Sma flexible fabric and gait guiding system and method based on edge ai

Abstract

The present application relates to the technical field of wearable motion auxiliary equipment, and discloses an SMA flexible fabric and gait guiding system and method based on edge AI, which comprises a power safety protection unit, a waist main control unit, a gait information acquisition unit, an environment monitoring unit, a user physiological detection unit and an SMA flexible fabric execution unit. The system uses double IMU and plantar pressure data to identify abnormal gaits such as internal eight and external eight in real time through edge AI one-dimensional convolutional neural network. The SMA flexible fabric is spirally wound on the lower limbs, and the main control unit drives the contraction of a specific traction path based on the identification result, generates a lamellar lateral force and a horizontal plane rotation force, and provides two-plane flexible auxiliary guidance. At the same time, the system fuses visual terrain, environmental humidity and physiological state for multi-source safety gating, and cooperates with a four-stage thermal control mechanism to realize the rapid release of traction force. The present application can effectively correct abnormal gait without rigid joints, and has strong wearing adaptability and safety and reliability.

Description

Technical Field

[0001] This invention relates to the field of wearable motion assistive device technology, specifically to an edge AI-based SMA flexible fabric and gait guidance system and method. Background Technology

[0002] With the development of wearable smart devices, lower limb exoskeletons, and mountaineering aids, gait monitoring and guidance have expanded from simple medical assessments to become key issues in human movement health and wearable robot control. Conditions such as in-toeing, out-toeing, pronation / eversion, and abnormal toe angles can affect plantar load distribution and lower limb stability. Existing lower limb exoskeletons primarily focus on sagittal plane assistance, which can easily lead to a mismatch between the assistance direction and the user's actual gait trajectory during intervention, and may even reinforce the user's existing incorrect gait.

[0003] Existing lower limb assistive devices primarily assist flexion and extension movements in the sagittal plane around the hip, knee, and ankle joints, mainly addressing propulsion in the anterior-posterior direction. These devices generally employ rigid linkages, cables, or pneumatic structures, requiring additional joints, guide wheels, or multiple sets of fixation components, significantly limiting their adaptability. A true human gait simultaneously includes lateral stability in the frontal plane, foot inversion / pronation control, and coordination of toe angle and tibial rotation in the horizontal plane. Existing devices generally lack real-time recognition and guidance mechanisms for frontal and horizontal plane postures, making it difficult to effectively identify abnormal gaits such as in-toeing and out-toeing, and unable to release or dynamically adjust assistance in a timely manner in complex terrain and various movement states.

[0004] Furthermore, while shape memory alloys offer advantages such as compact structure, low noise, and high power density, making them suitable for traction feedback in flexible fabrics, they also possess inherent characteristics such as significant thermal inertia, small displacement, susceptibility to overheating, and large transient currents from the power supply. In practical applications, relying solely on a simple continuous power-on drive method can easily lead to safety hazards such as system temperature runaway, material fatigue, wearability discomfort, and battery voltage drops. Therefore, current technologies lack a closed-loop system for wearable low-power scenarios, making it difficult to provide perceptible and timely gait guidance while ensuring real-time performance and safety. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a SMA flexible fabric and gait guidance system and method based on edge AI. The technical problem solved by this invention is that existing lower limb movement assistive devices mostly provide forward and backward propulsion assistance around the sagittal plane, typically relying on rigid joints or fixed support structures, resulting in poor wearability. For abnormal gaits such as in-toeing and out-toeing, it is difficult to provide lateral stability and rotational guidance corresponding to the direction of gait deviation in the frontal and horizontal planes. Furthermore, shape memory alloy actuators in wearable applications suffer from high thermal inertia and slow cooling and reset, making it difficult to coordinate their actuation timing and traction release process after actuation with the actual gait stage, terrain changes, and safety conditions in real time.

[0006] To address the above problems, the present invention provides the following technical solution:

[0007] The first aspect of the present invention provides an SMA flexible fabric and gait guidance system based on edge AI, including a power safety protection unit, a waist main control unit, a gait information acquisition unit, an environmental monitoring unit, a user physiological detection unit, and an SMA flexible fabric execution unit;

[0008] The gait information acquisition unit is used to collect plantar pressure and inertial information;

[0009] The environmental monitoring unit is used to acquire images of the terrain ahead and environmental humidity information;

[0010] The user physiological detection unit is used to collect the user's heart rate status;

[0011] The power safety protection unit is used to monitor overcurrent, overtemperature and undervoltage conditions during the SMA driving process;

[0012] The waist-mounted main control unit is used to centrally perform data reception, feature processing, edge AI inference, pulse width modulation control, and security status management.

[0013] The SMA flexible fabric actuator has a spiral path, which is a clockwise or counterclockwise spiral path based on the longitudinal axis of the lower limb when viewed from the proximal end to the distal end in a standing posture. The SMA flexible fabric actuator is wrapped around the lower limb of the wearer and provides flexible assistance in the frontal and horizontal planes of the lower limb according to the output of the waist main control unit.

[0014] Furthermore, the SMA flexible fabric actuation unit includes a flexible fabric substrate and at least one set of shape memory alloy composite yarns disposed on the flexible fabric substrate. The shape memory alloy composite yarns extend along the flexible fabric substrate to define the traction direction and form a continuous leg-wrapping path.

[0015] Furthermore, the flexible fabric base is a long bandage-shaped knitted fabric; the shape memory alloy composite yarn is fixed to the flexible fabric base by embroidery, sewing, or directional fiber laying and is located on the side away from the skin; a heat-insulating fabric layer is provided between the shape memory alloy composite yarn and the skin.

[0016] Furthermore, the waist main control unit includes a waist main control board fixed fabric, and the gait information acquisition unit includes an ankle acquisition board fixed fabric; the two ends of the SMA flexible fabric execution unit are provided with Velcro connecting parts, and the Velcro connecting parts are respectively connected to the waist main control board fixed fabric and the ankle acquisition board fixed fabric, so that the SMA flexible fabric execution unit is spirally wound from the waist to the ankle along the lower limb curvature.

