A multi - garbage classification device, control and recognition method for self - supervised hybrid domain adaptation

Abstract

The present invention discloses a multi - garbage classification device, control and recognition method for self - supervised hybrid domain adaptation. The multi - garbage classification device includes a frame body, a multi - stage conveying unit, a compression unit, a rotating storage unit and an identification and display unit. The multi - stage conveying unit, the compression unit and the rotating storage unit are sequentially arranged inside the frame body from top to bottom, and the multi - stage conveying unit is located on one side of the central axis of the frame body, while the compression unit is located on the other side of the central axis of the frame body; a feeding port is arranged above the multi - stage conveying unit at the top of the frame body, balls are fixedly arranged at the four corners of the bottom of the frame body, a middle support plate is arranged in the middle of the frame body, a lower bottom plate is arranged at the bottom of the frame body, and a control unit is arranged on one side of the frame body. The present invention improves the success rate of garbage separation and compression efficiency through hybrid domain adaptation technology, and can meet the intelligent garbage classification requirements in complex scenarios such as communities and schools.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent environmental protection equipment. Specifically, it relates to a multi-waste classification device, control and recognition method based on self-supervised hybrid domain adaptation. Background Art

[0002] With the acceleration of urbanization and the improvement of environmental awareness, waste classification has become an important measure to improve the living environment and promote resource recycling. In the field of waste classification technology, traditional devices generally have bottlenecks such as poor adaptability to multiple scenarios, single detection dimension, and lack of interactive feedback. As the core equipment for automatic classification, intelligent waste classification devices have been widely used in public places such as communities and schools. However, due to the wide variety and different shapes of waste, and significant differences in waste characteristics in different regions, higher requirements are put forward for the recognition accuracy and environmental adaptability of classification devices. Especially in crowded places, the characteristics of frequent and random waste disposal make traditional devices face severe challenges.

[0003] Existing devices mostly adopt a single-channel transmission structure, which cannot handle the stacking interference problem caused by multiple wastes being put in at the same time, resulting in misjudgment of the recognition system due to occlusion or composite materials. Most mainstream classification algorithms rely on a single sensor modality (such as weight or optical detection), and it is difficult to handle the feature confusion problem brought by the diversity of waste shapes. For example, when transparent glass coexists with liquid containers, traditional single-sensor detection schemes are difficult to accurately identify. Although some devices improve the classification accuracy by introducing technical means such as a spacing maintaining mechanism or acoustic feature detection, they have not solved the problem of parallel multi-target processing. At the same time, although existing visual recognition technologies have improved the classification accuracy to a certain extent, they have not effectively solved the problem of cross-domain feature migration.

[0004] The technical bottlenecks of traditional devices are mainly reflected in three aspects: First, the adaptability to multiple scenarios is poor. The single-channel transmission structure cannot effectively handle the situation of multiple wastes being put in at the same time, and it is easy to cause misjudgment in recognition due to waste stacking. Second, the detection dimension is single. It is difficult to accurately identify composite material wastes only relying on a single sensor modality, and the lack of an effective domain adaptation mechanism leads to insufficient generalization performance of the model in different scenarios. Finally, the interactive feedback is missing. Most devices lack a real-time feedback system, and users cannot know the classification results and container status, resulting in an operation blind area. These problems seriously restrict the actual application effect of waste classification devices in complex scenarios, and there is an urgent need for a composite solution integrating multi-stage physical separation, intelligent compression storage, and self-supervised hybrid domain adaptation visual detection to break through the existing technical bottlenecks through the collaborative innovation of mechanical structures and algorithms.

[0005] Regarding the problems in the related art, no effective solution has been proposed yet. Summary of the Invention

[0006] In view of the problems in the related art, the present invention proposes a multi - garbage classification device, control and recognition method based on self - supervised hybrid domain adaptation, which has the advantages of mechanical separation optimization, intelligent compression, dynamic recognition feedback and modular architecture, forming significant advantages in terms of classification efficiency, space utilization rate, accuracy and environmental adaptability, and thus solving the problems of garbage accumulation misjudgment, cumbersome operation and high maintenance cost in the prior art.

[0007] For this reason, the specific technical solutions adopted by the present invention are as follows:

[0008] According to one aspect of the present invention, a multi - garbage classification device based on self - supervised hybrid domain adaptation is provided. The multi - garbage classification device based on self - supervised hybrid domain adaptation includes a frame main body, a multi - stage conveyor unit, a compression unit, a rotating storage unit and an identification and display unit. The multi - stage conveyor unit, the compression unit and the rotating storage unit are sequentially arranged inside the frame main body from top to bottom, and the multi - stage conveyor unit is located on one side of the central axis of the frame main body, while the compression unit is located on the other side of the central axis of the frame main body; a feeding port is arranged above the multi - stage conveyor unit at the top of the frame main body, ball bearings are fixedly arranged at the four corners of the bottom of the frame main body, a middle support plate is arranged in the middle of the frame main body, a lower bottom plate is arranged at the bottom of the frame main body, and a control unit is arranged on one side of the frame main body.

[0009] Furthermore, the multi - stage conveyor unit includes a V - shaped conveyor belt assembly arranged on one side of the top of the frame main body;

[0010] The V - shaped conveyor belt assembly includes two groups of symmetrically arranged composite profiles, and the two groups of composite profiles form a V - shaped obtuse - angle structure. A number of linearly distributed V - shaped conveyor belt connectors are arranged at the bottom ends of the two groups of composite profiles, and a vibration motor is commonly connected to one side of the bottoms of the two groups of composite profiles;

[0011] Both ends of the composite profile are provided with V - shaped conveyor belt rollers, and a first stepping motor is arranged at one end of the two groups of V - shaped conveyor belt rollers on the side of the vibration motor.

[0012] Furthermore, the multi - stage conveyor unit further includes a horizontal conveyor belt assembly arranged on the side of the middle of the frame main body away from the V - shaped conveyor belt assembly;

[0013] The horizontal conveyor belt assembly includes two groups of symmetrically arranged horizontal conveyor belt rollers. A horizontal conveyor belt is sleeved outside the horizontal conveyor belt rollers. Bearing seats are arranged at both ends of the horizontal conveyor belt rollers, and a connecting profile is arranged between the two groups of bearing seats;

[0014] A second stepping motor is arranged at the bottom end of the connecting profile. A synchronous pulley is sleeved at the output end of the second stepping motor, and the synchronous pulley is connected to the horizontal conveyor belt roller on the side away from the compression unit through a synchronous belt.

[0015] Further, the compression unit includes an electric push rod assembly disposed on one side of the middle part of the frame body;

[0016] The electric push rod assembly includes a rear support plate disposed on one side of the middle part of the frame body. An electric push rod is disposed on one side of the rear support plate. The output end of the electric push rod is provided with a main compression plate. First optical axes are disposed through both ends of the main compression plate. One end of each first optical axis is provided with an optical axis support frame. A linear motion bearing fixedly connected to the main compression plate is disposed in the middle of each first optical axis;

[0017] The other ends of the two groups of first optical axes are jointly provided with a front support plate. Optical axis support seats matched with the first optical axes are disposed on both sides of the front support plate. A passive compression plate is disposed in the middle of the front support plate.

[0018] Further, the compression unit further includes a sliding door assembly disposed below the electric push rod assembly in the middle part of the frame body;

[0019] The sliding door assembly includes two groups of sliding door tracks symmetrically disposed at the top end of the middle support plate. A sliding groove is formed on one side of each sliding door track. A sliding door is disposed inside the sliding groove;

[0020] A sliding door upper rack is disposed on one side of the top end of the sliding door. A third stepping motor is disposed on one side of the sliding door track far from the multi-stage conveyor unit. A sliding door gear matched with the sliding door upper rack is disposed at the output end of the third stepping motor.

[0021] Further, the rotary storage unit includes a rotary chassis disposed at the bottom of the frame body. A plurality of garbage can bases distributed in a circular pattern are formed at the top end of the rotary chassis. A garbage can body is disposed on the top of each garbage can base. A driven gear is disposed at the bottom end of the rotary chassis; A fourth stepping motor is disposed on one side of the rotary chassis below the compression unit. A driving gear meshed with the driven gear is disposed at the bottom of the output end of the fourth stepping motor;

[0022] An optical axis fastener is disposed through the middle of the rotary chassis. A second optical axis is disposed through the middle of the optical axis fastener. A middle fixing member matched with the optical axis fastener is sleeved outside the middle of the second optical axis. Tapered roller bearings are sleeved at both the top end and the bottom end of the second optical axis. Fixing members are disposed outside the tapered roller bearings.

[0023] According to another aspect of the present invention, a multi-class garbage classification control method for self-supervised hybrid domain adaptation is further provided. The control method includes:

[0024] SI. Differentially separate the input garbage through the V-shaped conveyor belt assembly, and dynamically adjust the differential ratio through the control unit;

[0025] SII. Pause the horizontal conveyor belt assembly, collect static images by using a camera and transmit them to the control unit;

[0026] SIII. The control unit completes garbage classification recognition and outputs the result to the control board when the confidence level reaches the threshold;

[0027] SIV. Select the operation mode according to the classification result. When it is recognized as recyclable garbage, execute the recyclable garbage mode, and sequentially execute the operations of closing the sliding door, compressing the electric push rod, rotating the rotating chassis, and opening the sliding door; when it is recognized as non-recyclable garbage, execute the non-recyclable garbage mode, and sequentially execute the operations of rotating the rotating chassis and opening the sliding door;

[0028] SV. Update the garbage classification information and capacity status through the display screen, and trigger a shutdown warning when the capacity of the garbage bin body exceeds the threshold.

