e.g., clockwise or counterclockwise) based on the collision angle and the spatial orientation of the obstacle, to maximize obstacle avoidance and optimize the motion trajectory. Based on the determined obstacle avoidance strategy type, the corresponding parameters are precisely calculated. For the boundary-following strategy, the obstacle avoidance distance and direction depend on the robot's current relative position to the window frame and the direction of the collision. The robot's trajectory must maintain a safe distance from the window frame while ensuring clean coverage. For the obstacle-avoiding strategy, the obstacle avoidance radius is directly related to the collision force. A greater force indicates a harder or larger obstacle, leading to a larger detour radius. The robot also dynamically determines an appropriate obstacle avoidance direction based on the actual environment to ensure the path is not bogged down or repeatedly backed off by local obstacles. All obstacle avoidance control parameters are aggregated and fed into the path planning module. During the path generation phase, the system uses the current position as the starting point, the expected return path point in the cleaning operation plan as the end point, and the positions of intermediate control points are designed based on the parameters determined by the obstacle avoidance strategy. These three points together form the set of key control points of a quadratic Bezier curve. For the boundary-following strategy, the intermediate control points are offset appropriately along the normal direction of the window frame, while for the obstacle-avoiding strategy, the intermediate control points are located at the spatial angle corresponding to the obstacle avoidance radius. This ensures a smooth, sharp-angle-free path while also meeting obstacle avoidance requirements. Using the parametric equation of a Bezier curve, the control point set is mapped into a continuous, curved initial obstacle avoidance path, effectively eliminating the abruptness and instability of traditional broken-line obstacle avoidance and improving the smoothness of the robot's motion. Based on the initial obstacle avoidance path, a comprehensive path evaluation function is constructed using path length, obstacle safety distance, and path smoothness as evaluation metrics. A gradient descent algorithm is then used to iteratively adjust and fine-tune the positions of intermediate control points. Each iteration minimizes the path distance while ensuring safety distance and smooth trajectory, resulting in a globally optimal obstacle avoidance path. The optimal obstacle avoidance path is discretized and sampled, generating a sequence of path points through equally spaced sampling.

[0016] S4 transmits the path point sequence to the motion control system of the window cleaning robot in real time, and adaptively adjusts the robot's running speed and direction according to the path tracking error and the window surface adsorption state, realizing intelligent obstacle avoidance and continuous execution of cleaning operations.

[0017] Specifically, a fuzzy logic controller was constructed with path deviation, deviation change rate, and window adhesion as input variables, and the left and right motor speed difference and forward speed as output variables, forming the basic controller structure. In this structure, the path deviation reflects the vertical distance between the robot's actual path and the target path, the deviation change rate dynamically reflects the current path correction trend, and the window adhesion reflects the robot's secure attachment to the window surface by collecting real-time data from the negative pressure system. To achieve efficient and intelligent mapping between input and output, a reasonable fuzzy membership function was assigned to each input variable. The path deviation was divided into linguistic values ​​such as negative large, negative small, zero, positive small, and positive large. The deviation change rate and adhesion were further subdivided into linguistic intervals such as small, medium, and large. A multidimensional fuzzy rule base was established based on operational experience. For each actual input at each moment, the fuzzy membership function and rule base were defuzzified using the center of gravity method, resulting in clear and adaptive motor control instructions. In terms of path tracking control, the system uses the real-time pathpoint sequence as a reference and determines the nearest target pathpoint based on the robot's current position. Spatial geometry calculations are used to determine the straight-line distance and azimuth angle between the current position and the target point, generating a path tracking error vector that includes lateral, longitudinal, and directional errors. This vector drives the fuzzy controller to generate new motor commands. Based on the latest path tracking error vector and the fuzzy controller output, the speeds of the left and right motors are differentially adjusted, ensuring forward speed while precisely controlling the steering control amount. This enables the robot to efficiently correct path deviations and closely follow the pre-set obstacle avoidance path, improving obstacle avoidance and task consistency. Simultaneously, data streams from the side wheel displacement sensors and the negative pressure system are monitored in real time to continuously assess the robot's force state and adhesion security during operation. If abnormal side wheel displacement fluctuations are detected or the negative pressure value drops below a safety threshold, a safety protection mechanism is triggered, imposing hard limits on forward speed and steering range, or even initiating emergency braking, to prevent equipment damage or operation interruption due to instability, slippage, or collision. Motion control commands are generated within safety constraints, enabling intelligent obstacle avoidance and continuous cleaning operations.

[0018] In one example, the displacement data of the radial displacement sensor installed on the side wheel of the window cleaning robot and the status data of the photoelectric drop detector installed on the bottom of the robot are obtained to obtain the dual-modal sensor raw data, including: Install radial displacement sensors on the inner wall of the four side wheel sockets of the window cleaning robot, and keep the sensing rod in contact with the inner side of the side wheel to obtain the radial displacement electrical signal generated when the side wheel is subjected to external force; Photoelectric drop detectors that emit infrared light of a preset wavelength are installed at the four corners of the bottom of the window cleaning robot. The changes in the intensity of the reflected light are detected to obtain status signals when the robot is suspended in the air or partially leaves the window surface. The radial displacement electrical signal and the status signal are digitally processed respectively to obtain the displacement data of the radial displacement sensor and the status data of the photoelectric drop detector, and the displacement data and the status data are time-series aligned to obtain a time-series continuous sensor data stream; The time-series continuous sensor data stream is filtered to obtain target filtered data, and dynamic baseline correction is performed on the target filtered data to obtain dual-modal sensor raw data.

