Multi-robot cooperative path planning and anti-collision dynamic obstacle avoidance welding method and system

Abstract

The invention relates to the technical field of intelligent manufacturing and industrial robots, in particular to a multi-robot collaborative path planning and anti-collision dynamic obstacle avoidance welding method and system.According to the system, a micro edge calculation module is integrated in each robot joint driver, and millisecond-level local decision is triggered based on the health state change rate; through a four-stage closed-loop process, health dynamic monitoring, composite risk assessment, health coupling obstacle avoidance execution and adaptive evolution, accurate identification and rapid response of a high-curvature path, adjacent machine approximation, arc interference and joint deterioration concurrent scene are realized. The invention provides a multi-robot collaborative path planning and anti-collision dynamic obstacle avoidance welding method and system to solve or at least alleviate the problem of obstacle avoidance failure under a composite fault caused by health monitoring and obstacle avoidance logic separation and centralized architecture response delay in the prior art, and provides a multi-robot collaborative path planning and anti-collision dynamic obstacle avoidance welding method and a multi-robot collaborative path planning and anti-collision dynamic obstacle avoidance welding system.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing and industrial robot technology, and in particular relates to a multi-robot collaborative path planning and collision avoidance dynamic obstacle avoidance welding method and system. Background Technology

[0002] With the trend of intelligent manufacturing developing towards higher precision and higher reliability, multi-robot collaborative welding systems have become a key technological support for achieving integrated molding of complex components in high-end manufacturing fields such as medical devices and aerospace. These applications place extremely stringent requirements on operational continuity, path accuracy, and system safety. They not only require robots to perform millimeter-level positioning and welding trajectory tracking in confined spaces, but also to maintain absolute collision avoidance capabilities in highly constrained environments with frequent dynamic obstacle interventions and rapid changes in operating conditions. Against this backdrop, path planning and dynamic obstacle avoidance are no longer merely geometric spatial avoidance problems, but have evolved into a multi-dimensional coupled control challenge integrating the robot's health status, environmental perception, and collaborative decision-making.

[0003] In existing technologies, patent CN116619381B proposes a high-dimensional path planning method based on bidirectional node tree expansion, which significantly improves the efficiency and stability of collaborative path generation for gantry-type dual-welding robot systems under static or pre-set obstacle conditions. By constructing a feasible path tree in the global configuration space, it can effectively avoid motion interference between robotic arms, and its engineering applicability was quite good in early welding scenarios dominated by fixed tooling and low dynamic interference. However, as welding tasks have developed towards high curvature structures (curvature > 25° / m) and miniaturized components, this method has exposed fundamental limitations. The centralized decision-making architecture of the central controller it relies on is limited by the communication link latency (≥ 1200ms) in the path update cycle, making it impossible to respond to millisecond-level sudden disturbances. More critically, its obstacle avoidance logic is completely disconnected from the real-time operating status of the robot body. Health parameters such as joint vibration and temperature rise are only used as the basis for offline maintenance and are not embedded in the obstacle avoidance triggering mechanism. If lubrication failure and high-load welding occur simultaneously, causing positioning drift (measured up to 1.8mm), the system is particularly prone to unexpected collisions because it lacks the ability to detect latent faults.

[0004] Another representative solution, CN115202365B, employs an improved swarm optimization algorithm combined with 3D grid modeling. It optimizes the global path by introducing a welding torch rotation angle influence factor, showing some progress in dynamic obstacle avoidance. However, its core remains based on a centralized software processing paradigm. All sensor data must be uploaded to the main control unit for fusion calculation, resulting in a lack of local response capability. Furthermore, its risk assessment relies on fixed thresholds, lacking adaptability when switching between conventional and precision welding modes, leading to a false trigger rate exceeding 35% in high-sensitivity scenarios. Particularly noteworthy is that this solution fails to incorporate the dynamic evolution of joint health status into the obstacle avoidance decision-making loop, making it unable to identify soft collision risks caused by progressive degradation such as bearing wear and overheating. While these risks may not immediately cause downtime, they accumulate positioning errors, eventually leading to irreversible quality defects at critical welds, and even equipment damage.