[0029] According to another aspect of the present invention, there is also provided a multi-garbage classification recognition method based on self-supervised hybrid domain adaptation. The recognition method includes:

[0030] S1. Obtain multi-view images of garbage through a camera and perform preprocessing;

[0031] S2. Use the hybrid domain adaptation garbage recognition model to extract image features, realize spatial selective scanning and channel exchange, and update the model parameters using the transfer learning and incremental learning framework;

[0032] S3. Based on the hybrid domain adaptation garbage recognition model, output the garbage classification result and confidence level, and synchronously update them to the display screen and the control unit;

[0033] Among them, S2 includes:

[0034] S21. Construct a hybrid domain adaptation garbage recognition model, and improve the feature extraction ability by embedding a channel attention module and a contrast learning loss function;

[0035] S22. Process the feature map based on the spatial selective scanning mechanism and the channel exchange strategy, and realize spatial selective scanning and channel exchange by calculating the cross-domain feature entropy;

[0036] S23. Combine the transfer learning pre-training weights and the incremental learning framework to complete the dynamic update of the model parameters and improve the recognition performance of garbage categories.

[0037] Furthermore, constructing a hybrid domain adaptation garbage recognition model and improving the feature extraction ability by embedding a channel attention module and a contrast learning loss function includes:

[0038] S211. Extract the shallow features of the source domain and target domain images through two consecutive convolutional blocks, and generate a preliminary feature map through max-pooling downsampling;

[0039] S212. Process the preliminary feature map using depthwise separable convolution, and generate a similarity matrix by measuring the spatial consistency of the source domain features and the target domain feature map through cosine similarity;

[0040] S213. Divide the source domain feature map and the target domain feature map evenly along the channel dimension, calculate the channel-level similarity after swapping some channels, and generate features with spatial and channel alignment;

[0041] The expression for the channel-level similarity is:

[0042] ;

[0043] In the formula, S channel represents the channel-level similarity; , represents the feature vector after splicing the first and second segments of the source domain feature map and the third and fourth segments of the target domain feature map; , represents the feature vector after splicing the first and second segments of the target domain feature map and the third and fourth segments of the source domain feature map; ||·|| represents the norm of the vector.

[0044] Furthermore, process the feature map based on the spatial selective scanning mechanism and the channel swapping strategy, and realize spatial selective scanning and channel swapping by calculating the cross-domain feature entropy, including:

[0045] S221. Use self-attention to independently enhance the intra-domain features, generate source-dominated features and target-dominated features through cross-attention, and achieve soft alignment of the features;

[0046] S222. Transmit the source domain features, source-dominated features, target-dominated features, and target domain features to the next stage, and repeat the feature extraction and alignment process;

[0047] S223. Introduce entropy loss in the shallow, middle, and deep layers, calculate the cross-domain feature entropy through the edge activation function, and strengthen the cross-domain consistency of multi-level features;

[0048] The expression for the cross-domain feature entropy is:

[0049] ;

[0050] ;

[0051] In the formula, H t→ts (l) represents the entropy value of the source-dominated features in the l-th stage; k represents the proportionality coefficient; σ m represents the edge activation function; Z t→ts (l,k) represents the source-dominated features in the l-th stage; Z t (1,k)Represents the target domain features in the first stage; L entropy Represents the total entropy loss; N represents the number of feature layers; l represents the feature layer index; H t→st (l) Represents the entropy value of the target-dominant features in the l-th stage.

[0052] The beneficial effects of the present invention are as follows:

[0053] (1) The efficiency is improved by the multi-stage separation and parallel processing capabilities: Through the dual-stage collaborative design of the differential separation of the V-shaped conveyor belt assembly and the detection of the horizontal conveyor belt assembly, the present invention uses a V-shaped conveyor belt and adaptive differential control to achieve efficient physical separation of multiple types of garbage; at the same time, the V-shaped conveyor belt uses an infrared sensor to detect the stacking density of garbage in real time, and combines the differential drive of two stepper motors to automatically optimize the differential ratio according to the volume difference or the difference in material friction coefficient; in addition, silicone corrugated belts are used in both the horizontal conveyor belt assembly and the V-shaped conveyor belt assembly, and vibration motors are added at the bottom of the carbon fiber composite aluminum profiles. When the separation force is insufficient, high-frequency vibration is triggered to assist, ensuring that the garbage enters the detection area one by one, significantly improving the separation efficiency and the success rate of stacked separation, while reducing energy consumption and extending the service life of the conveyor belt.

[0054] (2) The storage space utilization rate is optimized by the compression function: A linkage system of an electric push rod assembly and a sliding door assembly is designed for recyclable garbage. Through the collaborative work of the active compression plate and the passive compression plate, the loose garbage is compressed into a shape; compared with the traditional device without a compression function, the present invention can significantly reduce the volume of garbage, extend the full-load cycle of the trash can, and reduce the user's dumping frequency, especially suitable for the high-load requirements in public scenarios.

[0055] (3) The classification accuracy is enhanced by intelligent recognition and dynamic interaction: The present invention realizes the coordination of algorithm recognition and mechanical control based on a dual-control system. The wide-angle camera provides high-precision image acquisition, and combines an autonomous classification algorithm to judge the type of garbage in real time; the rotating storage unit dynamically matches the corresponding trash can, and the display screen synchronously feeds back the classification result and the full-load state. Compared with the existing devices that rely on manual selection or a single sensor, the misjudgment rate is significantly reduced and the transparency of user operation is improved.

[0056] (4) The environmental adaptability is improved by the modular structure design: The present invention uses a composite aluminum profile frame and an autonomous PCB control unit to integrate multiple motors, sensors and a power supply system. The structure is compact and has strong anti-interference ability; the design of the V-shaped conveyor belt and the rotating base of the trash can reduces friction loss. Combined with a one-key start-stop switch and a wide-angle camera, it is suitable for various complex indoor and outdoor environments. Compared with traditional fixed or loosely assembled devices, it has both scalability and long-term operation stability. Description of the Drawings

[0057] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0058] Figure 1 It is a schematic structural diagram of a multi - garbage classification device with self - supervised hybrid domain adaptation according to an embodiment of the present invention;

[0059] Figure 2 It is a rear view of a multi - garbage classification device with self - supervised hybrid domain adaptation according to an embodiment of the present invention;

[0060] Figure 3 It is a right view of a multi - garbage classification device with self - supervised hybrid domain adaptation according to an embodiment of the present invention;

[0061] Figure 4 It is a schematic structural diagram of a V - type conveyor belt assembly in a multi - garbage classification device with self - supervised hybrid domain adaptation according to an embodiment of the present invention;

[0062] Figure 5 It is a schematic structural diagram of a horizontal conveyor belt assembly in a multi - garbage classification device with self - supervised hybrid domain adaptation according to an embodiment of the present invention;

[0063] Figure 6 It is a schematic structural diagram of an electric push rod assembly in a multi - garbage classification device with self - supervised hybrid domain adaptation according to an embodiment of the present invention;

[0064] Figure 7 It is a schematic structural diagram of a sliding door assembly in a multi - garbage classification device with self - supervised hybrid domain adaptation according to an embodiment of the present invention;

[0065] Figure 8 It is a schematic structural diagram of a rotating storage unit in a multi - garbage classification device with self - supervised hybrid domain adaptation according to an embodiment of the present invention;

[0066] Figure 9 It is a partial structural schematic diagram of a rotating storage unit in a multi - garbage classification device with self - supervised hybrid domain adaptation according to an embodiment of the present invention;

[0067] Figure 10 It is a schematic flow chart of a multi - garbage classification control method with self - supervised hybrid domain adaptation according to an embodiment of the present invention;

[0068] Figure 11 It is a specific implementation diagram of a multi - garbage classification control method with self - supervised hybrid domain adaptation according to an embodiment of the present invention;

[0069] Figure 12 It is a schematic flow chart of a multi - garbage classification recognition method for self - supervised hybrid domain adaptation according to an embodiment of the present invention.

[0070] In the figure:

[0071] 1. Frame main body; 101. Feeding port; 102. Ball; 103. Middle support plate; 104. Lower bottom plate; 105. Infrared distance sensor; 2. Multi - stage conveying unit; 201. First stepping motor; 202. Composite profile; 203. V - belt roller; 204. V - belt connecting piece; 205. Vibration motor; 206. Synchronous belt; 207. Horizontal belt roller; 208. Horizontal conveyor belt; 209. Second stepping motor; 210. Bearing seat; 211. Connecting profile; 212. Synchronous pulley; 3. Compression unit; 301. Electric push rod; 302. Rear support plate; 303. Active compression plate; 304. Optical axis support frame; 305. Linear motion bearing; 306. First optical axis; 307. Optical axis support seat; 308. Passive compression plate; 309. Front support plate; 310. Third stepping motor; 311. Sliding door gear; 312. Upper rack of sliding door; 313. Sliding door track; 314. Chute; 315. Sliding door; 4. Rotary storage unit; 401. Fourth stepping motor; 402. Active gear; 403. Rotary chassis; 404. Garbage can main body; 405. Garbage can base; 406. Driven gear; 407. Second optical axis; 408. Middle fixing piece; 409. Optical axis fastener; 410. Fixing piece; 411. Tapered roller bearing; 5. Identification and display unit; 501. Display screen; 502. Camera; 6. Control unit; 7. Photoelectric emergency stop sensor. Detailed implementation manners

[0072] To further illustrate each embodiment, the present invention provides accompanying drawings. These drawings are part of the disclosure of the present invention, mainly used to illustrate the embodiments, and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these contents, those of ordinary skill in the art should be able to understand other possible implementation manners and the advantages of the present invention. The components in the figures are not drawn to scale, and similar component symbols are usually used to represent similar components.