[0017] Furthermore, the SMA flexible fabric actuator includes a traction path for correcting inward-pointing SMA flexible fabric and a traction path for correcting outward-pointing SMA flexible fabric, which can be selected by the wearer. The traction path for correcting inward-pointing SMA flexible fabric and the traction path for correcting outward-pointing SMA flexible fabric are respectively counterclockwise spiral paths and clockwise spiral paths based on the wearer's standing posture and the longitudinal axis of the lower limb viewed from the proximal end to the distal end. According to the correction type predetermined by the user, the two are selectively configured into one of the traction paths for correcting inward-pointing SMA flexible fabric or the traction path for correcting outward-pointing SMA flexible fabric. Both traction paths extend from the waist main control plate fixed fabric to the ankle acquisition plate fixed fabric and form displacement in the length direction and circumferential direction of the lower limb.

[0018] Furthermore, when the traction path of the inward-to-the-eight SMA flexible fabric contracts, it forms a horizontal rotational force and a frontal lateral force to reduce the tendency of toe inversion and tibial internal rotation; when the traction path of the outward-to-the-eight SMA flexible fabric contracts, it forms a horizontal rotational force and a frontal lateral force in opposite directions.

[0019] Furthermore, the gait information acquisition unit includes four plantar piezoresistive pressure sensors, two six-axis motion processing sensors for acquiring the inertial information and forming a dual IMU, a protective shell for the ankle acquisition plate, a fixing fabric for the ankle acquisition plate, and a secondary board MCU equipped with Bluetooth; the four plantar piezoresistive pressure sensors are respectively located on the medial side of the heel, the lateral side of the heel, the medial side of the forefoot, and the lateral side of the forefoot, and the two six-axis motion processing sensors are respectively located near the tibia of the lower leg and in the arch area.

[0020] Furthermore, the waist main control unit includes a protective shell for the waist main control board, a fabric fixing the waist main control board, and a motherboard MCU equipped with Bluetooth; the environmental monitoring unit includes a visual look-ahead unit and an environmental humidity detection unit. The visual look-ahead unit uses a global shutter grayscale camera, and the output of the global shutter grayscale camera is transferred to a static image buffer through a digital camera interface and a direct memory access module. The visual look-ahead unit is used to identify staircase scenes, and the environmental humidity detection unit is used to acquire the environmental humidity information.

[0021] Furthermore, the waist control unit uses the sudden increase in plantar pressure or the peak value of inertial impact in the inertial information as the starting point of the gait cycle, the toe-off event as the end of the assist phase or the boundary of a short time window, and the cycle end event corresponding to the next sudden increase in plantar pressure or peak value of inertial impact as the end point of the complete gait cycle window; the waist control unit forms a multi-channel temporal feature matrix by combining the plantar pressure collected by the gait information acquisition unit and the triaxial acceleration and triaxial angular velocity output by the two six-axis motion processing sensors; the waist control unit combines the stair scene and environmental humidity information obtained by the environmental monitoring unit, the user heart rate status collected by the user physiological detection unit, and the detection results of the power safety protection unit, and obtains the gait recognition result and performs safety state management through quantized one-dimensional convolutional neural network classification and rule-gated state machine judgment.

[0022] A second aspect of the present invention provides an edge AI-based SMA flexible fabric and gait guidance method applied to the aforementioned system, wherein the SMA flexible fabric execution unit in the system includes one of selectively correcting an inward-pointing SMA flexible fabric traction path or correcting an outward-pointing SMA flexible fabric traction path, and the method includes:

[0023] Complete the loading of user parameters, the zero-point calibration of plantar pressure, and the zero-bias calibration of the dual IMUs;

[0024] The system collects the plantar pressure, the timing output of the dual IMUs, the stair scene identified by the visual forward-looking unit, and the ambient humidity information, and receives the user's heart rate status collected by the user's physiological detection unit.

[0025] Dynamic gait window slicing is performed based on the sudden increase in plantar pressure or the peak value of inertial impact in the inertial information, and the multi-channel temporal feature matrix is ​​generated.

[0026] The gait recognition result is obtained by quantizing a one-dimensional convolutional neural network for classification and a rule-gated state machine for judgment, and then a safe state management is performed.

[0027] Based on the gait recognition results, in the corresponding gait stage of the subsequent gait cycle, the corrected inward-pointing SMA flexible fabric traction path or the corrected outward-pointing SMA flexible fabric traction path in the SMA flexible fabric execution unit is selected, the guiding intensity and thermal control state are determined, and the pulse width modulation output is controlled to switch between dormancy, burst preheating, constant temperature maintenance and rapid release states according to the thermal control state.

[0028] In the rapid release state, the pulse width modulation output is stopped, which cools the shape memory alloy composite wire and reduces the traction force.

[0029] This invention provides a SMA flexible fabric and gait guidance system and method based on edge AI, which has the following beneficial effects:

[0030] 1. This invention, by setting a traction path of SMA flexible fabric spirally wound in a clockwise or counterclockwise direction, converts the linear traction force generated by the contraction of shape memory alloy composite yarns through electrical current into a horizontal rotational component around the longitudinal axis of the lower limb and a frontal lateral component acting along the inner and outer sides of the lower limb. Compared to traditional auxiliary devices that mainly rely on rigid joints or fixed structures to provide only sagittal propulsion force, this invention can provide directional matching dual-plane flexible guidance for abnormal gaits such as in-toeing and out-toeing without adding rigid multi-degree-of-freedom mechanisms, thus improving the wearability and movement coordination of the device.

[0031] 2. This invention integrates temporal features from plantar pressure and dual IMUs of the foot and lower leg, and relies on a quantized one-dimensional convolutional neural network and a rule-gated state machine deployed in the waist control unit for edge AI inference. This enables precise slicing of continuous dynamic gait cycles and low-latency recognition of gait categories such as in-toeing and out-toeing. Compared to traditional systems lacking real-time monitoring of frontal and horizontal plane posture, this invention can automatically learn temporal differences based on sensor input, determine whether the current gait matches the user-preset correction type within millisecond delays, and determine the corresponding driving timing and intensity, ensuring that the flexible assistance closely matches the actual movement phase of the human body.