[0005] The common flaw in these technical solutions stems from the disconnect between health monitoring and obstacle avoidance execution logic in their architectural design. Traditional systems treat health data as post-event diagnostic information and obstacle avoidance as a purely geometric spatial problem, neglecting the fact that in highly coupled scenarios like high-precision welding, sudden changes in joint health, such as a sudden increase in the rate of change of vibration waveforms or abnormal temperature rise rates, are themselves precursory collision warnings. When health deterioration, along with factors like high-curvature paths, proximity of adjacent machines, and arc interference, creates a "quadruple composite fault," the communication bottlenecks and rigid thresholds of the centralized architecture prevent the system from freezing dangerous movements in time or performing rapid compensation at the local joint level. Moreover, due to the lack of a hardware-level closed-loop feedback mechanism, existing solutions struggle to achieve dynamic coordination between obstacle avoidance actions and health status, and cannot maintain operational continuity while ensuring safety.

[0006] Therefore, how to construct a closed-loop hardware system at the joint actuator level, triggering the health state change rate, integrating millisecond-level local decision-making, multi-level risk verification, collaborative signal broadcasting, and adaptive parameter evolution, to achieve accurate identification, rapid response, and gradual recovery of complex fault scenarios without relying on a central controller, has become the key to breaking through the current safety bottleneck of multi-robot collaborative welding. Solving this technical challenge is not only about order-of-magnitude improvements in obstacle avoidance response speed, but also involves a fundamental shift in the system from a passive defense to an active immunity paradigm. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art, solve or at least alleviate the obstacle avoidance failure caused by the separation of health monitoring and obstacle avoidance execution logic and the rigidity of threshold setting in the prior art, and provide a multi-robot collaborative path planning and collision avoidance dynamic obstacle avoidance welding method and system.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a multi-robot cooperative path planning and collision avoidance dynamic obstacle avoidance welding method, wherein a micro edge computing module is deployed inside the housing of each joint actuator of the robot, and the signal output terminals of the joint encoder, temperature sensor and motor driver are connected through a hardware signal splitter; the method includes: (S1) Dynamic monitoring of health status: Synchronously collect joint vibration waveform, temperature rise rate and load torque change rate. After anti-interference filtering, when the vibration and temperature rise change rate continuously exceed the preset threshold, trigger the path freeze command and open the cross-joint data channel. (S2) Composite scenario risk assessment: Combine local spatial curvature and obstacle distance to determine composite risk scenarios, mark the risk level according to the combination of multiple parameter mutations, expand the safety margin threshold under the highest risk level, and broadcast collaborative obstacle avoidance instructions through wireless beacons. Nearby robots respond to the instructions and perform avoidance actions. (S3) Healthy Coupling Obstacle Avoidance Execution: Under the highest risk level, the joint speed is limited and the structural stability is verified based on the data of adjacent joints. The path is compensated for by vector offset according to the safety margin. The reliability of the corrected path is ensured through a multi-layer verification mechanism. Then, the speed is gradually restored according to the healthy state and cross-channel is closed. (S4) Adaptive evolution: Record key event parameters to the health database. When the amount of data reaches the threshold, optimize the safety margin mapping logic through a neural network model. Dynamically adjust the collaborative waiting period based on the response efficiency of neighboring machines, and adaptively update the health mutation threshold based on the historical trigger frequency.

[0009] To further realize the present invention, the following technical solutions may be preferred: Preferably, the path freeze command in S1 is sent via a high-priority bus protocol, which can interrupt the path update command of the central controller and activate the visual warning signal at the end of the welding torch.

[0010] Preferably, the cooperative obstacle avoidance command in S2 includes a hierarchical coding system, with the highest risk code triggering the neighboring machine to retreat a predetermined distance along a preset safe direction and return a confirmation signal.

[0011] Preferably, the path compensation in S3 includes: Construct a virtual reference frame that integrates the poses of adjacent joints to counteract trajectory drift; A smooth transition segment is inserted when the deviation between the original path and the corrected path exceeds the limit. Path point coordinate modulation is dynamically adjusted based on real-time health features.

[0012] Preferably, the unfreezing request must meet two conditions to be effective: Send a request frame with a state lock to the central controller; The visual signal at the end of the welding torch is switched to the preset safe state.

[0013] Preferably, the speed recovery in S3 is divided into multiple stages of gradual improvement. Each stage must meet a decreasing health mutation rate threshold; otherwise, it will regress to the previous stage.