[0073] According to an embodiment of the present invention, a multi - garbage classification device, control and recognition method for self - supervised hybrid domain adaptation are provided.

[0074] Now, the present invention will be further described in combination with the accompanying drawings and specific implementation manners, as Figures 1-9As shown, according to an embodiment of the present invention, a multi-class garbage sorting device for self-supervised hybrid domain adaptation is provided. The multi-class garbage sorting device for self-supervised hybrid domain adaptation includes a frame body 1, a multi-stage conveying unit 2, a compression unit 3, a rotating storage unit 4, and an identification and display unit 5. The multi-stage conveying unit 2, the compression unit 3, and the rotating storage unit 4 are sequentially arranged inside the frame body 1 from top to bottom, and the multi-stage conveying unit 2 is located on one side of the central axis of the frame body 1, and the compression unit 3 is located on the other side of the central axis of the frame body 1;

[0075] A feeding port 101 is arranged above the multi-stage conveying unit 2 at the top of the frame body 1. Ball rollers 102 are fixedly arranged at the four corners of the bottom end of the frame body 1. A middle support plate 103 is arranged in the middle of the frame body 1. A lower bottom plate 104 is arranged at the bottom of the frame body 1. A control unit 6 is arranged on one side of the frame body 1.

[0076] Specifically, as Figures 1-3 shown, a multi-class garbage sorting device for self-supervised hybrid domain adaptation proposed by the present invention includes a frame body 1, a multi-stage conveying unit 2, a compression unit 3, a rotating storage unit 4, and an identification and display unit 5; the multi-stage conveying unit 2, the compression unit 3, and the rotating storage unit 4 are sequentially installed inside the frame body 1 from top to bottom; a feeding port 101 is arranged above the V-shaped conveyor belt assembly on the frame body 1. Ball rollers 102 are fixed at the bottom of the frame body 1 to facilitate the movement of the device; the identification and display unit 5 includes a display screen 501 and a camera 502. The display screen 501 is installed above the frame body 1, and the camera 502 required for detection and identification is installed on the frame body 1 above the horizontal conveyor belt assembly; the control unit 6 includes a control board (the model of the control board in this embodiment is STM32F446RCT6) and a development board (the model of the development board in this embodiment is Jetson nano); the rotating storage unit 4 is equipped with four trash can bodies 404, which are respectively used to store recyclable garbage, hazardous garbage, kitchen waste, and other garbage.

[0077] It should be noted that the first stepping motor 201, the vibration motor 205, the second stepping motor 209, the electric push rod 301, the third stepping motor 310, the fourth stepping motor 401, the display screen 501, the camera 502 are electrically connected to the control unit 6. This is the prior art and will not be elaborated here.

[0078] After being powered on, the display screen 501 plays the garbage sorting publicity video in a loop. After multiple pieces of garbage are put in from the feeding port 101, the garbage passes through the V-shaped conveyor belt assembly.

[0079] Specifically, in the above embodiment, the multi-stage conveying unit 2 includes a horizontal conveyor belt assembly and a V-shaped conveyor belt assembly. The V-shaped conveyor belt assembly is composed of two conveyor belts arranged at an angle of 110° to 130°. It is differentially driven by two groups of first stepping motors 201 (in this embodiment, the first stepping motor is a 42-series two-phase stepping motor, model 42HS402004D, rated torque 0.4 N·m, peak current 1.5 A) (in this embodiment, the differential ratio of the two groups of first stepping motors is set to 1:1.2 to 1.8). The surface of the conveyor belt is coated with a polyurethane anti-slip layer (in this embodiment, the Shore hardness of the polyurethane anti-slip layer is 80A, and the friction coefficient is μ = 0.6). A vibration motor 205 is integrated at the bottom (in this embodiment, the model of the vibration motor is YZD206, frequency 30 Hz ± 5%, amplitude 1.5 mm); the horizontal conveyor belt assembly is driven by a second stepping motor 209 (in this embodiment, the second stepping motor is a 42-series two-phase stepping motor, model 42HS402004D), and a one-way baffle is provided at the end (in this embodiment, the inclination angle of the one-way baffle with the vertical direction is set to 15°), and the transmission speed is 0.25 to 1.25 m / s.

[0080] Specifically, in the above embodiment, the compression unit 3 includes an electric push rod assembly and a sliding door assembly. Electric push rod assembly: An electric push rod 301 (in this embodiment, the specification of the electric push rod is 24V / 1000N) drives the active compression plate 303, and a thin film pressure sensor is provided at the front end of the compression plate (in this embodiment, the range of the thin film pressure sensor is 0 to 1500N, accuracy ±1%); Sliding door assembly: It is controlled by a third stepping motor 310 (in this embodiment, the third stepping motor is a 42-series two-phase stepping motor, model 42HS402004D) through a gear and rack mechanism (in this embodiment, the module of the gear and rack mechanism is 1.5) to linearly move the sliding door (in this embodiment, the stroke of the sliding door is 120mm);

[0081] Specifically, in the above embodiment, the rotary storage unit 4 includes a fourth stepping motor 401 (in this embodiment, the fourth stepping motor is a 57-series three-phase stepping motor) driving a rotary chassis 403, positioned by tapered roller bearings 411 and a second optical axis 407 (in this embodiment, the diameter of the second optical axis is 17mm), and 32 Hall positioning slots are evenly distributed circumferentially on the chassis;

[0082] Specifically, in the above embodiment, the identification and display unit 5 includes a camera 502 (in this embodiment, the camera is a 4K wide-angle camera, field of view angle 120°), and an infrared ranging sensor 105 (in this embodiment, the specification of the infrared ranging sensor is detection accuracy ±2mm);

[0083] Control Unit 6: It adopts a dual-core architecture composed of a control board (the model of the control board in this embodiment is STM32F446RCT6) and a development board (the model of the development board in this embodiment is Jetson nano). The control board is connected to the actuator through a CAN bus (the baud rate of the CAN bus in this embodiment is 1 Mbps).

[0084] Frame Body 1: In this embodiment, an aluminum profile body of European standard 2020 is adopted, and ball bearings 102 (the diameter of the ball bearings in this embodiment is 25 mm) and a 6S lithium battery compartment (the capacity of the 6S lithium battery compartment in this embodiment is 10,000 mAh) are provided at the bottom.

[0085] In one embodiment, the multi-stage conveying unit 2 includes a V-shaped conveyor belt assembly provided on one side of the top of the frame body 1.

[0086] The V-shaped conveyor belt assembly includes two groups of symmetrically arranged composite profiles 202, and the two groups of composite profiles 202 form a V-shaped obtuse angle structure. A number of linearly distributed V-shaped conveyor belt connectors 204 are provided at the bottom ends of the two groups of composite profiles 202, and a vibration motor 205 is commonly connected to one side of the bottoms of the two groups of composite profiles 202.

[0087] V-shaped conveyor belt rollers 203 are provided at both ends of the composite profile 202, and a first stepping motor 201 is provided at one end of each of the two groups of V-shaped conveyor belt rollers 203 on the side of the vibration motor 205.

[0088] In one embodiment, the multi-stage conveying unit 2 further includes a horizontal conveyor belt assembly provided on one side of the middle part of the frame body 1 away from the V-shaped conveyor belt assembly.

[0089] The horizontal conveyor belt assembly includes two groups of symmetrically arranged horizontal conveyor belt rollers 207. A horizontal conveyor belt 208 is sleeved outside the horizontal conveyor belt rollers 207. Bearing seats 210 are provided at both ends of the horizontal conveyor belt rollers 207, and a connecting profile 211 is provided between the two groups of bearing seats 210.

[0090] A second stepping motor 209 is provided at the bottom end of the connecting profile 211. A synchronous pulley 212 is sleeved on the output end of the second stepping motor 209, and the synchronous pulley 212 is connected to the horizontal conveyor belt roller 207 on the side away from the compression unit 3 through a synchronous belt 206.

[0091] Specifically, as Figure 3 and Figure 4As shown in the figure, the V-shaped conveyor belt assembly includes two first stepping motors 201. The output shafts of the two first stepping motors 201 are respectively connected to two groups of V-shaped conveyor belt rollers 203 through couplings, so as to drive the two groups of V-shaped conveyor belt rollers 203 to rotate (in this embodiment, the reduction ratio between the output shaft of the first stepping motor 201 and the V-shaped conveyor belt roller 203 is set to 1:2). The two sides of the conveyor belt are fixed by a composite profile 202 (in this embodiment, the composite profile uses CF-AL6061 carbon fiber composite aluminum profile) and form an included angle of 110°-130° to drive the rotation of the conveyor belt. A pressure sensor and a silicone corrugated belt serving as the conveyor belt are arranged on the surface of the composite profile 202; the first stepping motor 201 adopts a closed-loop control mode (in this embodiment, the subdivision accuracy of the closed-loop control mode is 1600 pulses / revolution), and the control board adjusts the differential ratio in real time (in this embodiment, the differential ratio of the two first stepping motors is 1:1.2-1:5). The corresponding conveyor belt linear speed range is 0.25 m / s and 0.3-1.25 m / s, so that the stacked garbage generates a 5-15 N lateral separation force due to the speed difference. The surface of the silicone corrugated belt is provided with fishbone-shaped anti-slip patterns, and a vibration motor 205 is integrated at the bottom of the composite profile 202 to assist in separating the adhered garbage; the V-shaped conveyor belt connector 204 is forged from 7075-T6 aluminum alloy and is fixed to the European standard 2020 aluminum profile of the frame body 1 through M5 high-strength bolts. After testing, this assembly can separate 5 pieces of stacked garbage within 2 seconds (in this embodiment, the size of the garbage is 50-300 mm and the weight is 10-500 g), the separation success rate reaches 98.5%, and the continuous operation life ≥ 50,000 times.

