An intelligent cleaning device based on dirt shape perception and multi-dimensional feature fusion

Abstract

This invention discloses an intelligent cleaning device based on dirt morphology perception and multi-dimensional feature fusion, comprising: a multi-dimensional perception module for collecting multi-source perception data of the cleaning area; an AI decision processing module with a preset dirt morphology detection model for joint feature extraction of the multi-source perception data, outputting dirt morphology assessment results and cleaning intention labels through the dirt morphology detection model; an adaptive execution module for dynamically adjusting suction parameters, brush rotation speed parameters, and water output and shut-off parameters based on the dirt morphology assessment results and cleaning intention labels; a negative pressure feedback module for collecting the actual negative pressure value of the high-speed fluid power unit and forming comprehensive negative pressure feedback data; and a main control module for closed-loop control of the high-speed fluid power unit based on the deviation between the actual negative pressure value and the target suction parameters. This invention achieves high-precision dirt perception and adaptive suction adjustment, solving the technical problems of single-dimensional recognition and high noise of high-speed motors in traditional cleaning devices.

Description

Technical Field

[0001] This invention relates to the technical field of intelligent devices, and in particular to an intelligent cleaning device based on the perception of dirt morphology and the fusion of multi-dimensional features. Background Technology

[0002] The development of intelligent cleaning equipment with an electric motor as its core suction mechanism stems from the synergistic effect of vacuum suction principles, motor technology iterations, the popularization of smart homes, and intelligent control technologies. As the heart of the device, the motor generates negative pressure through high-speed rotation, creating suction to separate dust and air, and its performance directly determines the cleaning effect and user experience.

[0003] In terms of technological evolution, the suction mechanism motor of the device has undergone three generations of iteration: from the early brushed DC motor, characterized by low speed, short life and high noise, to the current mainstream brushless DC motor, and then to the flagship high-speed digital motor, gradually solving the pain points such as weak suction, short battery life and high noise.

[0004] However, existing smart cleaning equipment still has certain defects: (1) Single recognition and easily interfered with: It relies on a single vision camera. When encountering reflective floors and countertops, the vision algorithm fails and cannot identify dirt; (2) Lack of in-depth understanding of the physical form of dirt: It cannot distinguish between dry ash, water stains and solidified oil stains, resulting in "running around blindly" when facing stubborn stains and failing to clean them. When facing dust, it will over-wet the carpet; (3) Lack of human-like interactive behavior logic: When cleaning stubborn stains, users are used to stopping the equipment to let it "soak" for a while. Traditional equipment will automatically shut down or cut off the water when stopped, which interrupts the cleaning process. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and propose an intelligent cleaning device based on dirt morphology perception and multi-dimensional feature fusion. This device simultaneously collects multi-source sensory data, including visual, acoustic, and pressure data, and uses an attention-weighted fusion method based on a dirt morphology detection model to accurately output dirt morphology assessment results, such as dry particles, wet adhesion, or complex stubborn stains, as well as cleaning intention labels such as dynamic pushing / pulling or static floor contact. The system adopts dynamic temporal tracking logic, automatically executing a water-stopping and slow-down soaking strategy during the static period of stubborn stain cleaning, and triggering a completion interruption when the dirt characteristics completely disappear and the negative pressure stabilizes. Simultaneously, it innovatively introduces a multi-source cross-fault-tolerant mechanism to overcome the blind spots of single-vision perception on reflective surfaces.

[0006] The objective of this invention is achieved through the following technical solution: an intelligent cleaning device based on dirt morphology perception and multi-dimensional feature fusion, comprising:

[0007] The multi-dimensional sensing module is used to collect multi-source sensing data of the clean area, including visual image data, acoustic data, pressure data, and laser ranging data.

[0008] The AI ​​decision processing module has a pre-set dirt morphology detection model. It performs joint feature extraction on multi-source sensing data, generates a multi-dimensional fusion feature vector including dirt morphology features and dynamic inhalation features, and outputs dirt morphology assessment results and cleaning intention labels through the dirt morphology detection model.