[0014] Preferably, the neural network model in S4 takes the historical mutation rate and safety margin as inputs and outputs a nonlinear adjustment mapping table to replace the fixed-proportion logic.

[0015] Preferably, the adaptive update of the health mutation threshold in S4 includes: Increase the threshold when the frequency of consecutive task triggers is too high; Lower the threshold when no consecutive tasks are triggered; The new threshold will take effect after the system restarts and the change log will be recorded.

[0016] A system for implementing the above method includes: Multiple welding robot units, each with a micro edge computing module embedded in its joint actuator; The module is connected to the encoder, temperature sensor and motor driver via a hardware splitter. The module's internal circuitry is configured to execute the entire process from S1 to S4, achieving hardware-level coupling between health status and obstacle avoidance control.

[0017] Preferably, the micro edge computing module is integrated with the motor driver housing via a heat-conducting medium, and internally integrates analog-to-digital conversion, hardware filtering, and multi-unit processing circuits.

[0018] The beneficial effects of this invention are: This invention deploys a micro edge computing module at the joint actuator level, using the rate of change in health status as the core criterion for obstacle avoidance triggering. It constructs a complete hardware closed loop from signal acquisition, mutation determination, risk classification, cooperative broadcasting, forced intervention, path compensation to gradual recovery and parameter evolution. All decisions are made locally, completely avoiding the communication bottleneck of centralized architecture. At the same time, through a four-level dynamic coupling mechanism (monitoring-evaluation-execution-evolution), the obstacle avoidance strategy and the health status of the host are coordinated, ensuring that in welding scenarios with high curvature, high dynamics, and high reliability requirements, collisions can be effectively prevented while maximizing the maintenance of work continuity. Attached Figure Description

[0019] Figure 1 This is a diagram showing the overall architecture of the system of the present invention; Figure 2 This is a schematic diagram of the hardware signal processing of the micro edge computing module of the present invention; Figure 3 This is a flowchart of step S1 of the method of the present invention; Figure 4 This is a flowchart of step S2 of the method of the present invention; Figure 5 This is a flowchart of step S3 of the method of the present invention; Figure 6 This is a flowchart of step S4 of the method of the present invention. Detailed Implementation

[0020] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0021] The technical solutions of 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.

[0022] Example 1 This embodiment discloses a multi-robot collaborative path planning and collision avoidance dynamic obstacle avoidance welding system. Its core lies in constructing a hardware closed-loop system architecture based on a joint actuator hierarchy. The system includes at least two six-axis industrial welding robots, each working collaboratively along a preset weld seam trajectory. Each robot consists of a base, upper arm, forearm, wrist, and welding torch end effector. The upper arm, forearm, and wrist each contain several rotary joints, each equipped with an independent joint actuator. The joint actuator integrates a motor, reducer, encoder, temperature sensor, and motor driver PWM output unit, and embeds a miniature edge computing module on its metal casing. This module is connected to the internal circuitry of the joint actuator via a high-density pin interface, and is also externally connected to a UWB beacon, LiDAR, and welding torch end effector LED status encoder, forming a complete perception-decision-execution-feedback link. By embedding a micro edge computing module inside the joint actuator housing, and utilizing existing hardware signal resources such as encoder channels, temperature sensor interfaces, and CAN buses, an obstacle avoidance decision-making and execution mechanism driven by the rate of change of health status is realized. In the scenario of high-density collaborative operation of multiple robots, a four-level deeply coupled control process of localized freezing, cross-joint compensation, collaborative release, and parameter adaptive evolution is completed.

[0023] Reference Figure 2The micro edge computing module is directly embedded in a reserved cavity inside the robot joint actuator housing, and is tightly bonded to the metal shell using thermal grease for efficient heat dissipation. The module integrates an analog-to-digital converter (ADC), hardware filtering circuitry, a dedicated register array, a mutation rate determination logic unit, a risk level marking engine, a safety margin reset controller, a cross-joint data channel manager, a speed limit actuator, a path correction generator, an LED state lock controller, a progressive recovery scheduler, a health parameter storage unit, a neural network coprocessor, a collaborative efficiency monitor, and a threshold adaptive adjuster. These functional units are connected via a low-latency on-chip interconnect bus, forming a local processing closed loop independent of the main control system.