[0092] Specifically, the separated garbage falls onto the horizontal conveyor belt assembly in sequence, as Figure 5 shown. The horizontal conveyor belt assembly is driven by a second stepping motor 209 to drive the synchronous belt 206 to rotate through a rotating synchronous pulley 212. The rotation of the synchronous belt 206 drives the horizontal conveyor belt roller 207 to rotate through meshing, and the horizontal conveyor belt roller 207 drives the horizontal conveyor belt 208 to rotate through friction. Moreover, the horizontal conveyor belt 208 also uses a silicone corrugated belt, and the surface of the silicone corrugated belt is provided with fishbone-shaped anti-slip patterns, so that the horizontal conveyor belt 208 realizes transportation through the friction force generated by the surface and the garbage.

[0093] Specifically, before transporting the garbage, the horizontal conveyor belt assembly remains stationary. When garbage falls onto the horizontal conveyor belt assembly, the camera 502 detects it, and at the same time, the V-shaped conveyor belt assembly remains stationary to prevent multiple pieces of garbage from entering the horizontal conveyor belt assembly. When the camera 502 finishes detecting, the horizontal conveyor assembly starts to move and transports the garbage into the compression unit 3.

[0094] Specifically, the camera 502 obtains the images on the horizontal conveyor belt assembly in real time and feeds them back to the development board. The development board processes, identifies, and analyzes the obtained photos, and sends the analysis results to the control board. The control board controls the movement of the V-shaped conveyor belt assembly, the horizontal conveyor belt assembly, the subsequent compression unit 3, and the rotary storage unit 4.

[0095] Specifically, when the analysis result of the development board is recyclable waste (such as paper cups, iron cups, and plastic bottles, etc.), the control board controls the movement of the compression unit 3. Refer to Figure 6 and Figure 7 , in the electric push rod assembly, the electric push rod 301 is fixed to the rear support plate 302 by bolts. A passive compression plate 308 is fixed in front of the electric push rod 301 to increase the compression surface. Both sides of the active compression plate 303 are fixed by linear motion bearings 305 (in this embodiment, the linear motion bearing is a linear bearing edge flange motion bearing LUU). To enhance stability, the rear end of the first optical axis 306 (the diameter of the first optical axis in this embodiment is 8 mm) is fixed by an optical axis support frame 304, and the front end is fixed to the front support plate 309 by a linear optical axis support seat 307 (the model of the linear optical axis support seat in this embodiment is SHFSHF8), so as to realize the movement of the active compression plate 303 along the first optical axis 306; the passive compression plate 308 is fixed to the front support plate 309 by bolts.

[0096] In one embodiment, the compression unit 3 includes an electric push rod assembly disposed on one side of the middle part of the frame body 1;

[0097] The electric push rod assembly includes a rear support plate 302 disposed on one side of the middle part of the frame body 1. An electric push rod 301 is disposed on one side of the rear support plate 302. The output end of the electric push rod 301 is provided with an active compression plate 303. Both ends of the active compression plate 303 are penetrated with a first optical axis 306. One end of the first optical axis 306 is provided with an optical axis support frame 304, and a linear motion bearing 305 fixedly connected with the active compression plate 303 is disposed in the middle of the first optical axis 306;

[0098] The other ends of the two groups of first optical axes 306 are jointly provided with a front support plate 309, and optical axis support seats 307 matched with the first optical axes 306 are disposed on both sides of the front support plate 309. A passive compression plate 308 is disposed in the middle of the front support plate 309.

[0099] In one embodiment, the compression unit 3 further includes a sliding door assembly disposed below the electric push rod assembly in the middle of the frame body 1;

[0100] The sliding door assembly includes two groups of sliding door tracks 313 symmetrically disposed at the top of the middle support plate 103. A chute 314 is opened on one side of the sliding door track 313, and a sliding door 315 is disposed inside the chute 314;

[0101] On one side of the top end of the sliding door 315, there is a rack on the sliding door 312, and on one side of the sliding door track 313 far from the multi-stage conveying unit 2, there is a third stepping motor 310. The output end of the third stepping motor 310 is provided with a sliding door gear 311 that mates with the rack on the sliding door 312.

[0102] Specifically, as Figure 7 shown, the sliding door assembly is driven by the third stepping motor 310 to rotate the sliding door gear 311; the sliding door gear 311 meshes with the rack on the sliding door 312, thereby driving the sliding door 315 to move back and forth within the sliding door track 313. The electric push rod assembly is fixed to the frame body 1 through the rear support plate 302 and the front support plate 309; the sliding door assembly is fixed to the middle support plate 103 of the frame body 1 through the sliding door track 313.

[0103] Specifically, as Figure 6 shown, after the compression unit 3 receives the information from the control board, the third stepping motor 310 in the electric push rod assembly operates, driving the sliding door 315 to move towards one end close to the horizontal conveyor belt assembly. At this time, the sliding door 315 covers the falling port on the middle support plate 103, thus closing the downward sliding channel of the garbage; the electric push rod 301 starts to move, driving the active compression plate 303 to move forward to compress the recyclable garbage; when the compression is completed, the electric push rod 301 retracts, and the third stepping motor 310 rotates in the reverse direction, driving the rack on the sliding door 312 to retreat, driving the sliding door 315 to move towards the end far from the horizontal conveyor belt assembly. At this time, the sliding door 315 moves away, and the recyclable garbage automatically falls to the rotating storage unit 4 under the action of gravity through the falling port on the middle support plate 103.

[0104] Specifically, in the above embodiment, the electric push rod assembly is provided with a pressure feedback system: a thin film pressure sensor is arranged at the front end of the active compression plate 303 (the model of the thin film pressure sensor in this embodiment is FlexiForce A201), and the sampling rate is 1 kHz; two-stage compression thresholds are set: the first threshold of 500 N triggers slow-speed compression (the speed drops to 50%), and the second threshold of 800 N triggers emergency stop protection; after the compression is completed, 3 times of vibration rebound are performed (the frequency of the vibration rebound in this embodiment is 20 Hz, and the amplitude is 1 mm).

[0105] In one embodiment, the rotating storage unit 4 includes a rotating chassis 403 arranged at the bottom of the frame body 1. A plurality of garbage can bases 405 distributed in a circular pattern are opened at the top end of the rotating chassis 403. A garbage can body 404 is arranged on the top of the garbage can base 405. A driven gear 406 is arranged at the bottom end of the rotating chassis 403; on one side of the rotating chassis 403 below the compression unit 3, there is a fourth stepping motor 401. The bottom of the output end of the fourth stepping motor 401 is provided with a driving gear 402 that meshes with the driven gear 406;

[0106] A optical axis fastener 409 is disposed through the middle of the rotary chassis 403. A second optical axis 407 is disposed through the middle of the optical axis fastener 409. A middle fixing member 408 that cooperates with the optical axis fastener 409 is sleeved outside the middle of the second optical axis 407. Tapered roller bearings 411 are sleeved at both the top end and the bottom end of the second optical axis 407. A fixing member 410 is disposed outside the tapered roller bearings 411.

[0107] Specifically, before the recyclable waste drops, the control board controls the movement of the rotary storage unit 4 according to the received waste type information. Refer to Figure 8 With Figure 9 , in the rotary storage unit 4, the fourth stepping motor 401 drives the driving gear 402 to rotate. The driving gear 402 meshes with the driven gear 406. The rotary chassis 403 is fixed at the top of the driven gear 406; the entire rotary chassis 403 is fixed by the middle fixing member 408 and the fixing member 410 at the bottom end, and is tightly fixed to the second optical axis 407 through the optical axis fastener 409. When the rotary chassis 403 rotates, it drives the second optical axis 407 to rotate at the same time; the upper part of the second optical axis 407 is fixed to the middle support plate 103 of the frame body 1 through the tapered roller bearing 411 at the top end and the fixing member 410, and the lower part of the second optical axis 407 is fixed to the lower bottom plate 104 of the frame body 1 through the fixing member 410 at the bottom end and the tapered roller bearing 411 at the bottom end; the trash can body 404 is placed above the rotary chassis 403 for easy picking and placing; several infrared ranging sensors 105 distributed in a circle are installed on the middle support plate 103 of the frame body 1, which can detect the full load condition of the trash can body 404 for subsequent normal operation.

[0108] Specifically, after the rotary storage unit 4 rotates to a predetermined position, the waste naturally drops into the corresponding trash can body 404. At the same time, the display screen 501 displays the corresponding waste information, including: waste type, waste quantity, whether the waste has been put in place, etc. When the infrared ranging sensor 105 detects that the waste capacity in the trash can body 404 exceeds three-quarters of the trash can capacity, the infrared ranging sensor 105 sends information to the control board. The control board controls all devices to pause all operations and transmits the full load information to the development board. The development board sends the full load information to the display screen 501 for display.

[0109] Specifically, in the above embodiment, the positioning system of the rotary storage unit 4 includes: a driver of the fourth stepping motor 401 (the driver model in this embodiment is TMC5160 stepping motor driver) for realizing 256 micro-stepping control; an electromagnetic locking mechanism (the holding torque of the electromagnetic locking mechanism in this embodiment is ≥2 N·m) for starting after rotating in place; an RFID tag (the set frequency in this embodiment is 13.56 MHz) disposed on the trash can base 405 for identifying the trash can type.