[0032] 3. This invention establishes a four-stage thermal control mechanism—dormancy, burst preheating, constant temperature maintenance, and rapid release—and combines it with multi-source safety gating conditions such as visual staircase look-ahead, ambient humidity, and heart rate status to achieve dynamic adjustment and rapid interruption of the SMA pulse width modulation output. Compared to conventional simple continuous-energized shape memory alloy drive methods, this invention overcomes the inherent defects of SMA actuators, such as high thermal inertia and slow cooling reset. It ensures that the output can be quickly cut off and the traction force can be reduced during the cooling process when the gait phase ends or when facing complex terrain and safety risks, thus guaranteeing the safety of the equipment in various scenarios. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the overall structure of the SMA flexible fabric gait recognition and guidance system based on edge AI described in this invention;

[0034] Figure 2 This is a structural diagram of plantar pressure and dual IMU arrangement in this invention;

[0035] Figure 3 This is a schematic diagram of the SMA flexible fabric in this invention;

[0036] Figure 4 This is a schematic diagram of the leg-wrapping traction direction of the SMA flexible fabric actuator in this invention;

[0037] Figure 5 This is a schematic diagram of the staircase recognition process of the visual forward-looking unit in this invention;

[0038] Figure 6 This is a schematic diagram of the one-dimensional convolutional recognition process for edge AI decision-making in this invention;

[0039] Figure 7 This is a general flowchart of the method described in this invention.

[0040] The components include: 1. Power safety protection unit; 2. Waist main control unit; 23. Waist main control board protective shell; 24. Waist main control board fixing fabric; 25. Mainboard MCU with Bluetooth; 3. Gait information acquisition unit; 31. Four-channel plantar piezoresistive pressure sensor; 32. Two six-axis motion processing sensors; 33. Ankle acquisition board protective shell; 34. Ankle acquisition board fixing fabric; 35. Sub-board MCU with Bluetooth; 4. Environmental monitoring unit; 41. Visual forward-looking unit; 42. Environmental humidity detection unit; 5. User physiological detection unit; 6. SMA flexible fabric actuator; 61. Flexible fabric substrate; 62. Shape memory alloy composite yarn; 63. Velcro connector; 64. Heat insulation fabric layer; 66. Correction of inward-pointing SMA flexible fabric traction path; 68. Correction of outward-pointing SMA flexible fabric traction path. Detailed Implementation

[0041] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] See attached document Figure 1This embodiment provides an edge AI-based SMA flexible fabric and gait guidance system, which includes a power safety protection unit 1, a waist main control unit 2, a gait information acquisition unit 3, an environmental monitoring unit 4, a user physiological detection unit 5, and an SMA flexible fabric execution unit 6.

[0043] Specifically, the waist control unit 2 is located in the waist area of ​​the wearer and is used to centrally perform data reception, feature processing, edge AI inference, pulse width modulation control and safety status management; the gait information acquisition unit 3 is located in the ankle and foot area of ​​the wearer; the SMA flexible fabric execution unit 6 extends from the waist to the ankle and wraps around the lower limb.

[0044] The waist control unit 2 includes a waist control board protective shell 23, a waist control board fixing fabric 24, and a Bluetooth-enabled mainboard MCU 25. The waist control board fixing fabric 24 wraps around or partially wraps around the wearer's waist in a belt shape and is secured by Velcro, buckles, or elastic straps; the waist control board protective shell 23 is located in the belt area, and the Bluetooth-enabled mainboard MCU 25 is located in the wearer's waist area. The waist control board fixing fabric 24 also forms the proximal anchoring base for the SMA flexible fabric actuator 6. The visual forward-looking unit 41 in the environmental monitoring unit 4 is positioned facing forward of the wearer, and the environmental humidity detection unit 42 is located in the waist control area to acquire images of the terrain ahead and environmental humidity information, respectively.

[0045] The gait information acquisition unit 3 includes an ankle acquisition plate protective shell 33, an ankle acquisition plate fixing fabric 34, and a Bluetooth-enabled secondary board MCU 35. The ankle acquisition plate fixing fabric 34 is fixed to the wearer's ankle, and the ankle acquisition plate protective shell 33 and the Bluetooth-enabled secondary board MCU 35 are located on the inner side of the ankle to reduce the impact on walking sway. The ankle acquisition plate fixing fabric 34 constitutes the distal anchoring position of the SMA flexible fabric actuator 6 and cooperates with the waist main control board fixing fabric 24 to allow the traction force generated by the contraction of the SMA to act on the wearer's lower limbs through the flexible fabric.

[0046] See attached document Figure 2 The four-channel plantar piezoresistive pressure sensors 31 in the gait information acquisition unit 3 are respectively located on the inner heel, outer heel, inner forefoot, and outer forefoot. For the underlying signal processing where the pressure sensors convert resistance changes into voltage values ​​via a voltage divider circuit and are acquired by an analog-to-digital converter, those skilled in the art can employ conventional voltage divider conversion and analog-to-digital acquisition methods. The specific electrical signal conversion process is well-known in the field and will not be elaborated here. After the system is powered on, a zero-point calibration value is obtained under no-load conditions on the sole of the foot, and a piecewise linear interpolation table is obtained through weight calibration.

[0047] The lumbar control unit 2 calculates the pressure ratio between the medial and lateral sides of the forefoot, the pressure ratio between the medial and lateral sides of the heel, the trajectory of the pressure center, and the direction of the plantar line based on the aforementioned four types of plantar pressure. When the pressure center is continuously biased towards the medial or lateral side of the plantar surface, it indicates a shift in the load distribution on the frontal plane of the lower limb. The direction of the plantar line is used to help characterize the direction of force on the foot and the trend of plantar load changes within the gait cycle. Sudden pressure increase, forward shift of the pressure center, and decrease in plantar pressure are used to help determine gait events such as heel strike, mid-stance, and toe lift-off, respectively.