[0024] The micro edge computing module receives three key physical signals via a hardware signal splitter: the first extracts the vibration waveform characteristics of the joint bearing from the A / B phase quadrature output of the incremental photoelectric encoder; the second reads the output voltage of the thermistor through a temperature sensor interface; and the third decodes the load torque change rate from the PWM signal output of the motor driver. Specifically, the vibration signal processing uses a high-speed comparator to shape and extract features in the 50Hz to 2kHz frequency band; the temperature rise signal is sampled after conditioning by an instrumentation amplifier using the analog voltage output of a PT1000 platinum resistance temperature sensor; and the load torque change rate is obtained by demodulating the optocoupler-isolated PWM signal through a digital phase-locked loop. All signals are processed by a band-stop filter with a center frequency of 35kHz and a bandwidth of 40kHz before entering the analog-to-digital conversion. This filter uses an LC passive topology and can effectively attenuate welding arc interference components in the 15kHz to 55kHz frequency band, ensuring that only low-frequency characteristic signals related to the health status of the mechanical structure are retained.

[0025] In one specific embodiment, when the robot's end effector enters a preset high-precision welding area (e.g., a weld segment with a curvature radius of less than 200 mm), the micro-edge computing module automatically initiates a multi-dimensional health signal synchronous capture process. This process is triggered by an internal state machine and requires no external instruction intervention. Specifically, the module performs a health dynamic feature package generation operation with a period of 150 milliseconds: vibration signals are sampled at 4 kHz for 120 points, and the main frequency energy ratio and kurtosis coefficient are extracted after FFT transformation; temperature rise signals are sampled at 10 Hz for 15 points, and the temperature rise rate is obtained by calculating the first-order differential mean; the PWM signal is sampled at 100 kHz to capture the complete cycle waveform, and the instantaneous load torque is inferred from the duty cycle-torque mapping table, and its second derivative is calculated as the rate of change index. The three sets of feature values ​​are normalized and then packaged and written into the Health Dynamic Feature Register (HDFR) in the dedicated register group. Example

[0026] Based on the hardware of Embodiment 1, this embodiment further discloses a multi-robot cooperative path planning and collision-avoidance dynamic obstacle avoidance welding method, referring to... Figures 3-6 Specifically, it includes the following steps: S1: Dynamic Monitoring Phase of Health Status In the hardware-based health mutation rate determination stage, the micro-edge computing module executes mutation detection logic based on dynamic thresholds. This logic is implemented by a hardware comparator circuit, requiring no software interrupts or polling. The specific determination conditions are defined as follows: when the vibration waveform change rate exceeds 25% per second and the temperature rise rate exceeds 3 degrees Celsius per second, the system marks the current joint as being in a "health mutation" state. The vibration waveform change rate is obtained by normalizing the Euclidean distance between the current and previous cycle's vibration energy spectra; the temperature rise rate is obtained by linearly extrapolating the time derivatives of the current and previous temperature rise values. If the flag status remains active for a continuous 150-millisecond sampling period, a dual-path hardware action is triggered: The first action is to send a 30-millisecond "path freeze hard pulse" signal to the driver of the same joint via the CAN controller. This signal uses the CAN 2.0B protocol, has an identifier of 0x7E8, and sets the 0th byte of the data field to 0xFF. It has the highest arbitration priority and can immediately interrupt any path update commands received by the driver from the central controller. The second action is to activate the RGB LED status encoder integrated at the end of the welding torch, causing it to flash red at a high frequency of 6 Hz with a duty cycle of 50%, providing a visual warning to the operator and nearby robots. After the freeze signal is generated, the module automatically opens a cross-joint data channel. This channel is connected to the corresponding edge computing module of the adjacent joint via a reserved high-speed serial link (SPI mode, clock frequency 20MHz), reserving communication bandwidth and dual-port RAM cache resources (allocated addresses from 0x2000 to 0x20FF) for subsequent collaborative compensation operations.