[0019] In this example, highly sensitive radial displacement sensors are installed on the inner walls of the four side wheel sockets. Each radial displacement sensor utilizes capacitive or inductive micro-displacement detection principles, sensing radial movement of the side wheels under minute forces with extremely high resolution. To ensure real-time mechanical response and signal accuracy, the sensing rod maintains moderate elastic contact with the inner surface of each side wheel. Regardless of whether the robot is stationary or in motion, any slight deflection of the side wheels caused by external forces (such as impact with the window frame, obstacles, or vibration) is instantly transmitted to the sensor's core detection unit. The sensor internally converts this mechanical displacement into a continuous analog electrical output signal. The signal amplitude and rate of change reflect the actual force intensity and its dynamic evolution. The four sets of radial displacement sensors respond independently, generating four raw sensing signals with high sampling rate and strong interference resistance. In addition, infrared photoelectric drop detectors are installed at the four corners of the window cleaning robot's base. Each detector consists of a highly stable infrared transmitter and a sensitive infrared receiver. The transmitter continuously emits an infrared beam of a specific wavelength with strong penetration and interference resistance, making it suitable for surface distance monitoring in complex external lighting environments. The robot's working window reflects infrared light, and the receiver measures the intensity of the reflected infrared light in real time. When properly attached to the window, the received reflected signal intensity remains high. However, if the robot partially floats, deviates, or encounters a dangerous operating condition such as separation from the window, the reflected light intensity at the corresponding detection point will drop sharply or even disappear. The detector converts this reflected signal change into an explicit status output, distinguishing between different scenarios such as "normal attachment," "edge violation," and "dangling fall risk." The radial displacement electrical signal and status signal are digitized separately. All analog electrical signals (including analog voltages representing minute displacements and analog light intensity signals from the infrared receiver) are input into a multi-channel, high-speed, high-resolution analog-to-digital converter (ADC). Each signal is discretized at a fixed high sampling frequency, converting the continuous electrical signal into a standard digital data stream. After sampling by the ADC, the radial displacement sensor outputs a finely discretized displacement data stream. The photoelectric detector, through level discrimination and threshold setting, digitizes the light intensity changes into binary or multi-level status signals such as "attached" and "dangling." All collected digital data streams are synchronized according to a unified system clock timestamp, ensuring that the robot's sensory states at all four corners and all side wheels correspond at any given moment. Multi-channel data within a sampling period is organized into a time series matrix. Through time alignment, each frame of data contains synchronized sample values ​​for all displacement and status signals, constructing a time-continuous multimodal sensor raw data stream. Multi-stage filtering is applied to this time-continuous multi-channel data stream to minimize the impact of noise and non-target signals on the analysis results.A digital low-pass filter, such as a Butterworth filter, is selected with an appropriate cutoff frequency to effectively filter out high-frequency irrelevant components caused by mechanical vibration and ambient electromagnetic noise, while preserving the physical characteristics of key events such as collisions and drops to the greatest extent possible. The filter order and parameters are adjusted for different signal sources (such as displacement and photoelectric signals) to adapt to their respective noise spectrum characteristics. The filtered signal is defined as the target filtered data. Dynamic baseline correction is performed on the target filtered data. The baseline of each signal is dynamically corrected using algorithms such as preset static calibration points, sliding window averaging, or real-time baseline tracking. For example, the sensor output in a static state is periodically collected as a baseline, or the current baseline level is automatically updated when there is no significant signal disturbance. Furthermore, adaptive compensation mechanisms such as weighted averaging and recursive least squares are used to continuously correct for extreme value drift within a short-term window, correcting for zero-point drift and nonlinear response of the sensor output curve, ultimately generating raw dual-modal sensor data.

[0020] In one example, the raw data of the dual-modal sensor is similarly matched with a preset window frame collision feature template library and an obstacle collision feature template library to obtain a collision determination result, including: According to the original data of the dual-modal sensor, the displacement data is constructed into a displacement vector, and the state data of the photoelectric drop detector is constructed into a state vector; The displacement vector and state vector are segmented and fused into continuous sampling periods according to the preset time window length to obtain the initial spatiotemporal feature matrix; Perform differential calculation on the data of adjacent sampling points in the initial spatiotemporal feature matrix to obtain the time-domain differential displacement vector representing the displacement change rate and the time-domain differential state vector representing the state change; The initial spatiotemporal feature matrix is ​​spliced ​​and fused with the time domain differential displacement vector and the time domain differential state vector to obtain an enhanced spatiotemporal feature matrix; Perform principal component analysis on the enhanced spatiotemporal feature matrix to obtain a feature subspace after dimensionality reduction, retain the principal components whose cumulative explained variance reaches a preset threshold in the feature subspace after dimensionality reduction, and perform projection transformation on the enhanced spatiotemporal feature matrix to obtain a feature fusion matrix; The feature fusion matrix is ​​matched with the preset window frame collision feature template library and obstacle collision feature template library for similarity to obtain the collision judgment result.