[0009] The adaptive execution module, including a high-speed fluid power unit and a function adjustment mechanism, dynamically adjusts the suction parameters, brush speed parameters, and water output and shut-off parameters based on the dirt morphology assessment results and cleaning intention labels output by the decision processing module.

[0010] The negative pressure feedback module communicates with the AI ​​decision processing module to collect the actual negative pressure value of the high-speed fluid power unit, form comprehensive negative pressure feedback data, and feed it back to the AI ​​decision processing module to form a suction closed-loop control.

[0011] The main control module performs closed-loop control of the high-speed fluid power unit based on the deviation between the actual negative pressure value and the target suction parameter.

[0012] Furthermore, the AI ​​decision processing module performs the following operations:

[0013] Based on a preset time sliding window, dirt morphology features are extracted from visual image data, and dynamic inhalation features are extracted from acoustic or pressure data.

[0014] Attention weights are assigned based on the real-time confidence of each multi-source sensing data, and the morphological features of dirt and dynamic inhalation features are fused into a multi-dimensional fusion feature vector.

[0015] The fused feature vector is input into the dirt morphology detection model, and the model outputs simultaneously through a multi-task learning architecture: dirt morphology assessment results that characterize the physical properties of dirt and cleaning intention labels that characterize the current state of the equipment. The dirt morphology assessment results include dry particulates, wet adhesives, and complex stubborn dirt. The cleaning intention labels include dynamic push-pull interaction, static ground-adhesive, and cleaning completed blocking status.

[0016] Furthermore, the determination of the cleaning completion blocking state includes:

[0017] During the cleaning process, continuously track the percentage of dirt area in the visual characteristics of dirt or the rate of change of particle concentration over time in the dynamic inhalation characteristics.

[0018] When the percentage of dirty area shows a monotonically decreasing trend and the rate of change is within the preset effective cleaning range, it is determined to be in the process of cleaning.

[0019] When the rate of change is lower than the stagnation threshold and continues for a preset duration, and the comprehensive negative pressure feedback data indicates that the current actual negative pressure has stabilized within the limit safety threshold representing no-load, an intention label indicating that cleaning is complete is generated and output to the main control module to control the device to exit the high-power cleaning mode or trigger a shutdown prompt.

[0020] Furthermore, the determination of the composite stubborn class includes:

[0021] When the dirt morphology assessment result is complex and stubborn, and the cleaning intention label is static ground-adsorption type, the displacement variance in the dynamic suction characteristics is continuously monitored.

[0022] If the displacement variance drops below the first preset threshold and continues to exceed the first soaking time, it is determined to be a stubborn stain softening interaction pause state.

[0023] The control function adjustment mechanism stops water output or reduces the roller brush speed to the preset maintenance speed to ensure that the cleaning liquid fully acts on the stains; when the displacement variance recovers, it determines that the softening is over and instantly restores the high suction power and high speed output.

[0024] Furthermore, the AI ​​decision processing module performs the following operations:

[0025] When visual image data shows that the morphological characteristics of dirt are lower than the physical judgment threshold due to the reflectivity or uniformity of the ground material, if acoustic data or pressure data confirm the physical fluctuations caused by the inhalation of foreign objects, and the negative integrated negative pressure feedback data indicates that the actual negative pressure deviates from the no-load reference value, then the visual judgment limit is crossed, and the current cleaning intention label for dynamic push-pull interactive cleaning is maintained.

[0026] Furthermore, the high-speed hydrodynamic unit includes:

[0027] The stator assembly includes multiple sets of stator core units, which are wound in parallel and assembled to form a closed structure that forms a continuous and uninterrupted magnetic circuit loop.

[0028] The rotor assembly is located inside the stator assembly;

[0029] The dynamic balancing mechanism is located at both ends of the rotor assembly and is used to automatically generate a restoring torque that resists the eccentric displacement when the rotor assembly undergoes eccentric displacement due to high-speed rotation.