[0048] The first six-axis motion processing sensor is located near the tibia of the lower leg, and the second six-axis motion processing sensor is located in the arch region of the foot. The two six-axis motion processing sensors 32 constitute a dual IMU, and acquire triaxial acceleration and triaxial angular velocity at a frequency of not less than 100Hz. In a preferred implementation, the lumbar main control unit 2 does not rely on long-term integration to obtain the absolute foot heading angle or the absolute tibia rotation angle relative to the foot. Instead, it directly uses the raw time-series data of the dual IMU after zero-bias calibration and low-pass filtering as the input features of the end-side AI. The quantized one-dimensional convolutional neural network automatically learns the temporal differences in toe inversion, toe outversion, impact phase, and support phase.

[0049] For implementations requiring interpretable attitude parameters, the waist control unit 2 can calculate the angular velocity increment within a short time window of one gait cycle or several consecutive sampling points, and combine it with the acceleration and gravity direction to form the foot heading angle increment, the tibia rotation angle increment relative to the foot, or corresponding auxiliary attitude features. These auxiliary features are preferably used for rule gating or log interpretation, but can also be used as optional derived features to form a multi-channel temporal feature matrix with the basic input. They are not the sole necessary condition for the system to complete in-toeing or out-toeing recognition. This avoids the cumulative drift caused by long-term gyroscope integration while retaining the ability to perform gait classification using the original dynamic features from the dual IMUs.

[0050] The gait information acquisition unit 3 acquires the outputs of four plantar pressure sensors and two six-axis motion processing sensors 32 via a Bluetooth-enabled sub-board MCU 35, and receives the heart rate status acquired by the user's physiological detection unit 5, before sending it to the Bluetooth-enabled mainboard MCU 25. For data transmission between the Bluetooth-enabled sub-board MCU 35 and the Bluetooth-enabled mainboard MCU 25 based on Bluetooth Low Energy, those skilled in the art can use conventional Bluetooth Low Energy communication methods; the specific wireless communication process is well-known in the field and will not be described further here. The data packets include timestamps and sampling sequence numbers, and the waist-mounted main control unit 2 aligns the multi-source data according to the timestamps.

[0051] The waist control unit 2 uses the sudden increase in plantar pressure or the peak of inertial impact as the starting point of the gait cycle, and the toe-off event or the cycle end event corresponding to the next sudden increase in plantar pressure or the peak of inertial impact as the window end to obtain the true gait cycle window. In the complete gait cycle slice used to quantize the input of the one-dimensional convolutional neural network, the next sudden increase in plantar pressure or the peak of inertial impact is preferred as the cycle end; the toe-off event is used to help determine the end of the assist phase or the boundary of the short time window. For gait cycles with different durations, linear interpolation resampling is used to obtain a fixed-length sequence, so that data under different speed conditions can be input into the same edge AI model.

[0052] In an exemplary embodiment, the gait cycle slice uses the heel strike impact as the cycle starting point. Pressure triggering occurs when the sum of the four plantar pressure signals is greater than 700mV, the pressure increment between adjacent frames is greater than 120mV, and the sum of the four plantar pressure signals in the previous low-pressure phase is less than 180mV; or inertial impact triggering occurs when the absolute value of the Z-axis acceleration of either of the two six-axis motion processing sensors 32 is greater than 12000 LSB. An HS event is generated if either of these triggering conditions is met. To avoid repeated identification of the same landing, a 300ms refractory period is set for the HS event. The original number of frames in an effective gait cycle is limited to 24 to 120 frames, followed by linear interpolation into a 60-frame timing matrix.

[0053] See attached document Figure 6 The dynamic gait recognition branch uses a one-dimensional convolutional neural network. Its input is a multi-channel time-series matrix of size (B, 16, 60), where B is the batch size, the 16 channels include four plantar pressure resistance signals and three-axis acceleration and three-axis angular velocity output by two six-axis motion processing sensors 32, and 60 is the number of time sampling points within one gait cycle. The first layer of the network is Conv1d(16, 16, kernel_size=5, padding=2), followed by BatchNorm and ReLU, with an output size of (B, 16, 60); the second layer is Conv1d(16, 32, kernel_size=5, padding=2), followed by BatchNorm and ReLU, with an output size of (B, 32, 60); then AdaptiveAvgPool1d(8) is used to compress the time dimension to 8, with an output size of (B, 32, 8); after being expanded to (B, 256) by Flatten, it is sequentially fed into Linear(256, 64), ReLU, Dropout(0.3) and Linear(64, 4), outputting four types of dynamic gait results: normal walking, running, pigeon-toed walking and pigeon-toed walking.

[0054] The aforementioned 16 channels are the basic input features for the prototype's dynamic gait recognition branch; the forefoot medial-lateral pressure ratio, heel medial-lateral pressure ratio, pressure center trajectory, plantar line direction, and auxiliary posture features within a short time window are synchronously calculated by the waist main control unit 2 and used for rule-gated state machine judgment, or as optional derived features to form a multi-channel temporal feature matrix together with the aforementioned basic inputs.

[0055] The rule-gated state machine performs a secondary judgment on the output of the quantized one-dimensional convolutional neural network. When the maximum class confidence Pmax is not less than 0.70, and the difference between the confidence scores of the first and second classes is not less than 0.15, and the temporal change directions of the plantar pressure distribution, pressure center trajectory, plantar line direction, and the two six-axis motion processing sensors 32 are not contradictory, inward or outward gait assist control is allowed. When Pmax is less than 0.55, the classes are inconsistent in two consecutive cycles, the sensor results contradict each other, or an unsuitable assist state such as running is detected, the SMA flexible fabric actuator 6 is kept dormant or enters a rapid release state. In the prototype dataset, the complete gait cycle is used as the test unit. The training set, validation set, and test set are divided into independent acquisition batches to avoid leakage across sets within the same consecutive acquisition batch. The prototype has accumulated 1326 gait cycles tested, of which 884 effective test windows were included in the statistics of the four types of dynamic gait, including 282 normal walking cycles, 100 running cycles, 242 inward gait cycles, and 260 outward gait cycles. The overall recognition accuracy was 85.55% in the test results; the recall rate for inward-pointing gait was 89.17%, and the recall rate for outward-pointing gait was 93.36%. Under the INT8 inference conditions at the edge of the GD32F470ZGT6, the inference latency for a single 60-frame gait cycle is approximately 6.9ms, which meets the real-time recognition requirements triggered by gait cycles.