[0027] S2: Risk Assessment Phase for Complex Scenarios During the risk assessment phase of the composite scenario, upon receiving a self-generated freeze signal, the micro edge computing module immediately initiates the environment and health coupling verification process. This process first reads the local curvature data of the current workspace via the CAN interface (identifier 0x450) of the LiDAR. This data is calculated by real-time scanning of the weld contour by a 2D LiDAR mounted on the robot's base, followed by fitting a cubic spline curve. If the curvature value is greater than 5 degrees per meter, the robot is determined to be in a high-precision welding scenario. Simultaneously, the module detects the Euclidean distance between the current joint and the nearest static obstacle (like a tooling fixture) or dynamic obstacle (such as a neighboring robot arm) using UWB beacon ranging results or inverse kinematics model analysis. If this distance is less than 40 centimeters, combined with the existing "health mutation" state, the system is confirmed to have entered a "quadruple composite scenario," meaning that high curvature space, nearby obstacles, joint health mutation, and high-precision welding requirements all occur simultaneously. After successful verification, the module performs hardware labeling of the risk level based on the combination of mutation parameters. If only vibration changes abruptly, it's marked as a Level 1 risk, triggering a yellow LED to flash at 1 Hz. If both vibration and temperature rise change abruptly, it's marked as a Level 2 risk, triggering an orange LED to flash at 3 Hz. If vibration, temperature rise, and load torque all change abruptly, it's marked as a Level 3 risk, maintaining a red LED flashing at 6 Hz. Simultaneously with marking the risk level, the module generates a dynamic safety margin reset command. In Level 3 risk, the joint's safety distance threshold is forcibly increased to 220% of the standard value. This value is directly written to the specified offset address (0x3010) of the path planning coordinate correction register for subsequent use by the motion control unit during trajectory interpolation.

[0028] In the collaborative obstacle avoidance signal broadcasting mechanism, the micro edge computing module broadcasts risk level encoded signals through the integrated UWB beacon interface (based on the IEEE 802.15.4z standard). The specific encoding rules are as follows: Level 1 risk corresponds to hexadecimal code 0x12, which the neighboring robot does not need to perform any action upon receiving; Level 2 risk corresponds to code 0x24, which causes the neighboring robot to enter a path release preparation state and preload a safety retreat subroutine; Level 3 risk corresponds to code 0x36, which requires the neighboring robot to immediately perform a path release operation. When a neighboring robot receives the 0x36 code, its own edge computing module controls the driver to make the robot retreat 6 cm along a preset safe direction (usually the direction away from the robot itself, determined by offline calibration), with the retreat speed limited to 70% of the rated speed. It also illuminates a green LED flashing at a high frequency of 8 Hz as an acknowledgment signal and simultaneously returns a "Release Complete" response frame through the UWB channel. The frame format includes the source ID, target ID, action type, and timestamp. The local module continuously monitors the UWB channel. If it does not receive an acknowledgment signal from any neighboring machine within 150 milliseconds, it extends the path freeze period from 30 milliseconds to 100 milliseconds to enhance local protection and records communication timeout events to the log buffer.

[0029] S3: Health Coupling Obstacle Avoidance Execution Phase During the healthy coupling obstacle avoidance execution phase, after confirming the three levels of risk, the micro edge computing module executes dual hardware protection actions. The first action is to directly intervene in the duty cycle output of the PWM signal line, forcibly limiting the movement speed of the current joint to 50% of the rated speed, but not lower than the minimum speed threshold required by the welding process (usually 8 mm / s). This speed limit is achieved by adjusting the proportional gain of the speed loop inside the driver. The second action is to enable the aforementioned cross-joint data channel, reading the position feedback data of adjacent joints in real time at a sampling rate of 10 kHz (including angle, angular velocity, and encoder count) to build a local deformation compensation model. After the speed limit takes effect, the module will initiate a triple structural stability verification mechanism. The first verification requires that the fluctuation amplitude of the adjacent joint position (based on the smallest resolution unit of the encoder) is less than 0.25 mm and last for 30 milliseconds. Meeting this condition is considered "instantaneous stability". The second verification requires that the above condition be met twice consecutively, that is, within 60 milliseconds, which is considered "short-term stability". The third verification requires that the condition be met five times consecutively, that is, within 150 milliseconds, which is considered "long-term stability". If any verification fails, the rate-limiting period will be automatically extended by 0 milliseconds, and the verification process will restart. A maximum of three extensions are allowed, and the cumulative rate-limiting time cannot exceed 390 milliseconds.