[0110] Specifically, the control unit 6 in the present invention is provided with three - level safety protection: a) Mechanical protection: An optoelectronic emergency stop sensor 7 is provided at the entrance of the conveyor belt (in this embodiment, the response time of the optoelectronic emergency stop sensor is <50 ms); b) Electrical protection: Leakage protection (in this embodiment, the operating current of the leakage protection is 30 mA) and overload protection (in this embodiment, the threshold current of the overload protection is 5 A); c) Data protection: AES256 encryption is used to transmit data, and local storage is retained for ≥30 days.

[0111] Specifically, the frame body 1 in the present invention adopts a modular design: The V - type conveyor belt assembly is installed through a quick - release interface (in this embodiment, the quick - release interface uses an M8 bolt), and the included - angle adjustment of 60° - 150° is supported; The trash can base is provided with an expansion interface (in this embodiment, the expansion interface is TypeC), and the adaptation of 26 types of trash containers is supported; The top is integrated with a solar power supply interface (in this embodiment, the solar power supply interface is an MPPT controller, and the conversion efficiency is ≥95%).

[0112] Specifically, in the present invention, the shape photo data of the trash is obtained through the recognition and display unit 5, and the types of trash are recognized through the internal algorithm model of the development board (i.e., the multi - garbage classification recognition method of self - supervised hybrid domain adaptation in another embodiment); Through the cooperation of the development board and the control board, the movement of the whole device is further controlled to realize the recognition, classification and storage of multiple types of trash. Compared with the classification only through shape data, volume data and weight data, the recognition accuracy is higher, and the precise cooperation between the upper computer and the lower computer of the device makes the classification more accurate. The present invention solves the industry pain points of low recognition accuracy and limited classification efficiency of traditional garbage classification equipment in complex scenarios through the deep integration of hardware cooperation and intelligent algorithms.

[0113] As Figures 10-11 shown, according to another embodiment of the present invention, a multi - garbage classification control method of self - supervised hybrid domain adaptation is further provided, and the control method includes:

[0114] SI. Differentially separate the input trash through the V - type conveyor belt assembly, and dynamically adjust the differential ratio through the control unit;

[0115] SII. Pause the horizontal conveyor belt assembly, use the camera to collect static images and transmit them to the control unit;

[0116] SIII. The control unit completes the garbage classification recognition, and when the confidence level reaches the threshold, the result is output to the control board;

[0117] SIV. Select the operation mode according to the classification result. When it is identified as recyclable waste, execute the recyclable waste mode, and sequentially execute the operations of closing the sliding door, compressing the electric push rod, rotating the rotating chassis, and opening the sliding door; when it is identified as non-recyclable waste, execute the non-recyclable waste mode, and sequentially execute the operations of rotating the rotating chassis and opening the sliding door.

[0118] SV. Update the garbage classification information and capacity status through the display screen, and trigger a shutdown warning when the capacity of the trash can body exceeds the threshold.

[0119] Specifically, ①Differentially separate multiple simultaneously input garbage through the V-shaped conveyor belt assembly, and dynamically adjust the differential ratio to 1:1.2~1:1.8 (automatically matched according to the material friction coefficient and volume difference);

[0120] ②When the horizontal conveyor belt assembly pauses for 3 seconds, the camera 502 captures a static image and transmits it to the development board (delay ≤ 50ms);

[0121] ③The development board completes the classification within 500ms, and outputs the result to the control board when the confidence level ≥ 0.85;

[0122] ④Select the operation mode according to the classification result:

[0123] Recyclable waste mode: Close the sliding door 315 (in this embodiment, the sliding door response time is set < 100ms) → Start the 24V / 1000N electric push rod 301 to compress (in this embodiment, the compression force of the electric push rod is 200 ± 5N, and the stroke is 150mm) → Rotate the rotating chassis 403 to the corresponding trash can (in this embodiment, the positioning accuracy is set ± 0.5°) → Open the sliding door 315;

[0124] Non-recyclable waste mode: Directly rotate the rotating chassis 403 to the corresponding trash can (in this embodiment, the set rotational angular velocity is 90° / s) → Open the sliding door 315;

[0125] ⑤The display screen 501 updates the classification information and the trash can capacity status in real time, and the infrared ranging sensor 105 triggers a shutdown warning when the detected capacity ≥ 75% (in this embodiment, the set beep frequency is 2kHz).

[0126] As Figure 12 shown, according to another embodiment of the present invention, a multi-garbage classification recognition method based on self-supervised mixed-domain adaptation is further provided. The recognition method includes:

[0127] S1. Obtain multi-view images of the garbage through the camera 502 and perform preprocessing;

[0128] S2. Use the mixed-domain adaptation garbage recognition model to extract image features, realize spatial selective scanning and channel exchange, and update the model parameters using the transfer learning and incremental learning framework.

[0129] S3. Output the garbage classification results and confidence levels based on the hybrid domain adaptation garbage recognition model, and synchronously update them to the display screen and the control unit;

[0130] Specifically, ① Obtain multi-view images of the garbage through the camera 502, with a resolution of not less than 3840×2160 and a frame rate ≥ 30fps;

[0131] ② Preprocess the images, including deblurring (using Wiener filter in this embodiment), illumination equalization (using CLAHE algorithm in this embodiment), and background segmentation (using MaskRCNN model in this embodiment);

[0132] ③ Use the hybrid domain adaptation garbage recognition model (using the improved YOLOv8 model in this embodiment) for feature extraction. The model embeds the SE attention module (the channel compression ratio of the SE attention module is 16:1 in this embodiment) and adds a contrastive learning term to the loss function (the temperature coefficient τ of the contrastive learning loss is 0.07 in this embodiment);

[0133] ④ Load the pre-trained weights through transfer learning (using the ImageNet dataset in this embodiment), and update the parameters of the fully connected layer using the incremental learning framework (the learning rate of the incremental learning optimizer is set to 1e - 4, and the momentum coefficient is set to 0.9 in this embodiment);

[0134] ⑤ Output the classification results and confidence levels (the threshold of the model prediction confidence level is set to 0.85 in this embodiment), and synchronously update them to the display screen 501 and the control board.

[0135] In one embodiment, use the hybrid domain adaptation garbage recognition model to extract image features, implement spatial selective scanning and channel exchange, and update the model parameters using the transfer learning and incremental learning frameworks, including:

[0136] S21. Build a hybrid domain adaptation garbage recognition model to improve the feature extraction ability by embedding a channel attention module and a contrastive learning loss function;

[0137] S22. Process the feature map based on the spatial selective scanning mechanism and the channel exchange strategy, and achieve spatial selective scanning and channel exchange by calculating the cross-domain feature entropy;

[0138] S23. Combine the transfer learning pre-trained weights and the incremental learning framework to complete the dynamic update of the model parameters and improve the recognition performance of garbage categories.

[0139] Specifically, a) Build a spatial selective scanning mechanism to extract cross-domain spatial features through depthwise separable convolution (the convolution kernel size kernel = 3×3 and the stride stride = 1 are set in this embodiment);

[0140] b) Implement the channel exchange strategy, evenly divide the feature map into four segments along the channel dimension, and calculate the channel-level similarity (in this embodiment, weight update is triggered when the cosine similarity > 0.7);

[0141] c) Introduce an entropy loss layer before the detection head to calculate the cross-domain feature entropy value, and the expression is:

[0142] ;

[0143] In the formula, L entropy represents the entropy loss value; i = 2, 4, 6 represents the selected feature layer index; E[] represents the mathematical expectation; σ m represents the margin activation function (in this embodiment, the MarginReLU function is adopted, and the threshold θ = 0.2); F (i) represents the feature map output of different layers;

[0144] d) When updating the model parameters every week, retain 10% of the historical data (in this embodiment, the FIFO strategy is adopted), and the recognition accuracy of new categories is improved by ≥ 15%.

[0145] In one embodiment, to build a hybrid domain adaptation garbage recognition model and improve the feature extraction ability by embedding a channel attention module and a contrast learning loss function, it includes:

[0146] S211. Extract the shallow features of the source domain and target domain images through two consecutive convolutional blocks, and generate a preliminary feature map through max-pooling downsampling;

[0147] S212. Process the preliminary feature map using depthwise separable convolution, measure the spatial consistency of the source domain features and target domain feature maps through cosine similarity, and generate a similarity matrix;

[0148] S213. Evenly divide the source domain feature map and target domain feature map along the channel dimension, exchange some channels, and calculate the channel-level similarity to generate spatially and channel-aligned features;

[0149] The expression of the channel-level similarity is:

[0150] ;

[0151] In the formula, S channel represents the channel-level similarity; , represents the feature vector after splicing the 1st and 2nd segments of the source domain feature map and the 3rd and 4th segments of the target domain feature map; , represents the feature vector after splicing the 1st and 2nd segments of the target domain feature map and the 3rd and 4th segments of the source domain feature map; ||·|| represents the norm of the vector.

[0152] Specifically, the present invention adopts a hybrid-domain adaptation method combined with Mamba-Transformer to realize real-time garbage monitoring in the open world, enabling the model to adapt to the regionality of garbage samples. Let D s ={(x s , B s , C s )} represent a set of labeled images in the source domain, where B s and C s respectively represent the corresponding bounding box and class label of the source image x s . In the target domain, there is D t = {x t}, which contains N t images x t without bounding and class labels. The goal of recognition is to develop a domain-adaptive open-world garbage detection algorithm using the information of D s and D t .