[0030] Furthermore, the dynamic balancing mechanism includes elastic support members disposed at both ends of the rotor core along the axial direction; when the rotor core tilts, the elastic support member on the side closer to the rotation center is compressed to generate a first reverse elastic force, and the elastic support member on the side farther from the rotation center is stretched to generate a second reverse tensile force. The first reverse elastic force and the second reverse tensile force together constitute the restoring torque, which causes the rotor vibration to decay and return to the equilibrium position.

[0031] A control method for an intelligent cleaning device based on the above-mentioned dirt morphology perception and multi-dimensional feature fusion includes the following steps:

[0032] S1. Acquire multi-source sensing data of the cleaning area in real time;

[0033] S2. Multi-dimensional feature fusion is performed on multi-source sensing data through a dirt morphology detection model to output dirt morphology assessment results and cleaning intention labels.

[0034] S3. Based on the assessment results of dirt morphology, match the target suction power, target speed and target water output strategy, and execute intermittent start / stop or complete blocking logic based on the cleaning intention label;

[0035] S4. Perform closed-loop correction control of the fluid power unit based on the deviation between the target suction force and the actual negative pressure.

[0036] Furthermore, step S3 includes:

[0037] When the dirt morphology detection model outputs a dirt morphology assessment result of composite stubborn type and the cleaning intention label is static ground adsorption type, if the AI ​​decision processing module detects that the variance value representing the device displacement is lower than the preset displacement threshold and continues to exceed the set softening time threshold, it is determined to be a dwelling softening state.

[0038] In the softening state, the main control module controls the function adjustment mechanism to stop the output of cleaning fluid and controls the speed of the function adjustment mechanism to reduce to the preset maintenance speed so that the output cleaning fluid can fully adhere to the stubborn stains.

[0039] When the variance of the device displacement is detected to recover and the proportion of dirt area in the dirt morphology assessment results shows a monotonically decreasing trend, the main control module releases the dwell softening state, controls the high-speed fluid power unit to restore to the maximum target suction, and increases the function adjustment mechanism to the stripping speed.

[0040] Furthermore, step S3 includes:

[0041] A weak feature ground fault-tolerant cruise working mode is preset. When the confidence of visual feature extraction is lower than the effective judgment threshold due to the reflectivity of ground material being higher than the preset reflectivity threshold, the weak feature ground fault-tolerant cruise working mode is triggered.

[0042] In the weak feature ground fault-tolerant cruise working mode, the AI ​​decision processing module forcibly increases the weight ratio of acoustic data and pressure data in multi-dimensional feature fusion;

[0043] If the acoustic data indicates frequency domain energy fluctuations that match the characteristics of particle impact, or if the pressure data indicates that the actual negative pressure value of the air duct deviates from the no-load reference value to the preset deviation threshold, the main control module will bypass the visual judgment limitation, lock the current cleaning area, and maintain the current suction parameter output of the high-speed fluid power unit.

[0044] At the same time, the main control module controls the function adjustment mechanism to perform timed pre-spraying actions until the rate of change of the percentage of dirty area output by the dirty morphology detection model is lower than the stagnation threshold, at which point the cleaning is determined to be complete and the lock is released.

[0045] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0046] 1. This invention introduces a multi-dimensional feature fusion and state machine tracking strategy. Through multi-modal data such as vision, acoustics, and air pressure, it accurately identifies the dirt morphology assessment results and accurately judges the operation of the equipment based on the dirt morphology assessment results, thereby achieving biomimetic intelligent cleaning.

[0047] 2. This invention can achieve high-precision dirt detection and adaptive suction adjustment without network, taking into account cleaning efficiency, energy consumption and operational stability. It solves the technical problems of traditional cleaning equipment, such as cumbersome manual adjustment, single recognition dimension and large vibration and noise of high-speed motor. It is suitable for intelligent cleaning in multiple scenarios such as home, car, and office. Attached Figure Description

[0048] Figure 1 This is an architecture diagram of a smart cleaning device.

[0049] Figure 2 This is a control flowchart for intelligent cleaning equipment.

[0050] Figure 3 This is one of the structural schematic diagrams of a high-speed hydrodynamic unit.