[0056] See attached document Figure 3 The SMA flexible fabric actuator 6 includes a flexible fabric base 61, shape memory alloy composite yarns 62, Velcro fasteners 63, and an insulating fabric layer 64. The flexible fabric base 61 is a long, bandage-like knitted fabric, its length used to connect the waist and ankle, and its width adapting to the lower limb surface according to the circumference of the calf. The shape memory alloy composite yarns 62 are fixed to the side of the flexible fabric base 61 away from the skin by embroidery, stitching, or directional fiber placement. The fixing points are spaced apart along a predetermined spiral direction to define the traction direction of the shape memory alloy composite yarns 62, while allowing the flexible fabric base 61 to bend and deform locally with human movement. The insulating fabric layer 64 is disposed between the shape memory alloy composite yarns 62 and the skin, and can be made of heat-resistant and breathable textile materials. The insulating fabric layer 64 reduces the impact of the shape memory alloy composite yarns 62 heating up when energized on the skin and prevents direct contact between the shape memory alloy composite yarns 62 and the skin.

[0057] The flexible fabric base 61 has Velcro fasteners 63 at both ends. The proximal Velcro fastener 63 is connected to the waist control plate to fix the fabric 24, and the distal Velcro fastener 63 is connected to the ankle acquisition plate to fix the fabric 34. During installation, the flexible fabric base 61 is spirally wound from the waist side along the lower limb curve to the ankle; the fabric segment on the visible side of the wearer's front and the covering fabric segment on the back of the lower limb together form a continuous leg-wrapping path. After installation, the initial tension of the flexible fabric base 61 is adjusted so that the fabric fits the lower limb without causing significant restriction when no power is applied.

[0058] In this embodiment, the shape memory alloy composite wire 62 uses a nickel-titanium shape memory alloy wire with 50.0 wt% Ni, a diameter of 0.3 mm, and a single wire length of approximately 70 cm. A single nickel-titanium shape memory alloy wire generates a maximum contraction force of approximately 14 N under 7V, 0.8A driving conditions. Depending on the required assist level, one or more shape memory alloy composite wires 62 can be arranged in the same flexible fabric traction path and driven by the corresponding PWM drive channel within the current protection threshold range.

[0059] See attached document Figure 4 The flexible fabric traction path 66 for correcting inward-pointing SMA and the flexible fabric traction path 68 for correcting outward-pointing SMA are respectively a counterclockwise spiral path and a clockwise spiral path, viewed from the proximal end to the distal end along the longitudinal axis of the lower limb in a standing posture. Depending on the correction type predetermined by the user, one of the flexible fabric traction paths 66 for correcting inward-pointing SMA or 68 for correcting outward-pointing SMA is selectively constructed. Both traction paths extend from the waist main control plate fixing fabric 24 to the ankle acquisition plate fixing fabric 34, creating displacement in the length and circumferential directions of the lower limb.

[0060] When the SMA flexible fabric traction path 66 for correcting inward toeing is energized and contracts, the linear contraction force of the shape memory alloy composite yarn 62 acts in a counterclockwise helical tangential direction. This traction force, after being transmitted through the waist and ankle anchoring positions, forms a horizontal rotational component to reduce the tendency of toe inversion and tibial internal rotation, and a frontal lateral component to adjust the load and lateral force lines of the lower limbs. When the SMA flexible fabric traction path 68 for correcting outward toeing is energized and contracts, a traction force is generated in a clockwise helical tangential direction, thus forming a horizontal rotational component and a frontal lateral component opposite to the direction of the SMA flexible fabric traction path 66 for correcting inward toeing. Since the user knows whether they are inward or outward toeing, they wind the SMA flexible fabric actuator into the corresponding correction path when wearing the garment; the system only needs to recognize inward or outward toeing to initiate energized contraction assistance.

[0061] In an exemplary fabric arrangement, the longitudinal direction of the wearer's lower limbs is taken as the Z-axis, and the view along the longitudinal axis of the lower limbs from the proximal end to the distal end in the wearer's standing posture is taken as the reference. The circumferential angle φ of the horizontal section represents the position around the leg. φ=0° corresponds to the front side of the lower limb, φ=90° corresponds to the outer side, φ=180° corresponds to the rear side, and φ=270° corresponds to the inner side. If the wearer is pigeon-toed: the proximal anchor point of the SMA flexible fabric traction path 66 for correcting pigeon-toedness is set on the outer posterior region of the waist main control plate fixing fabric 24, with a circumferential angle of approximately 120° to 160°, and the distal anchor point is set on the anteromedial region of the ankle, with a circumferential angle of approximately 300° to 340°. If the wearer is splay-toed: the proximal anchor point of the SMA flexible fabric traction path 68 for correcting splay-toedness is set on the outer anterior region of the waist main control plate fixing fabric 24, with a circumferential angle of approximately 20° to 60°, and the distal anchor point is set on the posteromedial or posteromedial region of the ankle, with a circumferential angle of approximately 200° to 240°. When using the left or right leg, the circumferential angle is set mirror-image of the body midline.

[0062] The helix angle refers to the angle between the SMA traction path and the longitudinal axis of the lower limb. If the helix angle is too small, the horizontal rotational force is insufficient; if the helix angle is too large, the longitudinal traction is significant, and the wearing comfort decreases. In one embodiment, the helix angle is set to 8° to 28°, preferably 12° to 22°. When the effective length of the lower limb is 700mm to 900mm and the average radius is 45mm to 70mm, the circumferential angle of a single traction path can be set to 120° to 240°, thereby generating both horizontal rotational force and frontal lateral force when the SMA contracts.

[0063] Both the frontal and horizontal plane assist forces described above are distributed downwards through flexible fabric. Compared to rigid actuators that apply concentrated force at a single location, the long, bandage-like flexible fabric expands the contact area and allows the wearer to maintain natural knee and ankle movement during assistance. The magnitude and proportion of the dual-plane assist force are related to the helical path angle, the circumferential position of the path, the anchoring positions at the waist and ankle, the initial fabric tension, the number of shape memory alloy composite yarns 62, and the drive parameters. Before actual use, these parameters are adjusted according to the wearer's calf circumference and the target assist direction, and the actual traction direction after power is applied is verified.