[0021] In this example, the collected raw sensor data is organized. The radial displacement sensor outputs of the four side wheels at each moment are assembled according to their spatial arrangement into a four-dimensional displacement vector, representing the force distribution at the robot's four key contact points at that specific moment. Simultaneously, the state outputs of the photoelectric drop detectors at the four corners of the base are similarly assembled into a four-dimensional state vector, with each component taking the value 0 or 1, representing the current state of the detection point attached or suspended on the window surface. In this way, all raw data streams are converted into a time-varying high-dimensional vector sequence. Based on the duration of the actual collision event, the sensor response characteristics, and the real-time requirements of the task, a time window length is preset. The displacement and state vectors of all consecutive sampling periods are segmented. The data within each time window are then fused and aligned to form an initial spatiotemporal feature matrix. Each row of this matrix corresponds to a sampling moment, and each column represents the side wheel displacement and base state data. The entire matrix reflects the changes in the motion and attachment state of each key part of the robot during that time period. The data between any adjacent sampling points in the initial spatiotemporal feature matrix are differentially calculated. For each time point, the displacement difference between the current and previous sampling periods is calculated to form a new time-domain differential displacement vector, which reflects the rate and trend of change in the force on each side wheel. Similarly, the change in each component of the state vector is calculated to obtain a time-domain differential state vector. This effectively captures transient signals, such as when a drop detector suddenly changes from "attached" to "suspended" state, facilitating the timely identification of high-risk collision and drop events. The initial spatiotemporal feature matrix is ​​concatenated with the time-domain differential displacement vector and the time-domain differential state vector along the feature dimension to generate an enhanced spatiotemporal feature matrix. Feature extraction from the enhanced spatiotemporal feature matrix is ​​performed using dimensionality reduction algorithms such as principal component analysis (PCA). PCA calculates the covariance matrix between samples, solves for their eigenvalues ​​and eigenvectors, and ranks all original features according to their contribution to the overall variance. Only principal components whose cumulative explained variance reaches a preset threshold, such as 95%, are retained. This significantly compresses the feature space dimensionality while maximally preserving the core discriminant information in the original data. A projection transformation is then performed on the enhanced spatiotemporal feature matrix, mapping all original high-dimensional signals to a low-dimensional feature subspace composed of principal components, resulting in a feature fusion matrix. The reduced-dimensional feature fusion matrix is ​​then input into the preset window frame collision feature template library and obstacle collision feature template library for similarity matching. Each type of template in the template library is a feature subspace vector collected, standardized, and normalized from a large number of actual collisions, obstacle interference, and other events, representing the intrinsic characteristics of various typical collisions in terms of spatiotemporal and dynamic performance. Similarity matching uses algorithms such as Euclidean distance, cosine similarity, or dynamic time warping to calculate the similarity scores between the current feature fusion matrix and all templates. By comparing the template type with the maximum similarity, it is determined whether the event is a window frame collision, an obstacle collision, or an unknown type, and the collision judgment result is obtained.

[0022] In one example, similarity matching is performed between the feature fusion matrix and a preset window frame collision feature template library and an obstacle collision feature template library to obtain a collision determination result, including: Resample the feature fusion matrix at different time scales to obtain multi-scale feature representation; Perform a preliminary match between the multi-scale feature representation and the preset window frame collision feature template library, calculate the similarity of the horizontal window frame collision template, the vertical window frame collision template, and the corner window frame collision template, and obtain a set of window frame collision candidate templates; Perform a preliminary match between the multi-scale feature representation and the preset obstacle collision feature template library, calculate the similarity between the hard obstacle collision template and the soft obstacle collision template, and obtain a set of obstacle collision candidate templates; Execute the dynamic time warping algorithm on the feature fusion matrix and the window frame collision candidate template set and the obstacle collision candidate template set respectively to obtain the window frame collision similarity measurement value and the obstacle collision similarity measurement value; Extract the time series characteristic index of the collision signal from the feature fusion matrix, and perform weighted correction on the window frame collision similarity measurement value and the obstacle collision similarity measurement value based on the time series characteristic index to obtain the window frame collision correction matching probability and the obstacle collision correction matching probability; The collision determination result is generated based on the window frame collision correction matching probability and the obstacle collision correction matching probability.