[0051] Figure 4 This is the second schematic diagram of the high-speed hydrodynamic unit. Detailed Implementation

[0052] The present invention will be further described below with reference to specific embodiments.

[0053] Example 1

[0054] The intelligent cleaning device based on dirt morphology perception and multi-dimensional feature fusion provided in this embodiment includes:

[0055] 1) Multi-dimensional sensing module, used to collect multi-source sensing data of the clean area. The multi-source sensing data includes visual image data, acoustic data, pressure data and laser ranging data. At the same time, the multi-modal sensing module integrates visual sensors, acoustic sensors, pressure sensors and laser rangefinders.

[0056] 2) The AI ​​decision processing module has a pre-set dirt morphology detection model. It performs joint feature extraction on multi-source sensing data to generate a multi-dimensional fused feature vector that includes dirt morphology features and dynamic inhalation features. The module then outputs dirt morphology assessment results and cleaning intention labels through the dirt morphology detection model, including the following steps:

[0057] S2.1. Perform preprocessing steps including filtering, noise reduction, and normalization on the collected multi-source sensing data. Extract joint multi-dimensional fusion feature vectors from the multi-source sensing data. The multi-dimensional fusion feature vectors include dirt morphology features and dynamic inhalation features. Specifically, based on a preset time sliding window, extract dirt morphology features from visual image data and extract dynamic inhalation features from acoustic data or pressure data. Assign attention weights according to the real-time confidence of each multi-source sensing data, and fuse dirt morphology features and dynamic inhalation features into a multi-dimensional fusion feature vector.

[0058] S2.2 Input the multi-dimensional fusion features into the dirt morphology detection model and output the dirt morphology assessment results and cleaning intention labels. Specifically, through a multi-task learning architecture, the dirt morphology assessment results representing the physical properties of dirt and the cleaning intention labels representing the current state of the equipment are output simultaneously. The dirt morphology assessment results include dry particulates, wet adhesives, and composite stubborn dirt. The cleaning intention labels include dynamic push-pull interaction, static ground adhesion, and cleaning completed blocking status.

[0059] The determination of the cleaning completion interruption status includes: during the cleaning process, continuously tracking the percentage of dirt area in the visual characteristics of dirt or the rate of change of particle concentration in the dynamic inhalation characteristics over time; when the percentage of dirt area shows a monotonically decreasing trend and the rate of change is within the preset effective cleaning range, it is determined to be cleaning in progress; when the rate of change is lower than the stagnation threshold and continues for a preset duration, and the comprehensive negative pressure feedback data indicates that the current actual negative pressure has stabilized within the limit safety threshold representing no load, an intention label for cleaning completion interruption is generated and output to the main control module to control the device to exit the high-power cleaning mode or trigger a shutdown prompt.

[0060] The determination of complex and stubborn stains includes: when the stain morphology assessment result is complex and stubborn, and the cleaning intention label is static ground adsorption type, the displacement variance in the dynamic suction characteristics is continuously monitored; if the displacement variance drops below the first preset threshold and continues to exceed the first soaking time, it is determined to be a stubborn stain softening interactive pause state; the control function adjustment mechanism stops water output or reduces the roller brush speed to the preset maintenance speed so that the cleaning liquid can fully act on the stain; when the displacement variance recovers, it is determined that the softening ends, and the high suction power and high speed output are instantly restored.

[0061] S2.3 When the visual image data shows that the dirt morphology is below the physical judgment threshold due to the reflection of the ground material or the uniformity of the color, if the acoustic data or pressure data confirms the physical fluctuation caused by the inhalation of foreign objects, and the negative integrated negative pressure feedback data indicates that the actual negative pressure deviates from the no-load reference value, then the visual judgment limit is crossed, and the current cleaning intention label of dynamic push-pull interactive cleaning is maintained.