[0064] In one example of inward-toeing gait assistance, the waist control unit 2 drives the currently pre-set corresponding path: the corrected inward-toeing SMA flexible fabric traction path 66, after the gait cycle recognition result meets the inward-toeing walking category and the temporal changes of the medial and lateral plantar pressure distribution and the two six-axis motion processing sensors 32 meet the inward-toeing direction characteristics. The waist control unit 2 drives this path in the corresponding gait phase of subsequent gait cycles, so that it simultaneously provides horizontal rotational assistance and frontal lateral assistance, and stops outputting at the toe-off event or the end of the gait phase.

[0065] In one example of outward-pointing gait assistance, after the gait cycle recognition result satisfies the outward-pointing walking category, and the corresponding plantar load characteristics and the temporal characteristics of the two six-axis motion processing sensors 32 satisfy the outward-pointing direction characteristics, the lumbar main control unit 2 drives the currently pre-set corresponding path: the corrected outward-pointing SMA flexible fabric traction path 68, and drives this path in the corresponding gait phase of subsequent gait cycles. After the path contracts, it provides horizontal rotational assistance and frontal lateral assistance in opposite directions to reduce the influence of the outward-pointing tendency on the lower limb force line and the direction of exoskeleton assistance.

[0066] When this system is used in conjunction with a lower limb exoskeleton, the exoskeleton continues to provide sagittal flexion and extension assistance for the hip, knee, and ankle joints, while the SMA flexible fabric actuator 6 provides frontal and horizontal assistance based on edge AI recognition results. This reduces the likelihood of increasing sagittal propulsion force only along incorrect gait directions.

[0067] The SMA flexible fabric actuator 6 employs four thermal control states: dormancy, burst preheating, isothermal holding, and rapid release. In the dormancy state, the pulse width modulation output is zero; in the burst preheating state, a higher duty cycle is used to rapidly heat up and shrink the shape memory alloy composite yarn 62; in the isothermal holding state, the duty cycle is reduced to maintain assist during the corresponding gait phase; in the rapid release state, the pulse width modulation output is quickly stopped, allowing the shape memory alloy composite yarn 62 to cool down and reduce traction.

[0068] In one embodiment, the SMA drive employs PWM control. The PWM frequency is set to 200Hz to 2kHz, preferably 1kHz; during the preheating phase, the duty cycle is 70% to 100%, lasting 300ms to 900ms, during which the temperature rises above the phase transformation initiation temperature, and the martensite within the SMA undergoes a diffusionless shear reverse phase transformation, spontaneously transforming into high-symmetry, high-modulus high-temperature austenite, with the crystal lattice rearranging to form a strong pulling force; during the isothermal holding phase, the duty cycle decreases to 25% to 55%, continuing until the end of the corresponding gait phase; during the rapid release phase, the duty cycle is 0, power is stopped and cooling occurs, causing the austenite to transform into martensite, and the crystal spontaneously forms directional twins. The pre-set microscopic residual internal stress within the material causes the shape memory alloy composite wire 62 to reduce its pulling force as it cools. For scenarios requiring only clear prompts rather than continuous assistance, a 100% short pulse conduction mode can also be used, outputting a finite-duration contraction pulse after recognizing an inward or outward octagon.

[0069] In an exemplary prototype test, using a 0.3mm diameter, approximately 700mm long shape memory alloy composite wire 62, under 25°C room temperature and natural air convection conditions, perceptible contraction was observed after 0.6 to 0.8 seconds of energization. After energization was stopped, the time for the traction force to drop below 30% of its peak value was approximately 1.8 to 2.6 seconds, and the recovery time to a stable triggering state was approximately 2.5 to 3.8 seconds. When the wearer walks and swings, generating airflow, the reset time can be shortened by approximately 15% to 25%. The cooling reset time is used to limit the continuous assist frequency, preventing the SMA from continuously contracting or overheating due to thermal inertia. The rapid release state corresponds to the rapid cessation of the pulse width modulation output and the decrease in traction force during the cooling process.

[0070] In an exemplary gait correction study, in-toeing and out-toeing gaits were compared before and after wearing the garment. Without assistance, the average inward toe angle of the in-toeing samples was approximately 12.6°, which decreased to approximately 5.4° after activating the corresponding SMA traction path; the average outward toe angle of the out-toeing samples was approximately 14.2°, which decreased to approximately 6.1° after activating the corresponding SMA traction path. Simultaneously, the offset of the forefoot medial-lateral pressure ratio and the heel medial-lateral pressure ratio decreased by approximately 18% to 26%. These results demonstrate that the spiral SMA flexible fabric can provide directional, flexible guidance for abnormal toe angles and plantar load shifts without increasing the rigidity of the joints.

[0071] See attached document Figure 5 The visual look-ahead unit 41 uses a global shutter grayscale camera, and its output is transferred to the static image buffer through a digital camera interface and a direct memory access module. After a frame completion interruption, the waist-mounted main control unit 2 extracts the vertical low-resolution region of interest and sequentially performs one-dimensional vertical difference, dynamic threshold binarization, line projection statistics, thickness sliding window integration, peak coordinate extraction, and perspective compression ratio determination. When the visual look-ahead unit 41 detects a dense horizontal edge structure conforming to the geometric patterns of a staircase, and the plantar pressure and inertial information do not confirm a normal flat-ground gait, the rule-gated state machine prioritizes a rapid release or de-loading strategy. This avoids the spiral traction path restricting the wearer's autonomous adjustment during stair climbing or stepping.

[0072] In one embodiment, the visual look-ahead unit 41 uses a 160×120 grayscale image for lightweight stair recognition. The waist-level main control unit 2 calculates the grayscale difference between adjacent rows line by line. When the absolute value of the grayscale difference is greater than 30, it is recorded as an edge point. If the number of edge points in a certain image row accounts for no less than 40% of the image width, it is recorded as a strong horizontal line. The system averages three consecutive frames. If no less than three strong horizontal lines are detected, and the standard deviation of the distance between adjacent strong horizontal lines is no greater than 8 pixels, it is determined to be a stair scene. The confidence level of the stair scene can be calculated as 60 + 10 × (number of strong horizontal lines - 3), and the maximum value is limited to 95. When the stair scene confidence level exceeds 70, the state machine immediately disables SMA assistance or enters fast release.