[0023] In this example, multi-scale feature representations are extracted and generated through resampling operations at different time scales. Specifically, based on the original feature fusion matrix, the time series data stream is downsampled or aggregated according to different sampling intervals or data window lengths, allowing the same collision event to be represented as multiple levels of data representation, such as fine-grained, medium-grained, and coarse-grained. The multi-scale feature representations are then preliminarily matched against a pre-set library of window frame collision feature templates. This library is divided into three categories: horizontal, vertical, and corner window frames. Each template category is derived from a large number of real-world, normalized, and standardized feature subspaces of typical collision events. During the matching process, the multi-scale feature sequences are rapidly and batch-compared with each template type using basic metrics such as Euclidean distance, cosine similarity, and correlation coefficient. A similarity score is calculated for each time scale and each window frame template type, and the template type with the highest similarity at each scale is selected to form a set of candidate window frame collision templates containing candidate templates at multiple time scales. Similarly, the multi-scale feature representations are input into the obstacle collision feature template library for preliminarily matching. The obstacle template library is divided into at least two categories: hard and soft obstacles. Each template is derived from field-measured data processed, normalized, and dimensionality reduced. Multi-scale features are rapidly matched against the hard and soft obstacle templates, respectively. The templates with the highest similarity at each scale are selected to form a set of candidate obstacle collision templates. For the selected window frame collision candidate templates and obstacle collision candidate templates, a dynamic time warping algorithm is used to perform deep matching between the feature fusion matrix and each candidate template. The dynamic time warping algorithm intelligently and elastically aligns time series signals, effectively addressing scale inconsistencies and local stretching issues associated with event duration and response speed. Through dynamic time warping comparison, similarity measurements for window frame and obstacle collisions are obtained. These values ​​reflect the degree of global dynamic morphological match between the event to be identified and each template. Furthermore, the dynamic time warping matching results can distinguish between the force curves of short, high-intensity impacts and those of long, slow impacts, refining the granularity of type discrimination. The time series characteristic indicators of the collision signal are extracted from the feature fusion matrix, including collision duration, maximum amplitude, signal fluctuation frequency, peak occurrence time, and decay rate. These indicators reflect the actual collision physical process at the signal level. Different types of collision events have significant differences in their performance in these indicators. For example, window frame collisions have a shorter duration but a larger amplitude, while obstacle collisions last longer and have a smoother signal. Based on the time series characteristic indicators, the window frame collision similarity measurement values ​​and obstacle collision similarity measurement values ​​output by dynamic time warping are weighted and corrected. That is, a weight coefficient is set, and the time series characteristics and similarity are combined to form a corrected matching probability that better reflects the physical nature.For example, methods such as linear weighting, Bayesian posterior probability adjustment, and fuzzy normalization are used to superimpose the degree of conformance of a certain type of collision on the timing indicators onto the similarity discrimination score, ensuring that the final probability value reflects both the consistency of template matching and the physical rationality of the actual signal process. After the calculation of the corrected matching probability of the window frame collision and the corrected matching probability of the obstacle collision are completed, a comprehensive comparison is performed between the two, and the collision type with the highest probability that exceeds the preset threshold is selected as the judgment result for this event. If neither probability reaches the threshold, the event is judged as "unknown type" and its characteristics are automatically recorded for subsequent template library expansion and system self-learning upgrades.

[0024] In one example, based on the collision determination results, the force distribution and trigger timing differences of multiple side wheels are analyzed, and obstacle avoidance path parameters suitable for the current collision situation are calculated to generate a path point sequence, including: The collision type is determined based on the collision judgment result, and the displacement data of the four side wheels are weighted to obtain the coordinates of the collision force center. The triggering time of the radial displacement sensors of the four side wheels is recorded and sorted to obtain a time series. Based on the coordinates of the collision force center and the time series, the collision position is triangulated and the target position of the collision point is obtained; Based on the relative relationship between the target position of the collision point and the center of mass of the robot, the collision incident angle is calculated to obtain the collision angle value. The maximum displacement value of all side wheels is extracted and the stiffness coefficient is adjusted based on the characteristics of the collision type to obtain the collision force value. Integrate the collision type, collision point target position, collision angle value and collision force value into collision status information; Based on the collision state information, the obstacle avoidance path parameters suitable for the current collision situation are calculated and a path point sequence is generated.

[0025] In this example, sensor data analysis helps distinguish collision types, such as window frame collisions, obstacle collisions, or other special categories. The robot collects real-time displacement data from its four side wheel sensors. The spatial position and force data for each side wheel are systematically integrated. This data takes into account the wheel's actual installation position within the robot structure and its displacement amplitude at each moment. Through an overall weighted analysis, the center of force applied to the collision is inferred, thereby determining the dominant direction and impact area of ​​the external impact force on the entire robot. The robot also records the moment when each side wheel displacement sensor first detects a significant displacement change, creating a clear time series based on the trigger sequence. This time series information helps understand the collision propagation path and analyze whether the external force is acting from the front, side, or specific angle on the robot. The system combines the location of the collision force center with this trigger sequence to pinpoint the exact point of collision using a spatial positioning algorithm. The direction of the collision, or the angle of incidence, is analyzed based on the relative position of the collision point target and the robot's center of mass. The robot's obstacle avoidance strategies vary depending on whether the collision originates from the front, side, or rear. For example, a head-on collision requires backing off or maneuvering, while a side collision requires a lateral course adjustment. Furthermore, all displacement signals from all four side wheels are screened, the maximum displacement value extracted, and the force sensitivity dynamically adjusted for each scenario, taking into account the different physical characteristics of the collision type, to estimate the actual collision force. The collision type, target location of the collision point, collision angle, and collision force are integrated into collision status information. Based on this collision status information, the robot automatically selects the most appropriate obstacle avoidance strategy for the current scenario. For example, if a window frame collision is identified, the robot prioritizes following the edge of the window frame to maintain a safe distance and prevent repeated collisions. If the collision type is an obstacle, the robot evaluates the obstacle's volume, material, and force conditions, selecting a reasonable detour radius and optimal detour direction to avoid the obstacle as much as possible and ensure a smooth path. When generating an obstacle avoidance path, the robot's current position is used as the starting point, and the expected return point in the cleaning schedule is selected as the end point. Intermediate control points are dynamically set based on parameters such as the collision force center and obstacle avoidance radius. Path generation utilizes smooth curve planning to avoid abrupt turns in the robot's trajectory and enhance its stability and safety. The path undergoes a series of optimization processes, taking into account path length, minimum safe distance from obstacles, and path smoothness. An optimization algorithm continuously fine-tunes the positions of control points along the path to achieve the optimal obstacle-avoiding path. The optimized smooth path is discretized into a continuous sequence of path points, each containing spatial coordinates and the desired direction of motion. These path point sequences are then transmitted to the robot's motion control unit in real time.