[0062] 3) The adaptive execution module, including a high-speed fluid power unit and a function adjustment mechanism, dynamically adjusts suction parameters, brush speed parameters, and water output and shut-off parameters based on the dirt morphology assessment results and cleaning intention tags output by the decision processing module; see [link to relevant documentation]. Figures 3 to 4 As shown, the high-speed hydrodynamic unit includes a stator assembly 2, a rotor assembly 3, a housing 1, a dynamic balancing mechanism 4, turbine blades 5, and a turbine blade shroud 6. The stator assembly 2, rotor assembly 3, and dynamic balancing mechanism 4 are installed inside the housing 1 to form a high-speed motor a. The stator assembly 2 is a closed structure formed by the in-line winding and assembly of multiple sets of stator core units to form a continuous and uninterrupted magnetic circuit. The rotor assembly 3 is located inside the stator assembly 2. The dynamic balancing mechanism 4 is located at both ends of the rotor assembly 3 and is used to automatically generate a restoring torque that resists the eccentric displacement when the rotor assembly 3 undergoes eccentric displacement due to high-speed rotation. The dynamic balancing mechanism 4 includes elastic support members located at both ends of the magnetic core of the rotor assembly 3. When the rotor magnetic core tilts, the elastic support member closer to the rotation center is compressed to generate a first reverse elastic force, and the elastic support member farther from the rotation center is stretched to generate a second reverse tensile force. The first reverse elastic force and the second reverse tensile force together constitute the restoring torque, causing the rotor vibration to decay and return to the equilibrium position.

[0063] 4) The negative pressure feedback module is connected to the AI ​​decision processing module to collect the actual negative pressure value of the high-speed fluid power unit, form comprehensive negative pressure feedback data, and feed it back to the AI ​​decision processing module to form a suction closed-loop control.

[0064] 5) The main control module performs closed-loop control of the high-speed fluid power unit based on the deviation between the actual negative pressure value and the target suction parameters.

[0065] 6) Interaction module, used to receive user commands and report the device operating status to the main control module.

[0066] 7) Power management module, used to provide power management services for the device.

[0067] Example 2

[0068] This embodiment provides a control method for an intelligent cleaning device based on dirt morphology perception and multi-dimensional feature fusion as described in Embodiment 1, including the following steps:

[0069] S1. Acquire multi-source sensing data of the cleaning area in real time;

[0070] S2. Multi-dimensional feature fusion is performed on multi-source sensing data through a dirt morphology detection model to output dirt morphology assessment results and cleaning intention labels.

[0071] S3. Based on the assessment results of dirt morphology, match the target suction power, target speed and target water output strategy, and execute intermittent start / stop or complete blocking logic based on the cleaning intention label;

[0072] Specifically, when the dirt morphology assessment result output by the dirt morphology detection model is a complex stubborn type, and the cleaning intention label is a static ground-adhesive type, if the AI ​​decision processing module detects that the variance value representing the device displacement is lower than the preset displacement threshold and continues to exceed the set softening time threshold, it is determined to be a dwelling softening state. In the dwelling softening state, the main control module controls the function adjustment mechanism to stop the output of cleaning fluid and controls the speed of the function adjustment mechanism to reduce to the preset maintenance speed so that the output cleaning fluid can fully adhere to the stubborn stains. When the variance value of the device displacement is detected to recover and the proportion of dirt area in the dirt morphology assessment result shows a monotonically decreasing trend, the main control module releases the dwelling softening state, controls the high-speed fluid power unit to recover to the maximum target suction, and increases the function adjustment mechanism to the peeling speed.

[0073] A weak-feature ground fault-tolerant cruise working mode is preset. When the confidence level of visual feature extraction is lower than the effective judgment threshold due to the reflectivity of the ground material being higher than the preset reflectivity threshold, the weak-feature ground fault-tolerant cruise working mode is triggered. In the weak-feature ground fault-tolerant cruise working mode, the AI ​​decision processing module forcibly increases the weight ratio of acoustic data and pressure data in multi-dimensional feature fusion. If the acoustic data indicates the presence of frequency domain energy fluctuations that conform to the characteristics of particle impact, or if the pressure data indicates that the actual negative pressure value of the air duct deviates from the empty load reference value to the preset deviation threshold, the main control module will bypass the visual judgment limitation, lock the current cleaning area, and maintain the current suction parameter output of the high-speed fluid power unit. At the same time, the main control module controls the function adjustment mechanism to perform timed pre-spraying actions until the rate of change of the dirt area ratio output by the dirt morphology detection model is lower than the stagnation threshold, at which point the cleaning is determined to be completed and the lock is released.