[0030] Based on the stability verification results, the module performs a health compensation path generation operation. Specifically, it uses a safety margin value (e.g., 220%) to perform vector offset correction on the target path coordinates of the current joint. The offset direction is perpendicular to the tangent vector of the current motion trajectory and points towards the safe region. Simultaneously, it integrates real-time position data from adjacent joints to construct a virtual positioning reference system centered on this joint. This reference system fits the poses of adjacent joints using a weighted least squares method, with weight coefficients proportional to joint stiffness, to compensate for trajectory deviations caused by thermal deformation of the body or increased transmission clearance. After path correction is complete, a four-fold verification mechanism is initiated. The first layer of verification calculates the maximum deviation between the original path and the corrected path in Cartesian space. If it exceeds 0.5 mm, a cubic spline transition segment with 5 nodes is inserted between them, with the boundary condition being continuous position and velocity. The second layer of verification continuously modulates the path point coordinates during execution, with the modulation coefficient derived from the weighted sum of vibration energy and temperature rise rate in the health dynamic feature package, and the modulation amplitude not exceeding 0.15 mm. The third layer of verification generates a "health compensation complete" flag and writes it to the status register (address 0x4000) after the corrected path is fully executed, provided no new health mutation is triggered. The fourth layer of verification switches the LED status at the welding torch end from red 6 Hz flashing to green 2 Hz flashing, with the duty cycle remaining at 50%. Only after this LED status switch is completed does the module send a unfreeze request frame (identifier 0x7E9, data byte 0x01) to the central controller via the CAN bus. This request includes a state lock mechanism—the central controller only allows the restoration of path update permissions when it detects that the LED frequency has dropped from 6 Hz to 2 Hz (monitored in real-time by a photosensitive sensor array).

[0031] During the progressive recovery control phase, after thawing takes effect, joint movement speed gradually recovers according to a five-stage strategy: In the first stage (0 to 80 milliseconds), the speed increases to 65% while continuously monitoring the healthy mutation rate; in the second stage (80 to 200 milliseconds), if the mutation rate is below 20% per second, it increases to 80%; in the third stage (200 to 380 milliseconds), if the mutation rate is below 15% per second, it increases to 90%; in the fourth stage (380 to 600 milliseconds), if the mutation rate is below 10% per second, it increases to 98%; and in the fifth stage (600 to 850 milliseconds), if the mutation rate is below 5% per second, it fully recovers to 100%. In any stage, if the healthy mutation rate rises above the corresponding threshold, it immediately reverts to the speed setting of the previous stage and restarts the timing. Once the speed is restored to 100%, the module automatically closes the cross-joint data channel, clears the relevant cache flags (zeroing addresses 0x2000 to 0x20FF), and resets all status registers (including risk level, freeze flag, compensation completion flag, etc.) to prepare for the next health mutation event.

[0032] S4: Adaptive Evolutionary Phase During the adaptive evolution phase, after each complex scenario is processed, the micro-edge computing module executes a parameter solidification process. Specifically, key parameters such as the peak health mutation rate (the maximum value of vibration, temperature rise, and torque), safety margin expansion ratio (e.g., 220%), compensation path deviation (Euclidean distance in Cartesian space), and neighboring machine response time (time from broadcast to receipt of acknowledgment) are written into the health database in the internal ROM. Each node independently maintains a circular buffer with a capacity of 8 records, using a FIFO strategy to overwrite the oldest record. When the health database stores 4 sets of valid data, the module initiates a hardware-accelerated learning process. The neural network coprocessor loads a lightweight convolutional neural network model (3 nodes in the input layer, 16 nodes in the hidden layer, 1 node in the output layer, with ReLU activation function). The input is a historical mutation rate sequence (8 nodes in length) and the corresponding safety margin adjustment value, and the output is an optimized dynamic reset mapping table. This table is stored in dedicated SRAM in the form of a lookup table to overwrite the original fixed ratio logic (e.g., 220%), achieving non-linear adaptive adjustment of the safety margin. For example, if historical data shows that high vibration abrupt changes are often accompanied by small actual offsets, the new mapping table may adjust the safety margin to 180% instead of 220%.