[0026] In one example, based on the collision state information, obstacle avoidance path parameters suitable for the current collision situation are calculated to generate a path point sequence, including: According to the collision type and collision point target position in the collision status information, the obstacle avoidance strategy is determined. When the collision type is a window frame collision, the boundary following strategy is set; when the collision type is an obstacle collision, the obstacle avoidance strategy is set. The obstacle avoidance strategy type for the current collision situation is obtained; Parameters are calculated based on the obstacle avoidance strategy type. For boundary following strategies, the obstacle avoidance distance and parallel window frame movement direction are calculated. For obstacle avoidance strategies, the obstacle avoidance radius and direction are calculated based on the collision force value to obtain the obstacle avoidance control parameters. Construct a quadratic Bezier curve control point based on the obstacle avoidance control parameters, set the current position as the starting point, and the expected return path point as the end point. Calculate the position of the intermediate control point based on the obstacle avoidance strategy type to obtain a control point set. Apply the Bezier curve parametric equation to the set of control points to generate an initial obstacle avoidance path; Based on the initial obstacle avoidance path, the path length, obstacle safety distance and path smoothness are used as evaluation indicators. The positions of the intermediate control points are optimized and adjusted through the gradient descent algorithm to obtain the optimal obstacle avoidance path. Discrete sampling is performed on the optimal obstacle avoidance path to generate a path point sequence.

[0027] In this example, the obstacle avoidance strategy is determined based on the collision type and target location of the collision point in the collision status information. For example, if the system determines a collision with the window frame, the robot focuses on following the boundary area and maintaining a safe buffer. In this case, the obstacle avoidance strategy type is set to Boundary Following, meaning the robot should maintain stable movement along the edge of the window frame as much as possible, flexibly adjusting its direction of travel to prevent secondary collisions or leaving the work surface. If an obstacle collision is determined, a more aggressive obstacle avoidance strategy is adopted. In this scenario, the obstacle avoidance path must not only stay away from the collision point but also consider the obstacle's size, material, and the robot's dynamic safety margins to ensure a smooth and efficient obstacle avoidance process. After determining the obstacle avoidance strategy type, the parameterization phase begins. For the Boundary Following strategy, the key parameters are determining the obstacle avoidance distance and setting the direction of motion parallel to the window frame. Setting the obstacle avoidance distance requires careful consideration of factors such as the clearance between the window frame and the robot, the safe buffer zone, and the window surface adhesion, ensuring the robot always stays within a suitable area for efficient cleaning without frequent collisions. The direction of movement parallel to the window frame is dynamically adjusted based on the angle between the robot's current orientation and the edge of the window frame. By correcting the speed of the robot's left and right motors in real time, the robot can move smoothly along the edge. For the obstacle avoidance strategy, the obstacle avoidance radius and obstacle avoidance direction become the core control parameters. The size of the obstacle avoidance radius is directly related to the estimated collision force and the physical size of the obstacle. When the system detects a large collision force, it means that the obstacle is harder or larger in size. The system adaptively increases the detour radius to avoid further contact with the obstacle. The judgment of the obstacle avoidance direction depends on the collision incident angle, the robot's current trajectory, and the spatial distribution of obstacles. The direction with more open space and easier turning is selected to ensure the continuity and safety of the path. The current position of the robot is used as the starting point of the obstacle avoidance path, and the expected return path point in the cleaning operation plan is used as the end point. Based on the determined obstacle avoidance strategy type, the position of the intermediate control point is dynamically calculated. For the boundary following strategy, the intermediate control point is located at a spatial point offset a set distance outward from the window frame's normal direction from the robot's current position, resulting in smooth, edge-to-edge motion. For the obstacle avoidance strategy, the intermediate control point is set at a tangent line centered on the obstacle and with a radius equal to the obstacle avoidance radius, minimizing collision risk while ensuring controllable robot motion. After determining these three points, the system constructs a set of control points for a quadratic Bezier curve to generate an initial obstacle avoidance path. Based on this initial obstacle avoidance path, a comprehensive evaluation metric is introduced. The total path length, the minimum distance from any point to the obstacle, and the overall curvature or smoothness of the path are simultaneously calculated. These multiple objective weights are then combined to form a unified optimization objective function. Using optimization algorithms such as gradient descent, the positions of the intermediate control points of the Bezier curve are automatically adjusted, and iterations are performed to gradually shorten the path, increase the safety distance, and improve motion consistency. Each adjustment improves the overall performance of the path until all evaluation metrics meet the set thresholds or reach a global optimum, at which point the optimal obstacle avoidance path is output.The optimal obstacle avoidance path is sampled discretically, converting the continuous curve into a sequence of pathpoints for real-time robot navigation. This discretization process segments the Bezier curve into a number of equally spaced pathpoints, each with spatial coordinates and a desired direction of travel, based on the sampling interval and travel accuracy of the robot's motion control system. This pathpoint sequence is transmitted to the robot's underlying motion control system in real time. Combined with actual environmental perception and dynamic feedback, this enables seamless switching between intelligent obstacle avoidance and cleaning operations.