[0073] The ambient humidity detection unit 42 is used to obtain ambient humidity information and serves as an indirect gating condition for high humidity or slippery conditions. In one embodiment, the ambient humidity detection unit 42 uses a DHT11, and the humidity sampling position is located near the ventilation hole of the main control board housing to avoid misjudging the ambient humidity due to proximity to the SMA heating area. When the relative humidity is not lower than 85%RH, the system determines that there is a risk of high humidity or slippery conditions and shuts down the PWM output of the SMA flexible fabric actuator 6.

[0074] The user physiological detection unit 5 uses a photoelectric heart rate sensor module and is connected to the gait information acquisition unit 3. The signal is transmitted to the waist main control unit 2 via Bluetooth by the sub-board MCU 35 equipped with Bluetooth. When the user's heart rate exceeds the preset range, the waist main control unit 2 stops or reduces the assist and can control the SMA flexible fabric actuator 6 to enter the rapid release state.

[0075] The power safety protection unit 1 is located on the wearer's waist and connected to the drive circuit of the waist main control unit 2 and the SMA flexible fabric actuator unit 6. It is used to monitor overcurrent, overtemperature, and undervoltage conditions during the SMA driving process. When any safety condition is triggered, the power safety protection unit 1 outputs a safety signal to the waist main control unit 2, and the waist main control unit 2 shuts down or reduces the pulse width modulation output with a higher priority than the edge AI assist command. SMA temperature protection can be achieved through an NTC thermistor attached to the outer side of the middle section of the shape memory alloy composite yarn 62. The sampling point is preferentially set in the fabric interlayer area where heat dissipation is slowest. When the SMA surface temperature is higher than 55°C, the duty cycle is reduced, and when it is higher than 60°C, the drive is stopped immediately. A second temperature sampling point can be set near the drive MOSFET. When the driver temperature is higher than 80°C, the output is shut down. The current protection threshold can be set to a single-channel 1.2A current limit and 1.5A instantaneous shutdown. When any protection condition is triggered, the system enters a fast release state and records the fault flag.

[0076] See attached document Figure 6The edge AI decision-making process includes dynamic gait window slicing, linear interpolation resampling, multi-channel temporal feature matrix generation, quantized one-dimensional convolutional neural network classification, and rule-gated state machine judgment. The quantized one-dimensional convolutional neural network is deployed in the waist-mounted main control unit 2 through integer quantization, static memory allocation, and time-sharing inference scheduling to adapt to the computing power and power consumption limitations of wearable embedded devices. The waist-mounted main control unit 2 switches the SMA pulse width modulation output based on gait recognition results, safety status, and current thermal control status. When the current state is dormant, the recognition result meets the assistance conditions, and the gait enters the mid-support phase of the subsequent gait cycle, it enters the burst preheating stage. During normal walking, it takes 0.5-0.7 seconds to go from the mid-support phase to the mid-swing phase. After entering the burst preheating stage in the mid-support phase, the burst preheating takes 300-900ms. This time allows the force to be at its maximum in the mid-swing phase. After reaching the preset preheating time, it enters the constant temperature maintenance stage. When the gait phase ends, the recognition result changes, or the safety conditions are triggered, it is quickly released when the gait enters the swing phase. After the shape memory alloy composite wire 62 cools down to a state where it can be stably triggered again, it returns to dormant.

[0077] See attached document Figure 7 This embodiment also provides a method for SMA flexible fabric and gait guidance based on edge AI. First, user parameters are loaded, plantar pressure zero-point calibration and dual IMU zero-bias calibration are completed. Then, plantar pressure, dual IMU timing output, stair scene and ambient humidity identified by visual look-ahead unit 41 are collected, and heart rate status is received. Then, dynamic gait window slicing, feature matrix generation and edge AI recognition are performed. According to the recognition results, when entering the support phase of the subsequent gait cycle, the traction path 66 of the inward-pointing SMA flexible fabric is contracted to correct or the traction path 68 of the outward-pointing SMA flexible fabric is corrected to correct the inward-pointing SMA flexible fabric, the corresponding guidance intensity is determined, and dual-plane assistance is output according to the corresponding thermal control state.

[0078] After completing the current gait cycle, the system asynchronously synchronizes gait events, scene tags, guidance outputs, and safety status via Bluetooth Low Energy communication. The recorded data can be used to adjust user parameters, inward or outward gait determination thresholds, SMA drive parameters, and environmental safety gating thresholds.

[0079] The sensor model, sampling frequency, pulse width modulation parameters, number of shape memory alloy composite threads 62, and flexible fabric material in the above embodiments can all be adjusted according to the application scenario. Any scheme that uses SMA flexible fabric traction paths arranged in opposite directions or different circumferential directions, and provides flexible assistance to the frontal and horizontal planes of the lower limbs based on gait recognition results, as well as any variations, modifications, substitutions, and alterations made by those skilled in the art without departing from the principles and spirit of this invention, fall within the scope of the technical concept of this invention. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A SMA flexible fabric and gait guidance system based on edge AI, characterized in that, It includes a power safety protection unit (1), a waist main control unit (2), a gait information acquisition unit (3), an environmental monitoring unit (4), a user physiological detection unit (5), and an SMA flexible fabric actuator (6). The gait information acquisition unit (3) is used to collect plantar pressure and inertial information; The environmental monitoring unit (4) is used to acquire images of the terrain ahead and environmental humidity information; The user physiological detection unit (5) is used to collect the user's heart rate status; The power safety protection unit (1) is used to monitor overcurrent, overtemperature and undervoltage conditions during the SMA driving process; The waist main control unit (2) is used to centrally perform data reception, feature processing, edge AI inference, pulse width modulation control and security status management; The SMA flexible fabric actuator (6) has at least one clockwise or counterclockwise spiral path based on the longitudinal axis of the lower limb when viewed from the proximal end to the distal end in the wearer's standing posture, and is wrapped around the wearer's lower limb. The SMA flexible fabric actuator (6) provides flexible assistance in the frontal and horizontal planes of the lower limbs according to the output of the waist main control unit (2).