[0074] S4. Perform closed-loop correction control of the fluid power unit based on the deviation between the target suction force and the actual negative pressure.

[0075] Example 3

[0076] See Figure 2As shown in this embodiment, the intelligent cleaning device has soaking and softening logic and dynamic tracking logic. The specific execution of the soaking and softening logic and dynamic tracking logic is as follows: When a large, stubborn oil stain is detected and the device is held still by the user on the stain, the system enters the "soaking and softening state." The AI ​​decision processing module tracks the visual characteristics of the stain area. At this time, the area has not decreased, i.e., the rate of change approaches 0. The device determines that the interaction is paused, automatically closes the water outlet solenoid valve, and reduces the roller brush speed to allow the sprayed cleaning liquid to fully dissolve the oil stain. When the user pushes the device again, i.e., the displacement variance suddenly increases, the device instantly restores maximum water output and maximum speed within 50ms. As cleaning progresses, the visual tracking shows a monotonically decreasing oil stain area. When the area change rate drops to 0 and the air duct negative pressure returns to the no-load baseline, a "cleaning complete interruption" label is generated, and a voice prompt tells the user that the area has been cleaned.

[0077] Example 4

[0078] See Figure 1 As shown, this embodiment discloses an intelligent floor scrubber. The multi-dimensional sensing module includes an RGB-D camera facing the ground to capture visual image data; a microphone array within the intake duct to collect acoustic data, used to hear the sound of dust hitting the roller brush; and an air pressure sensor within the duct to collect pressure data. The adaptive execution module includes a 140,000 rpm high-speed fan composed of a linear wound stator and an adaptive tension rotor, serving as a high-speed fluid power unit; and an independently controllable water outlet solenoid valve and roller brush drive motor, serving as a function adjustment mechanism.

[0079] When the dirt morphology detection model outputs a dirt morphology assessment result of "complex stubborn" and the cleaning intention label is "static ground-adhesive," if the AI ​​decision processing module detects that the variance value representing the device displacement is lower than the preset displacement threshold and continues to exceed the set softening time threshold, it is determined to be in a dwelling softening state. In the dwelling softening state, the main control module controls the function adjustment mechanism to stop the output of cleaning fluid and controls the speed of the roller brush drive motor to decrease to the preset maintenance speed so that the output cleaning fluid can fully adhere to the stubborn stains. When the device displacement variance value is detected to recover and the proportion of dirt area in the dirt morphology assessment result shows a monotonically decreasing trend, the main control module instantly releases the dwelling softening state, controls the high-speed fluid power unit to restore to the maximum target suction, and increases the roller brush drive motor to the peeling speed.

[0080] When the confidence level of visual feature extraction is lower than the effective judgment threshold due to the reflectivity of the ground material being higher than the preset reflectivity threshold, the weak feature ground fault-tolerant cruise working mode is triggered. In this mode, the AI ​​decision processing module forcibly increases the weight ratio of acoustic data and pressure data in multi-dimensional feature fusion. If the acoustic data indicates the presence of frequency domain energy fluctuations that conform to the characteristics of particle impact, or if the pressure data indicates that the actual negative pressure value of the air duct deviates from the empty load reference value to the preset deviation threshold, the main control module will bypass the visual judgment limitation, lock the current cleaning area, and maintain the current suction parameter output of the high-speed fluid power unit. At the same time, the main control module controls the function adjustment mechanism to perform timed pre-spraying actions until the rate of change of the dirt area ratio output by the dirt morphology detection model is lower than the stagnation threshold, at which point the cleaning is determined to be completed and the lock is released.

[0081] The above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Therefore, any changes made in accordance with the shape and principle of the present invention should be covered within the protection scope of the present invention.