[0028] In one example, a pathpoint sequence is transmitted to the motion control system of a window cleaning robot in real time. The robot's speed and direction are adaptively adjusted based on the path tracking error and the window surface adsorption status, enabling intelligent obstacle avoidance and continuous cleaning operations. This includes: The path deviation, deviation change rate, and window adhesion force are set as input variables of the fuzzy logic controller, and the left and right motor speed difference and forward speed are set as output variables of the fuzzy logic controller to obtain the basic controller structure of the window cleaning robot. Based on the basic controller structure, the fuzzy membership function and fuzzy rule base are set, and the center of gravity method is used to perform defuzzification operation on the fuzzy membership function and fuzzy rule base to obtain the motor control command; Determine the target path point based on the path point sequence and the current position of the robot, and calculate the position error and direction error based on the target path point to obtain the path tracking error vector; Based on the path tracking error vector and motor control instructions, the left and right motor speeds are differentially adjusted to obtain the robot's steering control value and forward speed control value; The side wheel displacement sensor data and negative pressure system data are monitored in real time. When an abnormal state is detected, the safety protection mechanism is triggered to limit the robot's steering control amount and forward speed control amount, and obtain motion control instructions under safety constraints to achieve intelligent obstacle avoidance and continuous execution of cleaning operations.

[0029] In this example, a fuzzy logic-based adaptive controller architecture is designed. It uses three key factors—path deviation, deviation change rate, and window adhesion—as input variables, while the left and right motor speed difference and overall forward speed serve as output variables. This basic controller architecture flexibly reflects the robot's current trajectory tracking status and adhesion security. By fuzzy modeling the relationship between input and output variables, it improves the system's adaptability and control accuracy. Appropriate fuzzy membership functions are established for the input and output variables. Path deviation is categorized as "negatively large," "negatively small," "zero," "positively small," and "positively large," while the deviation change rate and window adhesion are categorized as "small," "medium," and "large," respectively. The fuzzy set for each input variable is parameterized based on the actual sensor range and the system's physical response range, ensuring that the fuzzified result covers common operating conditions while remaining sensitive to extreme or unexpected conditions. After the input signal is fuzzified, the system develops a fuzzy rule base based on the actual physical scenario and operational experience. Real-time fuzzy inference and defuzzification operations are performed based on the currently collected sensor signals. After each set of input variables is transformed by the membership function, the system retrieves all matching fuzzy rules and performs a weighted fusion of the output variables based on their weights and membership values. To convert the fuzzy output variables into usable drive commands for the robot, a center of gravity method is used for defuzzification. The outputs of multiple fuzzy inference results are continuously quantized to generate the target speed difference between the left and right motors and the overall forward speed. During the path tracking phase, the robot continuously compares its current actual spatial position with the next closest target point in the pathpoint sequence. Using spatial geometry calculations, it calculates the linear distance and the angle between the current position and the target pathpoint. It then outputs the lateral and longitudinal path tracking errors and derives the required steering correction and the deviation from the current direction of travel. All error information is integrated to form a path tracking error vector. Based on this new path tracking error vector, combined with the left and right motor speed difference and forward speed command already calculated by the fuzzy controller, the motor speeds are adjusted in real time, enabling the robot to continuously approach the next pathpoint while maintaining a reasonable operating speed and motion posture. Through the differential control strategy, the robot moves forward at high speed when facing a straight path, and can adaptively reduce speed and flexibly turn when turning, circumventing obstacles, or when fine-tuning the direction, thereby effectively improving the accuracy of path tracking and operating efficiency. During actual operation, the displacement sensor data and negative pressure system data of the four side wheels are monitored in real time. The side wheel displacement sensor can promptly reflect the abnormal force of the robot under conditions such as external force, collision, slippage or offset, while the data of the negative pressure system can directly reflect the safety status of the robot adsorbed on the window surface. If the system detects abnormal fluctuations in the displacement of the side wheels, or the negative pressure value is lower than the safety threshold, it indicates that the robot is facing the risk of falling off, unstable adsorption, or increased risk due to excessive path tracking errors.The system immediately triggers a safety protection mechanism, dynamically limiting the control output to the motor. For example, if the suction force is insufficient, the system reduces forward speed and reduces or suspends left and right steering to prevent the robot from sliding out of the window due to inertia. If the side wheels experience serious abnormalities, the system directly issues a stop command, requests manual intervention, or enters safety self-check mode to ensure the safety of equipment and personnel during operation.