[0153] Specifically, in the Step1 stage: the shallow feature extraction stage, the model receives images from the source domain and the target domain as inputs. First, low-level features are extracted through two convolutional blocks. Each convolutional block contains a convolutional layer, batch normalization (BatchNorm), and a ReLU activation function. The size of the input image is (H, W, C), where H and W respectively represent the height and width of the image, and C represents the number of channels. The convolutional kernel size is 3×3, and the stride is 1. The size of the output feature map is (H, W, C1), where C1 is the number of output channels of the convolutional layer in the first stage. Max pooling is used for downsampling, with a pooling kernel size of 2×2 and a stride of 2. The size of the output feature map is (H / 2, W / 2, C1), and its operation can be expressed as:

[0154] ;

[0155] In the formula, represents the output of the first stage in the source domain and the target domain; ConvBlock represents the convolutional block; x s,t represents the images in the source domain and the target domain, represents the set of real numbers.

[0156] Specifically, in the Step2 stage: the convolutional layer is used again to extract low-level spatial features, and the size of the output feature map is (H / 2, W / 2, C2), where C2 is the number of output channels of the convolutional layer in the second stage. Max pooling is used for downsampling, and the size of the output feature map is (H / 4, W / 4, C2), and its operation can be expressed as:

[0157] ;

[0158] In the formula, represents the output of the second stage of the source domain and the target domain; the convolutional block captures low-level spatial features such as edges and corners through local receptive fields, providing a basis for subsequent domain adaptation. Here, the inductive bias of convolution is used to retain the local structure, avoiding the premature loss of fine-grained information in the deep network and laying a foundation for domain adaptation.

[0159] Specifically, in Step3: Advanced feature extraction and alignment. After the feature maps of the source domain and the target domain processed in the first two convolutional stages , they are input into the hybrid domain adaptation module composed of Mamba-Transformer. The hybrid domain adaptation module sequentially passes through the domain adaptation Mamba module, the attention mechanism, and feature transfer and output. Specifically, the following steps:

[0160] Specifically, in Step31: The domain adaptation Mamba module. The domain adaptation Mamba module includes spatial selective scanning and channel exchange dimension mixing to adaptively model domain changes. Specifically: The control is independently applied to the source and target feature maps through depth convolution, retaining detailed spatial cues.

[0161] Specifically, in the Step311 stage, the spatial selective scanner calculates the cross-domain spatial similarity: for the source domain feature and the target domain feature depthwise separable convolutions are respectively performed on them, and the spatial consistency of the two-domain feature maps is measured through cosine similarity to generate a similarity matrix. This process focuses on the shared spatial patterns to capture the local structure information crucial for domain adaptation. The expression of the similarity matrix is:

[0162] ;

[0163] In the formula, DWConv represents the separable convolution operation, represents the cosine similarity calculation for the feature vectors after the first and second stages;

[0164] Based on the similarity, the feature maps are reweighted to enhance the spatial consistency region, enhance the shared spatial pattern, and suppress the domain-specific noise. The weighting expression is:

[0165] ;

[0166] ;

[0167] In the formula, ⊙ represents element-wise multiplication; represents the output feature of the source domain in the third stage after similarity weighting; represents the output feature of the source domain in the second stage; Represents the output of the target domain in the third stage after similarity weighting; Represents the output features of the target domain in the second stage;

[0168] The weighted features focus on the local structures shared across domains (such as texture and shape), enhancing the feature fusion ability in the intermediate layer, thereby improving the domain adaptation performance. Output the spatially aligned features.

[0169] Specifically, Step312: Channel-swapped spatial model, input the source domain feature map and the target domain feature map Are evenly divided into four segments along the channel dimension, , calculate the channel-level similarity after swapping some channels:

[0170] ;

[0171] The channel similarity vector S channel Is used to adjust the channel weights and suppress domain-specific noise, and the expression is:

[0172] ;

[0173] ;

[0174] In the formula, ChannelAttn represents the channel attention function; Represents adjusting the weights based on the channel similarity vector in the third stage; Represents the outputs of the source domain and the target domain of the features after using cross-attention.

[0175] Specifically, Step313: Features aligned in space and channel and .

[0176] In one embodiment, process the feature map based on the spatial selective scanning mechanism and the channel swapping strategy, and realize spatial selective scanning and channel swapping by calculating the cross-domain feature entropy, including:

[0177] S221. Independently enhance the features within the domain using self-attention, generate source-dominated features and target-dominated features through cross-attention, and achieve soft alignment of the features;

[0178] S222. Transmit the source domain features, source-dominated features, target-dominated features, and target domain features to the next stage, and repeat the feature extraction and alignment process;

[0179] S223. Introduce entropy loss in the shallow, middle, and deep layers, calculate the cross-domain feature entropy through the edge activation function, and strengthen the cross-domain consistency of multi-level features;

[0180] The expression of the cross-domain feature entropy is:

[0181] ;

[0182] ;

[0183] In the formula, H t→ts (l) represents the entropy value of the source-dominated feature in the l-th stage; k represents the proportionality coefficient; σ m represents the edge activation function, which is used to suppress negative responses; Z t→ts (l,k) represents the source-dominated feature in the l-th stage; Z t (1,k) represents the target domain feature in the 1st stage; L entropy represents the total entropy loss; N represents the number of feature layers; l represents the feature layer index, including the shallow layer (l = 2), the middle layer (l = 4), and the deep layer (l = 6); H t→st (l) represents the entropy value of the target-dominated feature in the l-th stage.

[0184] Specifically, in the Step32 stage: The features obtained through the spatial scanning mechanism are fed into the self-attention module, where the features within the domain are independently enhanced to enhance spatial consistency:

[0185] ;

[0186] ;

[0187] In the formula, the source domain key values are respectively , and the target domain key value is , which is used for internal feature refinement. SelfAtten represents the self-attention operation. The self-attention operation expression is as follows:

[0188] ;

[0189] The cross-attention generates the source-dominated feature Z t→s and the target-dominated feature Z s→t through cross-domain interaction of key-value pairs to softly align the feature distributions and blur the domain boundaries:

[0190] ;

[0191] Here, Crossattn represents the cross-attention operation, and the cross-attention operation is as follows:

[0192] ;

[0193] Wherein, CrossAttn represents the result of the cross-attention operation; Q represents the query matrix; K represents the key matrix; V represents the value matrix; d represents the feature dimension; softmax represents the softmax function; T represents the matrix transpose;

[0194] Here, when calculating , Q comes from the source domain, Q = Z s , K and V come from the target domain features, K = Z t , V = Z t , the calculation is the opposite.

[0195] Specifically, in the Step33 stage, four groups of feature streams (Z s , Z t→s , Z s→t , Z t ) are passed to the next stage, but only the source-dominated feature Z t→s is used for the final detection.

[0196] Specifically, in the Step4 - 6 stages: Repeat the Step3 stage to gradually deepen the feature abstraction and alignment.

[0197] Specifically, in the Step7 stage: Based on the low-level features and high-level features, use the SSD one-way detector for training. At the same time, introduce multi-level optimization in stages 2 (shallow), 4 (middle), and 6 (deep) to calculate the entropy loss of each stage (shallow, middle, deep): Suppress negative responses through the MarginReLU function σm(x)=, and calculate the cross-domain feature entropy:

[0198] ;

[0199] Wherein, represents the l-th layer feature entropy from the target domain to the source domain; k represents the scaling coefficient (usually a positive constant); σ m represents the margin activation function (MarginReLU function is adopted in this embodiment); represents the k-th feature of the l-th layer from the target domain to the source domain;

[0200] The total entropy loss is the mean of each layer, and this loss forces the model to reduce feature uncertainty and strengthen cross-domain consistency:

[0201] ;

[0202] Wherein, L entropy represents the total entropy loss; N represents the number of selected feature layers; l = 2, 4, 6 represents the selected feature layer indices; represents the l-th layer feature entropy from the target domain to the source domain; represents the l-th layer feature entropy from the target domain to the source domain.

[0203] In one embodiment, by combining pre-trained weights of transfer learning and an incremental learning framework to complete dynamic update of model parameters and improve the recognition performance of new categories, the steps include:

[0204] S231. Input the optimized source-dominant features into the detection head for training to generate bounding box regression results and class scores;

[0205] S232. Optimize domain-invariant features through an end-to-end training method to achieve robust detection in the target domain;

[0206] S233. Adopt an incremental learning strategy to dynamically update model parameters, while maintaining historical knowledge and adapting to new categories.

[0207] Specifically, in the Step8 stage: Add detection head training. The source-dominant features output in the sixth stage are input into the detection head of the Single Shot Detector (SSD) to generate bounding box regression results Bpred and class scores Cpred. Through end-to-end training, the model gradually optimizes domain-invariant features and finally achieves robust detection in the target domain. In each stage, through fine-grained alignment of the Spatial Selective Scanning Mechanism (SSM), cross-domain interaction of the attention mechanism, and joint optimization of entropy and adversarial, the domain shift problem is systematically solved.

[0208] Specifically, after the classification device of the present invention is started, the display screen 501 will automatically play a garbage classification publicity video, with a space occupation ratio of 3 / 4, and the information to be classified is left on the right. When garbage is put in, the V-shaped conveyor belt assembly starts to work. Through the differential rotation of two first stepping motors 201, the synchronous belt 206 is driven to rotate to complete garbage separation and transportation; when garbage enters the horizontal conveyor belt assembly from the V-shaped conveyor belt assembly, the V-shaped conveyor belt assembly immediately stops working and enters the detection link.

[0209] Specifically, the camera 502 takes pictures of the garbage on the horizontal conveyor belt assembly and transmits it to the development board for processing. Through a hybrid domain adaptation method combining Mamba-Transformer, the recognition and classification of the taken garbage photos are realized, and the classification information is transmitted to the control board. The control board controls the movement of the horizontal conveyor belt assembly, the compression unit 3, and the rotary storage unit 4 according to the information.