[0033] Regarding the evolution of cooperative obstacle avoidance capabilities, the module continuously monitors the timestamps of "release complete" signals returned by neighboring robots in the UWB channel and calculates the average time taken for path release operations. If this time is less than 150 milliseconds (based on statistics from the last 5 events), the current cooperative efficiency is deemed to be up to standard; otherwise, the confirmation waiting period is automatically shortened from 150 milliseconds to 120 milliseconds, and a "cooperative optimization request" data frame (CAN identifier 0x7EA, data byte 0x02) is sent to the central controller. After this request is parsed by the central controller, the "path release" subroutine (containing a 6 cm backward trajectory and velocity curve) is preloaded into the motion planning task queue of the low-priority robot, allowing it to jump directly to execution upon receiving the 0x36 encoding, thus reducing decision latency by approximately 40 milliseconds. Simultaneously, when constructing the virtual positioning reference frame, the sampling point density of the robot's coordinate interpolation algorithm is increased by 35% from the default 5 points per centimeter, to 6.75 points per centimeter (rounded up to 7 points), thereby improving compensation accuracy. Actual measurements show that this can reduce trajectory deviation by 18%.

[0034] Regarding adaptive adjustment of health thresholds, the module periodically analyzes the statistical distribution of historical mutation rate data in the health database. If the health mutation judgment condition is triggered more than 7 times in 10 consecutive welding tasks (i.e., the trigger rate is greater than 70%), the current threshold is deemed too sensitive, and the vibration waveform change rate threshold is automatically increased by 5% from 25% per second to 26.25%. If it is not triggered in 20 consecutive tasks, the threshold is deemed too lenient, and it is automatically decreased by 3% to 24.25%. The adjusted threshold is written to the hardware comparator's configuration register (0x5000) and synchronized to the backup register via the I²C interface. Simultaneously, a "threshold change" log record containing the old threshold, new threshold, adjustment reason ("over-triggered" or "under-triggered"), and timestamp is generated and stored in non-volatile EEPROM (starting from address 0x8000). This new threshold takes effect after the next system power-on self-test, ensuring that the adjustment process does not affect the continuity of the current job.

[0035] Example 3 To verify the technical effectiveness of this invention, the following reference examples and comparative examples were designed for comparative testing.

[0036] The test platform consisted of two six-axis industrial robots (model KUKA KR AGILUS R900) collaboratively welding the B-pillar reinforcement plate of an automotive body-in-white. The total length of the weld was 1.2 meters, containing multiple high-curvature areas (minimum radius of curvature 180 mm). The minimum distance between the end effectors of the two robots was 35 cm, and the cycle time for each piece was required to be no more than 45 seconds.

[0037] Referring to Example 1: Using the system described in this invention, the joint actuator integrates a micro edge computing module to execute the aforementioned four-level deep coupling control process. During the test, a joint health mutation event was artificially injected, applying an external impact to the shoulder joint of robot A, causing its vibration waveform change rate to rise to 28.7% / s within 150ms, and the temperature rise rate to reach 3.8℃ / s simultaneously.

[0038] Comparative Example 1: A traditional centralized obstacle avoidance scheme with a central controller is adopted. Health monitoring is periodically queried by the host computer software (every 500 milliseconds), and obstacle avoidance decisions rely on global path replanning.

[0039] Comparative Example 2: Joint-level health monitoring is adopted, but there is no collaborative broadcast mechanism; only local freezing is performed, and there is no linkage with neighboring machines.

[0040] Test metrics include stealth collision interception success rate (the proportion of obstacle avoidance completed before successful physical contact), average obstacle avoidance response delay, welding trajectory deviation (RMS error relative to the ideal weld), and number of mission interruptions (due to emergency stops).

[0041] The test results are shown in the table below: Test Items Reference example 1 Comparative Example 1 Comparative Example 2 Hidden collision interception success rate 98.7% 76.3% 82.1% Average obstacle avoidance response time (ms) 152 510 168 Welding trajectory deviation (mm RMS) 0.21 0.72 0.56 Number of task interruptions (per 1000 items) 0.8 12.5 7.3 Maximum positioning drift (mm) 0.21 0.92 0.68 Test results show that Reference Example 1 is significantly better than the Comparison Example 2 in all metrics. Particularly in terms of response latency, Reference Example 1 can control it within 150ms, while Comparison Example 1, relying on a central solver, suffers from a latency as high as 510ms, making it impossible to intercept rapidly evolving latent faults. While Comparison Example 2 possesses local response capabilities, it lacks a collaborative mechanism, leading to multiple potential collisions when neighboring machines fail to avoid collisions, resulting in a relatively high task interruption rate. Reference Example 1, through a health-obstacle avoidance hardware closed loop and four-level coupling control, achieves safe and continuous operation in high-density collaborative scenarios.