Claims

1. An intelligent cleaning device based on dirt morphology perception and multi-dimensional feature fusion, characterized in that, include: The multi-dimensional sensing module is used to collect multi-source sensing data of the clean area, including visual image data, acoustic data, pressure data, and laser ranging data. The AI ​​decision processing module has a pre-set dirt morphology detection model. It performs joint feature extraction on multi-source sensing data, generates a multi-dimensional fusion feature vector including dirt morphology features and dynamic inhalation features, and outputs dirt morphology assessment results and cleaning intention labels through the dirt morphology detection model. The adaptive execution module, including a high-speed fluid power unit and a function adjustment mechanism, dynamically adjusts the suction parameters, brush speed parameters, and water output and shut-off parameters based on the dirt morphology assessment results and cleaning intention labels output by the decision processing module. The negative pressure feedback module is used to collect the actual negative pressure value of the high-speed fluid power unit and form comprehensive negative pressure feedback data; The main control module performs closed-loop control of the high-speed fluid power unit based on the deviation between the actual negative pressure value and the target suction force parameter.

2. The intelligent cleaning device based on dirt morphology perception and multi-dimensional feature fusion according to claim 1, characterized in that, The AI ​​decision processing module performs the following operations: Based on a preset time sliding window, dirt morphology features are extracted from visual image data, and dynamic inhalation features are extracted from acoustic or pressure data. Attention weights are assigned based on the real-time confidence of each multi-source sensing data, and the morphological features of dirt and dynamic inhalation features are fused into a multi-dimensional fusion feature vector. The fused feature vector is input into the dirt morphology detection model, and the model outputs simultaneously through a multi-task learning architecture: dirt morphology assessment results that characterize the physical properties of dirt and cleaning intention labels that characterize the current state of the equipment. The dirt morphology assessment results include dry particulates, wet adhesives, and complex stubborn dirt. The cleaning intention labels include dynamic push-pull interaction, static ground-adhesive, and cleaning completed blocking status.

3. The intelligent cleaning device based on dirt morphology perception and multi-dimensional feature fusion according to claim 2, characterized in that, The determination of the cleaning completion and blocking status includes: During the cleaning process, continuously track the percentage of dirt area in the visual characteristics of dirt or the rate of change of particle concentration over time in the dynamic inhalation characteristics. When the percentage of dirty area shows a monotonically decreasing trend and the rate of change is within the preset effective cleaning range, it is determined to be in the process of cleaning. When the rate of change is lower than the stagnation threshold and continues for a preset duration, and the comprehensive negative pressure feedback data indicates that the current actual negative pressure has stabilized within the limit safety threshold representing no-load, an intention label indicating that cleaning is complete is generated and output to the main control module to control the device to exit the high-power cleaning mode or trigger a shutdown prompt.

4. The intelligent cleaning device based on dirt morphology perception and multi-dimensional feature fusion according to claim 2, characterized in that, The determination of the complex stubborn class includes: When the dirt morphology assessment result is complex and stubborn, and the cleaning intention label is static ground-adsorption type, the displacement variance in the dynamic suction characteristics is continuously monitored. If the displacement variance drops below the first preset threshold and continues to exceed the first soaking time, it is determined to be a stubborn stain softening interaction pause state. The control function adjustment mechanism stops water output or reduces the roller brush speed to the preset maintenance speed to ensure that the cleaning liquid fully acts on the stains; when the displacement variance recovers, it determines that the softening is over and instantly restores the high suction power and high speed output.

5. The intelligent cleaning device based on dirt morphology perception and multi-dimensional feature fusion according to claim 2, characterized in that, The AI ​​decision processing module performs the following operations: When visual image data shows that the morphological characteristics of dirt are lower than the physical judgment threshold due to the reflectivity or uniformity of the ground material, if acoustic data or pressure data confirm the physical fluctuations caused by the inhalation of foreign objects, and the negative integrated negative pressure feedback data indicates that the actual negative pressure deviates from the no-load reference value, then the visual judgment limit is crossed, and the current cleaning intention label for dynamic push-pull interactive cleaning is maintained.