[0210] Specifically, according to the recognized types of garbage, it is divided into: recyclable garbage, kitchen waste, hazardous waste, and other waste; if the recognized type is recyclable garbage, the control board controls the movement of the horizontal conveyor belt assembly, the sliding door 315 closes, the electric push rod 301 compresses, the rotary storage unit 4 rotates until the recyclable trash can coincides with the lowering port, and after the compression is completed, the sliding door 315 opens to complete the classification; if the recognized type is non-recyclable garbage, the control board controls the horizontal conveyor belt to move, the sliding door 315 opens, and the rotary storage unit 4 rotates until the corresponding trash can coincides with the lowering port to complete the classification.

[0211] Specifically, after the recognition and classification are completed, the display screen automatically displays the garbage classification information, including: garbage type, garbage quantity, whether the garbage has been put in place, etc.; an infrared distance sensor 105 is installed on the middle support plate 103 above the trash can main body 404 to detect the loading state of the trash can. When the garbage capacity exceeds 75%, the display screen 501 actively displays that the trash can is full, facilitating users to empty the trash in time.

[0212] To facilitate the understanding of the above technical solution of the present invention, the following is a specific description in conjunction with embodiments:

[0213] The garbage classification device proposed by the present invention works in coordination through the multi-stage conveyor unit 2, the compression unit 3, the rotary storage unit 4, and the recognition and display unit 5. The linkage control between the units is realized by a control unit 6 composed of a development board (such as the NVIDIA Jetson Nano development board) and a self-designed main control PCB board. The control unit 6 controls the working states of the conveyor belt, the electric push rod 301, the rotary chassis 403, and the display screen 501 according to the recognized types of garbage, realizing the function of classifying multiple pieces of garbage simultaneously and improving the reliability and accuracy of classification.

[0214] The multi-stage conveyor unit 2 includes a V-shaped conveyor belt assembly and a horizontal conveyor belt assembly. The V-shaped conveyor belt assembly includes a left conveyor belt module and a right conveyor belt module, which have exactly the same structure and separate garbage through differential rotation to form a pair of transmission modules; the V-shaped conveyor belt assembly is composed of two conveyor belts with differential operation and is differentially driven by two first stepping motors 201. The included angle between the conveyor belts is 110° - 130°. The angle adjustment is realized by an integrated stepping motor through a worm and gear mechanism. The composite profile 202 as the support frame adopts a carbon fiber composite aluminum profile, with a polyurethane anti-slip layer on the surface, and an infrared density sensor (model E3Z-T61 is used in this embodiment) is integrated to detect the stacking density of garbage in real time (the detection range is set to 0 - 200 cm in this embodiment 2 ), when the density ≥ 3 pieces / 100 cm², the included angle is automatically adjusted to 110°, and when the density ≤ 1 piece / 100 cm 2 it is adjusted to 130°.

[0215] The compression unit 3 includes an electric push rod assembly and a sliding door assembly. Among them, the electric push rod 301 adopts a 24V / 1000N specification, providing the force required for compression; the sliding door 315 is powered by the third stepping motor 310, converting the rotational motion of the third stepping motor 310 into a linear motion of the sliding door, providing a compression platform for recyclable waste; the overall frame of the device provides a fixed space for the electric push rod and the sliding door, improving the stability during compression.

[0216] The rotating storage unit 4 includes a rotating chassis 403 and a trash can main body 404, providing storage space for the sorted waste; the rotating chassis 403 is powered by the fourth stepping motor 401, and the power is transmitted through gear meshing to drive the rotation of the rotating chassis 403. There are balls at the bottom of the rotating chassis 403 to reduce friction and provide support. The trash can main body 404 adopts a cylindrical design to facilitate providing a larger capacity.

[0217] The identification and display unit 5 consists of a camera 502 and a display screen 501, providing the reliability of identification and interactivity; the camera 502 adopts a 4K / USB wide-angle camera module to increase the detection range; the display screen 501 adopts a resolution of 1280*800, providing a high-quality interactive experience for users.

[0218] The V-shaped conveyor belt assembly and the horizontal conveyor belt assembly provide reliable conveying functions; two first stepping motors 201 achieve closed-loop differential speed control through a control board, and the dynamic adjustment range of the differential speed ratio is 1:1.2 to 1:1.8. The corresponding conveyor belt linear speeds are 0.25m / s and 0.3 - 1.25m / s respectively. The control logic is based on the volume difference of the waste identified by the camera (in this embodiment, when the difference is set to >30%, the differential speed ratio is increased to 1:1.8) and the material friction coefficient (such as the speed reduction to 1:1.3 for glass materials), combined with the real-time feedback of the separation force by a pressure sensor (in this embodiment, the measuring range is set to 0 - 50N, and the accuracy is ±0.5%). When the separation force <5N, the bottom vibration motor 205 is triggered to assist in separation. The composite profile 202 as the conveyor belt support frame adopts a European standard 2060 carbon fiber composite aluminum profile (in this embodiment, the composite profile model is CF-AL6061, the cross-sectional size is 20mm×60mm, and the bending strength ≥200MPa). The conveyor belt on the surface is designed with fishbone-shaped anti-slip patterns (in this embodiment, the depth of the fishbone-shaped anti-slip patterns is 1.2mm, the spacing is 8mm, and the inclination angle is 45°). The vibration motor 205 is integrated at the bottom (in this embodiment, the installation distance between the vibration motor and the V-shaped conveyor belt assembly is 150mm). The conveyor belt is fixed to the European standard 2020 aluminum profile of the frame main body 1 (in this embodiment, the wall thickness of the frame main body is 1.5mm) through M8 high-strength bolts (in this embodiment, the bolt strength is 8.8 grade).

[0219] The sliding door 315 is driven by a third stepping motor 310 to drive the gear and rack to move, realizing the opening and closing function of the lowering port; the electric push rod assembly consists of an electric push rod 301 and a compression plate, and the compression plate consists of an active compression plate 303 and a passive compression plate 308; the active compression plate 303 is connected to the electric push rod 301 through bolts and is fixed by connecting with two first optical axes 306 on the left and right through a linear motion bearing 305 and an optical axis support seat 307, so as to have good stability; the passive compression plate 308 is fixed on the outer frame through a fiberglass board, facilitating the adjustment of the compression space and compression capacity of the electric push rod 301.

[0220] The rotating chassis 403 is powered by a fourth stepping motor 401 and is fixed and supported by bolts and ball bearings, and saves the bottom space; the chassis of the rotating storage unit 4 consists of the rotating chassis 403 and a driven gear 406, and the two are radially positioned through a bearing and the second optical axis 407 and axially positioned through ball bearings, and the two cooperate to provide reliable positioning; the fourth stepping motor 401 drives the driving gear 402 to rotate, and the gear meshing drives the driven gear 406 to rotate. The driven gear 406 is connected to the rotating chassis 403 through bolts, and then drives the trash can main body 404 to rotate, so as to realize the classification of different trash.

[0221] This device uses a self-developed control board to complete the power supply and control work of two first stepping motors 201, one second stepping motor 209, one third stepping motor 310, one fourth stepping motor 401, a display screen 501 and a development board; the camera 502 and the display screen 501 are connected to the development board, and the development board and the control board cooperate with each other to complete the identification and classification work of trash types; the power supply is completed by a Graupner 6S lithium battery for RC models, and an optoelectronic emergency stop sensor 7 is provided as a one-key start-stop switch to meet the requirements of daily use.

[0222] Finally, the trash classification device of the present invention is encapsulated by European standard 2020 aluminum profiles, and a power switch is left on the side wall of the device, ensuring the versatility and adaptability of the device in various application environments.

[0223] To sum up, with the above technical solutions of the present invention, through the deep integration of hardware collaboration and intelligent algorithms, the industry pain points of low recognition accuracy and limited classification efficiency of traditional trash classification equipment in complex scenarios are solved. Compared with the prior art, this device has the following innovative advantages:

[0224] 1) Dynamic separation-identification collaborative architecture:

[0225] Adopt the cascaded structure of a V-shaped differential conveyor belt and a horizontal conveyor belt, and pioneer the "separation - pause - recognition" working mechanism. Through the differential control of two first stepping motors 201, the physical dispersion of garbage is realized, and high-definition image acquisition is completed by the camera in a static state, significantly improving the quality of feature extraction. After testing, this design increases the recognition accuracy of small target garbage by 37%, effectively overcoming the problem of blurred images caused by traditional continuous conveying.

[0226] 2) Intelligent compression - storage linkage system:

[0227] Innovatively, the electric push rod compression module is dynamically coupled with the rotating storage unit 4, and millimeter-level timing control is achieved through the control board. When recyclables are recognized, the system completes the linkage operations of sliding door closing, compression execution (pressure reaching 200N), and storage bin positioning within 0.5 seconds. The compression efficiency is increased by 60% compared with the traditional pneumatic solution, and the energy consumption is reduced by 45%.

[0228] 3) Multi-dimensional state perception network:

[0229] Construct a multi-modal monitoring system composed of an infrared ranging sensor, motor torque detection, and visual feedback to real-time perceive the trash can capacity (accuracy ±2mm), equipment operation status, and abnormal vibration. An LSTM prediction model is established through the development board, which can give early warning of equipment failures 10 minutes in advance, and the operation and maintenance response time is shortened by 80%.

[0230] 4) Adaptive learning framework:

[0231] The device is built-in with an incremental learning algorithm, which uses tens of thousands of garbage image data processed daily to continuously optimize the YOLOv8 model through a contrastive learning strategy. The actual deployment data shows that the weekly model iteration of the system can increase the recognition accuracy of new categories of garbage by 15%, effectively solving the problem of performance degradation of traditional fixed models due to changes in the shape of garbage.