[0030] Reference Figure 2 This embodiment provides a collision recognition device for a window cleaning robot, comprising: Acquisition module 1 is used to acquire the displacement data of the radial displacement sensor installed on the side wheel of the window cleaning robot and the status data of the photoelectric drop detector installed at the bottom of the robot to obtain the raw data of the dual-modal sensor; Similarity matching module 2 is used to perform similarity matching between the raw data of the dual-modal sensor and the preset window frame collision feature template library and the obstacle collision feature template library to obtain a collision judgment result; Calculation module 3 is used to analyze the force distribution and trigger timing differences of multiple side wheels according to the collision judgment results, calculate the obstacle avoidance path parameters suitable for the current collision situation, and generate a path point sequence; The adaptive adjustment module 4 is used to transmit the path point sequence to the motion control system of the window cleaning robot in real time, and adaptively adjust the robot's running speed and direction according to the path tracking error and the window surface adsorption state, so as to realize intelligent obstacle avoidance and continuous execution of the cleaning operation.

[0031] In this embodiment, for the specific implementation of each unit in the above device embodiment, please refer to the above method embodiment, which will not be repeated here.

[0032] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, apparatus, article, or method comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, apparatus, article, or method. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, apparatus, article, or method comprising the element.

[0033] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A collision recognition method for a window cleaning robot, characterized in that: include: Obtain the displacement data of the radial displacement sensor installed on the side wheel of the window cleaning robot and the status data of the photoelectric drop detector installed on the bottom of the robot to obtain the raw data of the dual-modal sensor; Performing similarity matching on the raw data of the dual-modal sensor with a preset window frame collision feature template library and an obstacle collision feature template library to obtain a collision determination result; Analyzing the force distribution and trigger timing differences of multiple side wheels according to the collision determination results and calculating obstacle avoidance path parameters suitable for the current collision situation to generate a path point sequence; The path point sequence is transmitted to the motion control system of the window cleaning robot in real time, and the robot's running speed and direction are adaptively adjusted according to the path tracking error and the window surface adsorption state, so as to realize intelligent obstacle avoidance and continuous execution of the cleaning operation.

2. The collision recognition method for a window cleaning robot according to claim 1, characterized in that: The method of obtaining the displacement data of the radial displacement sensor installed at the side wheel of the window cleaning robot and the status data of the photoelectric drop detector installed at the bottom of the robot to obtain the dual-modal sensor raw data includes: Install radial displacement sensors on the inner wall of the four side wheel sockets of the window cleaning robot, and keep the sensing rod in contact with the inner side of the side wheel to obtain the radial displacement electrical signal generated when the side wheel is subjected to external force; Photoelectric drop detectors that emit infrared light of a preset wavelength are installed at the four corners of the bottom of the window cleaning robot. The changes in the intensity of the reflected light are detected to obtain status signals when the robot is suspended in the air or partially leaves the window surface. performing digital processing on the radial displacement electrical signal and the status signal respectively to obtain displacement data of the radial displacement sensor and status data of the photoelectric drop detector, and performing time sequence alignment processing on the displacement data and the status data to obtain a sensor data stream with a continuous time sequence; The time-series continuous sensor data stream is filtered to obtain target filtered data, and dynamic baseline correction is performed on the target filtered data to obtain dual-modal sensor raw data.

3. The collision recognition method for a window cleaning robot according to claim 2, characterized in that: The raw data of the dual-modal sensor is matched with a preset window frame collision feature template library and an obstacle collision feature template library for similarity to obtain a collision determination result, including: According to the raw data of the dual-modal sensor, the displacement data is constructed into a displacement vector, and the state data of the photoelectric drop detector is constructed into a state vector; Performing segmentation and fusion alignment of the displacement vector and the state vector into continuous sampling periods according to a preset time window length to obtain an initial spatiotemporal feature matrix; Performing differential calculation on adjacent sampling point data in the initial spatiotemporal feature matrix to obtain a time-domain differential displacement vector representing a displacement change rate and a time-domain differential state vector representing a state change; The initial spatiotemporal feature matrix is ​​concatenated with the time-domain differential displacement vector and the time-domain differential state vector to obtain an enhanced spatiotemporal feature matrix; Performing principal component analysis on the enhanced spatiotemporal feature matrix to obtain a feature subspace after dimensionality reduction, retaining principal components whose cumulative explained variance reaches a preset threshold in the feature subspace after dimensionality reduction, and performing projection transformation on the enhanced spatiotemporal feature matrix to obtain a feature fusion matrix; The feature fusion matrix is ​​matched with a preset window frame collision feature template library and an obstacle collision feature template library for similarity to obtain a collision determination result.

4. The collision recognition method for a window cleaning robot according to claim 3, characterized in that: The feature fusion matrix is ​​matched with a preset window frame collision feature template library and an obstacle collision feature template library for similarity to obtain a collision determination result, including: Resampling the feature fusion matrix at different time scales to obtain a multi-scale feature representation; Preliminarily matching the multi-scale feature representation with a preset window frame collision feature template library, calculating the similarity of horizontal window frame collision templates, vertical window frame collision templates, and corner window frame collision templates, and obtaining a set of window frame collision candidate templates; Preliminarily matching the multi-scale feature representation with a preset obstacle collision feature template library, calculating the similarity between the hard obstacle collision template and the soft obstacle collision template, and obtaining a set of obstacle collision candidate templates; Performing a dynamic time warping algorithm on the feature fusion matrix and the window frame collision candidate template set and the obstacle collision candidate template set, respectively, to obtain a window frame collision similarity measurement value and an obstacle collision similarity measurement value; Extracting a time series characteristic index of the collision signal from the feature fusion matrix, and performing weighted correction on the window frame collision similarity measurement value and the obstacle collision similarity measurement value based on the time series characteristic index to obtain a window frame collision correction matching probability and an obstacle collision correction matching probability; A collision determination result is generated according to the window frame collision correction matching probability and the obstacle collision correction matching probability.