6. The intelligent cleaning device based on dirt morphology perception and multi-dimensional feature fusion according to claim 1, characterized in that, The high-speed hydrodynamic unit includes: The stator assembly includes multiple sets of stator core units, which are wound in parallel and assembled to form a closed structure that forms a continuous and uninterrupted magnetic circuit loop. The rotor assembly is located inside the stator assembly; The dynamic balancing mechanism is located at both ends of the rotor assembly and is used to automatically generate a restoring torque that resists the eccentric displacement when the rotor assembly undergoes eccentric displacement due to high-speed rotation.

7. The intelligent cleaning device based on dirt morphology perception and multi-dimensional feature fusion according to claim 6, characterized in that, The dynamic balancing mechanism includes elastic support members disposed at both ends of the magnetic core of the rotor assembly. When the rotor magnetic core tilts, the elastic support member on the side closer to the rotation center is compressed to generate a first reverse elastic force, and the elastic support member on the side farther from the rotation center is stretched to generate a second reverse tensile force. The first reverse elastic force and the second reverse tensile force together constitute the restoring torque, which causes the rotor vibration to decay and return to the equilibrium position.

8. A control method for an intelligent cleaning device based on dirt morphology perception and multi-dimensional feature fusion according to any one of claims 1-7, characterized in that, Includes the following steps: S1. Acquire multi-source sensing data of the cleaning area in real time; S2. Multi-dimensional feature fusion is performed on multi-source sensing data through a dirt morphology detection model to output dirt morphology assessment results and cleaning intention labels. S3. Based on the assessment results of dirt morphology, match the target suction power, target speed and target water output strategy, and execute intermittent start / stop or complete blocking logic based on the cleaning intention label; S4. Perform closed-loop correction control of the fluid power unit based on the deviation between the target suction force and the actual negative pressure.

9. The control method for an intelligent cleaning device based on dirt morphology perception and multi-dimensional feature fusion according to claim 8, characterized in that, Step S3 includes: When the dirt morphology detection model outputs a dirt morphology assessment result of composite stubborn type and the cleaning intention label is static ground adsorption type, if the AI ​​decision processing module detects that the variance value representing the device displacement is lower than the preset displacement threshold and continues to exceed the set softening time threshold, it is determined to be a dwelling softening state. In the softening state, the main control module controls the function adjustment mechanism to stop the output of cleaning fluid and controls the speed of the function adjustment mechanism to reduce to the preset maintenance speed so that the output cleaning fluid can fully adhere to the stubborn stains. When the variance of the device displacement is detected to recover and the proportion of dirt area in the dirt morphology assessment results shows a monotonically decreasing trend, the main control module releases the dwell softening state, controls the high-speed fluid power unit to restore to the maximum target suction, and increases the function adjustment mechanism to the stripping speed.

10. The control method for an intelligent cleaning device based on dirt morphology perception and multi-dimensional feature fusion according to claim 8, characterized in that, Step S3 includes: A weak feature ground fault-tolerant cruise working mode is preset. When the confidence of visual feature extraction is lower than the effective judgment threshold due to the reflectivity of ground material being higher than the preset reflectivity threshold, the weak feature ground fault-tolerant cruise working mode is triggered. In the weak feature ground fault-tolerant cruise working mode, the AI ​​decision processing module forcibly increases the weight ratio of acoustic data and pressure data in multi-dimensional feature fusion; If the acoustic data indicates frequency domain energy fluctuations that match the characteristics of particle impact, or if the pressure data indicates that the actual negative pressure value of the air duct deviates from the no-load reference value to the preset deviation threshold, the main control module will bypass the visual judgment limitation, lock the current cleaning area, and maintain the current suction parameter output of the high-speed fluid power unit. At the same time, the main control module controls the function adjustment mechanism to perform timed pre-spraying actions until the rate of change of the percentage of dirty area output by the dirty morphology detection model is lower than the stagnation threshold, at which point the cleaning is determined to be complete and the lock is released.

Images