[0232] This device has been verified through laboratory simulation and field tests, achieving a classification accuracy of 95.3% in a test set containing more than 200 garbage categories, and the processing speed reaches 120 pieces per minute. The compact modular design (occupying an area of <1.5m 2 ) combined with the universal ball mechanism is especially suitable for space-limited scenarios such as schools and communities, providing an efficient and reliable waste classification solution for the construction of smart cities.

[0233] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection scope of the present invention.

Claims

1. A multi - garbage classification device for self - supervised hybrid domain adaptation, characterized in that, It includes a frame body, a multi-stage conveying unit, a compression unit, a rotating storage unit and an identification and display unit. The multi-stage conveying unit, the compression unit and the rotating storage unit are sequentially arranged inside the frame body from top to bottom. And the multi-stage conveying unit is located on one side of the central axis of the frame body, and the compression unit is located on the other side of the central axis of the frame body; A feeding port is arranged above the multi-stage conveying unit at the top end of the frame body. Ball bearings are fixedly arranged at the four corners of the bottom end of the frame body. A middle support plate is arranged in the middle of the frame body. A lower bottom plate is arranged at the bottom of the frame body. A control unit is arranged on one side of the frame body; When the multi-class garbage classification device for self-supervised hybrid domain adaptation realizes multi-class garbage classification and identification, it includes: S1. Obtain multi-view images of garbage through a camera and perform preprocessing; S2. Use the hybrid domain adaptation garbage recognition model to extract image features, realize spatial selective scanning and channel exchange, and update the model parameters using the transfer learning and incremental learning framework; S3. Based on the hybrid domain adaptation garbage recognition model, output the garbage classification result and confidence level, and synchronously update them to the display screen and the control unit; Among them, the S2 includes: S21. Construct a hybrid domain adaptation garbage recognition model, and improve the feature extraction ability by embedding a channel attention module and a contrast learning loss function; S22. Process the feature map based on the spatial selective scanning mechanism and the channel exchange strategy, and realize spatial selective scanning and channel exchange by calculating the cross-domain feature entropy; S23. Combine the pre-trained weights of transfer learning and the incremental learning framework to complete the dynamic update of the model parameters and improve the recognition performance of garbage categories.

2. The multi - garbage classification device for self - supervised hybrid domain adaptation according to claim 1, wherein, The multi-stage conveying unit includes a V-shaped conveyor belt assembly arranged on one side of the top of the frame body; The V-shaped conveyor belt assembly includes two groups of symmetrically arranged composite profiles, and the two groups of composite profiles are in a V-shaped obtuse angle structure. A number of linearly distributed V-shaped conveyor belt connectors are arranged at the bottom ends of the two groups of composite profiles. A vibration motor is commonly connected to one side of the bottoms of the two groups of composite profiles; V-shaped conveyor belt rollers are arranged at both ends of the composite profile, and a first stepping motor is arranged at one end of the two V-shaped conveyor belt rollers located on the side of the vibration motor; 3. The multi - garbage classification device for self - supervised hybrid domain adaptation according to claim 2, wherein, The multi-stage conveying unit further includes a horizontal conveyor belt assembly arranged on the side of the middle of the frame body away from the V-shaped conveyor belt assembly; The horizontal conveyor belt assembly includes two groups of symmetrically arranged horizontal conveyor belt rollers. A horizontal conveyor belt is sleeved outside the horizontal conveyor belt rollers. Bearing seats are arranged at both ends of the horizontal conveyor belt rollers. A connecting profile is arranged between the two groups of bearing seats; A second stepping motor is arranged at the bottom end of the connecting profile. A synchronous pulley is sleeved on the output end of the second stepping motor. The synchronous pulley is connected to the horizontal conveyor belt roller on the side away from the compression unit through a synchronous belt; 4. The multi - garbage classification device for self - supervised hybrid domain adaptation according to claim 1, characterized in that, The compression unit includes an electric push rod assembly arranged on one side of the middle of the frame body; The electric push rod assembly includes a rear support plate arranged on one side of the middle part of the frame body. An electric push rod is arranged on one side of the rear support plate. The output end of the electric push rod is provided with an active compression plate. First optical axes are respectively arranged through both ends of the active compression plate. One end of each first optical axis is provided with an optical axis support frame, and a linear motion bearing fixedly connected with the active compression plate is arranged in the middle of each first optical axis; The other ends of the two groups of first optical axes are jointly provided with a front support plate. Optical axis support seats matched with the first optical axes are arranged on both sides of the front support plate, and a passive compression plate is arranged in the middle of the front support plate.

5. The multi - garbage classification device for self - supervised hybrid domain adaptation according to claim 4, characterized in that, The compression unit further includes a sliding door assembly arranged below the electric push rod assembly in the middle of the frame body; The sliding door assembly includes two groups of sliding door tracks symmetrically arranged at the top of the middle support plate. A sliding groove is formed on one side of each sliding door track, and a sliding door is arranged inside the sliding groove; A sliding door upper rack is arranged on one side of the top of the sliding door. A third stepping motor is arranged on one side of the sliding door track far away from the multi-stage conveying unit. A sliding door gear matched with the sliding door upper rack is arranged at the output end of the third stepping motor.

6. The multi - garbage classification device for self - supervised hybrid domain adaptation according to claim 1, characterized in that, The rotary storage unit includes a rotary chassis arranged at the bottom of the frame body. A plurality of garbage can bases distributed in a circular pattern are formed at the top of the rotary chassis. A garbage can body is arranged on the top of each garbage can base. A driven gear is arranged at the bottom end of the rotary chassis; A fourth stepping motor is arranged on one side of the rotary chassis below the compression unit. A driving gear meshed with the driven gear is arranged at the bottom of the output end of the fourth stepping motor; A light axis fastener is arranged through the middle of the rotary chassis. A second optical axis is arranged through the middle of the light axis fastener. A middle fixing piece matched with the light axis fastener is sleeved on the outer side of the middle of the second optical axis. Tapered roller bearings are sleeved on both the top end and the bottom end of the second optical axis, and fixing pieces are arranged on the outer sides of the tapered roller bearings.

7. A multi - garbage classification control method based on self - supervised hybrid domain adaptation, which uses the multi - garbage classification device based on self - supervised hybrid domain adaptation described in any one of claims 1 - 6 to achieve multi - garbage classification control, characterized in that, including: SI. Differentially separate the input garbage through the V-belt conveyor assembly, and dynamically adjust the differential ratio through the control unit; SII. Pause the horizontal conveyor belt assembly, collect static images by using a camera and transmit them to the control unit; SIII. The control unit completes garbage classification recognition, and outputs the result to the control board when the confidence level reaches the threshold; SIV. Select an operation mode according to the classification result. When it is identified as recyclable garbage, execute the recyclable garbage mode, and sequentially execute the operations of closing the sliding door, compressing the electric push rod, rotating the rotary chassis, and opening the sliding door; When it is identified as non-recyclable garbage, execute the non-recyclable garbage mode, and sequentially execute the operations of rotating the rotary chassis and opening the sliding door; SV. Update the garbage classification information and the capacity status through the display screen, and trigger a shutdown warning when the capacity of the garbage can body exceeds the threshold.

8. A multi - garbage classification recognition method based on self - supervised hybrid domain adaptation, which uses the multi - garbage classification device based on self - supervised hybrid domain adaptation described in any one of claims 1 - 6 to realize multi - garbage classification recognition, characterized in that, The construction of the hybrid domain adaptation garbage recognition model to improve the feature extraction ability by embedding a channel attention module and a contrast learning loss function includes: S211. Extract the shallow features of the source domain and target domain images through two consecutive convolutional blocks, and generate preliminary feature maps through max-pooling downsampling; S212. Process the preliminary feature map using depthwise separable convolution, and generate a similarity matrix by measuring the spatial consistency between the source domain feature and the target domain feature map through cosine similarity; S213. Divide the source domain feature map and the target domain feature map equally along the channel dimension, calculate the channel-level similarity after swapping some channels, and generate features with spatial and channel alignment; The expression of the channel-level similarity is: where S channel represents the channel-level similarity; represents the feature vector after concatenating the first and second segments of the source-domain feature map and the third and fourth segments of the target-domain feature map; represents the feature vector after concatenating the first and second segments of the target-domain feature map and the third and fourth segments of the source-domain feature map; ||·|| represents the norm of the vector.

9. A multi - garbage classification recognition method based on self - supervised hybrid domain adaptation according to claim 8, characterized in that, Processing the feature map based on the spatial selective scanning mechanism and the channel swapping strategy, and realizing spatial selective scanning and channel swapping by calculating the cross-domain feature entropy includes: S221. Independently enhance the intra-domain features using self-attention, generate source-dominated features and target-dominated features through cross-attention, and achieve soft alignment of features; S222. Transmit the source domain features, source-dominated features, target-dominated features and target domain features to the next stage, and repeat the feature extraction and alignment process; S223. Introduce entropy loss in the shallow, middle and deep layers, calculate the cross-domain feature entropy through the edge activation function, and strengthen the cross-domain consistency of multi-level features; The expression of the cross-domain feature entropy is: H t→ts (l) = -k∑σ m (Z t→s (l,k) ) log(σ m (Z t (1,k) )); Where, H t→ts (l) represents the entropy value of the source-dominated feature in the l-th stage; k represents the proportionality coefficient; σ m represents the edge activation function; Z t→ts (l,k) represents the source-dominated feature in the l-th stage; Z t (1,k) represents the target-domain feature in the first stage; L entropy represents the total entropy loss; N represents the number of feature layers; l represents the feature layer index; H t→st (l) represents the entropy value of the target-dominated feature in the l-th stage.

Images