5. The collision recognition method for a window cleaning robot according to claim 1, characterized in that: The method of analyzing the force distribution and trigger timing differences of multiple side wheels according to the collision determination result and calculating obstacle avoidance path parameters suitable for the current collision situation to generate a path point sequence includes: Determine the collision type according to the collision determination result, perform weighted calculation on the displacement data of the four side wheels to obtain the coordinates of the collision force center, and record and sort the triggering times of the radial displacement sensors of the four side wheels to obtain a time sequence; Based on the collision force center coordinates and the time series, a triangulation calculation is performed on the collision position to obtain a target position of the collision point; According to the relative relationship between the target position of the collision point and the position of the robot's center of mass, the collision incident angle is calculated to obtain the collision angle value, the maximum displacement value of all side wheels is extracted and the stiffness coefficient is adjusted in combination with the collision type characteristics to obtain the collision force value; Integrating the collision type, the collision point target position, the collision angle value, and the collision force value into collision state information; Obstacle avoidance path parameters suitable for the current collision situation are calculated based on the collision state information, and a path point sequence is generated.

6. The collision recognition method for a window cleaning robot according to claim 5, characterized in that: The calculating of obstacle avoidance path parameters suitable for the current collision situation based on the collision state information and generating a path point sequence includes: Determine the obstacle avoidance strategy based on the collision type and collision point target position in the collision state information. When the collision type is a window frame collision, set it to a boundary following strategy. When the collision type is an obstacle collision, set it to an obstacle avoidance strategy. Obtain the obstacle avoidance strategy type for the current collision situation. Parameters are calculated based on the obstacle avoidance strategy type. When the boundary following strategy is used, the obstacle avoidance distance and the direction of movement of the parallel window frame are calculated. When the obstacle avoidance strategy is used, the obstacle avoidance radius and obstacle avoidance direction are calculated according to the collision force value to obtain the obstacle avoidance control parameters. Constructing quadratic Bezier curve control points according to the obstacle avoidance control parameters, setting the current position as the starting point, setting the expected return path point as the end point, calculating the position of the intermediate control point based on the obstacle avoidance strategy type, and obtaining a control point set; Applying a Bezier curve parametric equation to the set of control points to generate an initial obstacle avoidance path; Based on the initial obstacle avoidance path, the positions of the intermediate control points are optimized and adjusted using a gradient descent algorithm with path length, obstacle safety distance, and path smoothness as evaluation indicators to obtain the optimal obstacle avoidance path; Discrete sampling is performed on the optimal obstacle avoidance path to generate a path point sequence.

7. The collision recognition method for a window cleaning robot according to claim 1, characterized in that: The method of transmitting the path point sequence to the motion control system of the window cleaning robot in real time and adaptively adjusting the robot's running speed and direction according to the path tracking error and the window surface adsorption state to achieve intelligent obstacle avoidance and continuous execution of the cleaning operation includes: The path deviation, deviation change rate, and window adhesion force are set as input variables of the fuzzy logic controller, and the left and right motor speed difference and forward speed are set as output variables of the fuzzy logic controller to obtain the basic controller structure of the window cleaning robot. A fuzzy membership function and a fuzzy rule base are set based on the basic controller structure, and a center of gravity method is used to perform a defuzzification operation on the fuzzy membership function and the fuzzy rule base to obtain a motor control instruction; Determine a target path point according to the path point sequence and the current position of the robot, and calculate a position error and a direction error based on the target path point to obtain a path tracking error vector; Based on the path tracking error vector and the motor control instruction, differentially adjusting the speeds of the left and right motors to obtain a steering control amount and a forward speed control amount of the robot; The side wheel displacement sensor data and the negative pressure system data are monitored in real time. When an abnormal state is detected, the safety protection mechanism is triggered to limit the steering control amount and the forward speed control amount of the robot, and obtain motion control instructions under safety constraints to achieve intelligent obstacle avoidance and continuous execution of cleaning operations.

8. A collision recognition device for a window cleaning robot, characterized in that: The steps for implementing the collision recognition method of the window cleaning robot according to any one of claims 1 to 7, wherein the collision recognition device of the window cleaning robot comprises: An acquisition module is used to acquire the displacement data of the radial displacement sensor installed on the side wheel of the window cleaning robot and the status data of the photoelectric drop detector installed at the bottom of the robot to obtain the raw data of the dual-modal sensor; A similarity matching module is used to perform similarity matching on the raw data of the dual-modal sensor with a preset window frame collision feature template library and an obstacle collision feature template library to obtain a collision determination result; A calculation module, configured to analyze the force distribution and trigger timing differences of the multiple side wheels according to the collision determination results, calculate obstacle avoidance path parameters suitable for the current collision situation, and generate a path point sequence; The adaptive adjustment module is used to transmit the path point sequence to the motion control system of the window cleaning robot in real time, and adaptively adjust the robot's running speed and direction according to the path tracking error and the window surface adsorption state, so as to realize intelligent obstacle avoidance and continuous execution of cleaning operations.

Images