2. The system according to claim 1, characterized in that: The SMA flexible fabric actuation unit (6) includes a flexible fabric substrate (61) and at least one set of shape memory alloy composite yarns (62) disposed on the flexible fabric substrate (61), the shape memory alloy composite yarns (62) extending along the flexible fabric substrate (61) to define the traction direction and form a continuous leg-wrapping path.

3. The system according to claim 2, characterized in that: The flexible fabric substrate (61) is a long strip of bandage-shaped knitted fabric; The shape memory alloy composite yarn (62) is fixed to the flexible fabric substrate (61) by embroidery, sewing or directional fiber laying and is located on the side away from the skin. A heat insulation fabric layer (64) is provided between the shape memory alloy composite yarn (62) and the skin.

4. The system according to claim 2, characterized in that: The waist main control unit (2) includes a waist main control plate fixing fabric (24), and the gait information acquisition unit (3) includes an ankle acquisition plate fixing fabric (34). The SMA flexible fabric actuator (6) is provided with Velcro connectors (63) at both ends. The Velcro connectors (63) are respectively connected to the waist main control board fixing fabric (24) and the ankle acquisition board fixing fabric (34), so that the SMA flexible fabric actuator (6) is spirally wound from the waist to the ankle along the lower limb curve.

5. The system according to claim 4, characterized in that: The SMA flexible fabric execution unit (6) includes a corrective inward-pointing SMA flexible fabric traction path (66) and a corrective outward-pointing SMA flexible fabric traction path (68). The corrective inward-pointing SMA flexible fabric traction path (66) and the corrective outward-pointing SMA flexible fabric traction path (68) are respectively counterclockwise spiral paths and clockwise spiral paths based on the wearer's standing posture and the longitudinal axis of the lower limb from the proximal end to the distal end. According to the correction type predetermined by the user, the two are selectively constructed into one of the corrective inward-pointing SMA flexible fabric traction path (66) or the corrective outward-pointing SMA flexible fabric traction path (68). The path extends from the waist main control plate fixing fabric (24) to the ankle acquisition plate fixing fabric (34) and forms displacement in the length direction and circumferential direction of the lower limb.

6. The system according to claim 5, characterized in that: When the corrective inward-pointing SMA flexible fabric traction path (66) contracts, it forms a horizontal rotational force and a frontal lateral force to reduce the tendency of toe inward rotation and tibial internal rotation. When the corrective outward-pointing SMA flexible fabric traction path (68) contracts, it forms a horizontal rotational force and a frontal lateral force in opposite directions.

7. The system according to claim 1, characterized in that: The gait information acquisition unit (3) includes a four-channel plantar thin-film piezoresistive pressure sensor (31), two six-axis motion processing sensors (32) for acquiring the inertial information and forming a dual IMU, an ankle acquisition plate protective shell (33), an ankle acquisition plate fixing fabric (34), and a secondary board MCU (35) equipped with Bluetooth. The four-channel plantar piezoresistive pressure sensors (31) are respectively located on the inner side of the heel, the outer side of the heel, the inner side of the forefoot, and the outer side of the forefoot, and the two six-axis motion processing sensors (32) are respectively located near the tibia of the lower leg and in the arch area.

8. The system according to claim 7, characterized in that: The waist main control unit (2) includes a waist main control board protective shell (23), a waist main control board fixing fabric (24), and a motherboard MCU (25) equipped with Bluetooth. The environmental monitoring unit (4) includes a visual look-ahead unit (41) and an environmental humidity detection unit (42). The visual look-ahead unit (41) uses a global shutter grayscale camera. The output of the global shutter grayscale camera is transferred to the static image buffer through a digital camera interface and a direct memory access module. The visual look-ahead unit (41) is used to identify staircase scenes, and the environmental humidity detection unit (42) is used to acquire the environmental humidity information.

9. The system according to claim 8, characterized in that: The waist main control unit (2) takes the sudden increase in plantar pressure or the peak of inertial impact as the starting point of the gait cycle, and takes the toe-off event or the end event of the cycle corresponding to the next sudden increase in plantar pressure or the peak of inertial impact as the end point of the window. It also forms a multi-channel time sequence feature matrix by using the plantar pressure collected by the gait information acquisition unit (3) and the triaxial acceleration and triaxial angular velocity output by the two six-axis motion processing sensors (32). The waist main control unit (2) combines the stair scene and the environmental humidity information obtained by the environmental monitoring unit (4), the user heart rate status collected by the user physiological detection unit (5), and the detection results of the power safety protection unit (1) to obtain gait recognition results and perform safety status management through quantitative one-dimensional convolutional neural network classification and rule gating state machine judgment.

10. A method for guiding gait using SMA flexible fabric based on edge AI, applied to the system described in claim 9, wherein, The SMA flexible fabric actuation unit (6) in the system can be selected to include either a corrective inward-pointing SMA flexible fabric traction path (66) or a corrective outward-pointing SMA flexible fabric traction path (68), characterized in that it includes: Complete user parameter loading, plantar pressure zero-point calibration, and dual IMU zero-bias calibration; Collect the plantar pressure, dual IMU timing output, stair scene identified by the visual look-ahead unit (41), and environmental humidity information, and receive the user's heart rate status collected by the user physiological detection unit (5); Dynamic gait window slicing is performed based on the sudden increase in plantar pressure or the peak value of inertial impact, and the multi-channel temporal feature matrix is ​​generated. Gait recognition results are obtained and safety state management is performed by quantizing a one-dimensional convolutional neural network for classification and rule-gated state machine judgment. After selecting either the corrected inner eight SMA flexible fabric traction path (66) or the corrected outer eight SMA flexible fabric traction path (68), the guiding intensity and thermal control state are determined in the corresponding gait stage of the subsequent gait cycle according to the gait recognition result, and the pulse width modulation output is switched between dormancy, burst preheating, constant temperature maintenance and rapid release according to the thermal control state. Specifically, pulse width modulation output is stopped in the rapid release state.

Images