[0042] Furthermore, in the verification of the adaptive evolution function, referring to Example 1, after running continuously for 500 hours, the safety margin adjustment mapping table successfully optimized the average safety distance from the initial 220% to 195%. While maintaining an interception rate of over 98%, it reduced the average trajectory from 0.28 mm to 0.23 mm, which proves that the parameter adaptive mechanism is effective.

[0043] In summary, this invention achieves deep hardware-level coupling of health status and obstacle avoidance control by deploying miniature edge computing modules at the joint actuator level, capable of signal acquisition, mutation detection, local freezing, cooperative broadcasting, compensation generation, progressive recovery, and parameter evolution. This system does not rely on global solutions from a central controller; all critical decisions can be made within 150 milliseconds, effectively intercepting latent collision precursors caused by internal joint wear, thermal deformation, or increased transmission clearance. This significantly improves the safety boundaries and operational continuity of multi-robot collaborative welding in confined curvature spaces.

[0044] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multi-robot cooperative path planning and collision avoidance dynamic obstacle avoidance welding method, characterized in that, A miniature edge computing module is deployed inside the housing of each joint actuator of the robot, and connected to the signal output terminals of the joint encoder, temperature sensor, and motor driver via a hardware signal splitter; the method includes: (S1) Dynamic monitoring of health status: Synchronously collect joint vibration waveform, temperature rise rate and load torque change rate. After anti-interference filtering, when the vibration and temperature rise change rate continuously exceed the preset threshold, trigger the path freeze command and open the cross-joint data channel. (S2) Composite scenario risk assessment: Combine local spatial curvature and obstacle distance to determine composite risk scenarios, mark the risk level according to the combination of multiple parameter mutations, expand the safety margin threshold under the highest risk level, and broadcast collaborative obstacle avoidance instructions through wireless beacons. Nearby robots respond to the instructions and perform avoidance actions. (S3) Healthy Coupling Obstacle Avoidance Execution: Under the highest risk level, the joint speed is limited and the structural stability is verified based on the data of adjacent joints. The path is compensated for by vector offset according to the safety margin. The reliability of the corrected path is ensured through a multi-layer verification mechanism. Then, the speed is gradually restored according to the healthy state and cross-channel is closed. (S4) Adaptive evolution: Record key event parameters to the health database. When the amount of data reaches the threshold, optimize the safety margin mapping logic through a neural network model. Dynamically adjust the collaborative waiting period based on the response efficiency of neighboring machines, and adaptively update the health mutation threshold based on the historical trigger frequency.

2. The method according to claim 1, characterized in that, The path freeze command described in S1 is sent via a high-priority bus protocol, which can interrupt the path update command of the central controller and activate the visual warning signal at the end of the welding torch.

3. The method according to claim 1, characterized in that, The cooperative obstacle avoidance command described in S2 includes hierarchical coding. The highest risk code triggers the neighboring machine to retreat a predetermined distance along a preset safe direction and return a confirmation signal.

4. The method according to claim 1, characterized in that, The path compensation described in S3 includes: Construct a virtual reference frame that integrates the poses of adjacent joints to counteract trajectory drift; A smooth transition segment is inserted when the deviation between the original path and the corrected path exceeds the limit. Path point coordinate modulation is dynamically adjusted based on real-time health features.

5. The method according to claim 5, characterized in that, For a freeze request to take effect, two conditions must be met: Send a request frame with a state lock to the central controller; The visual signal at the end of the welding torch is switched to the preset safe state.

6. The method according to claim 1, characterized in that, In S3, speed recovery is divided into multiple stages of gradual improvement. Each stage must meet a decreasing health mutation rate threshold; otherwise, it will regress to the previous stage.

7. The method according to claim 1, characterized in that, The neural network model described in S4 takes the historical mutation rate and safety margin as inputs and outputs a nonlinear adjustment mapping table to replace the fixed-proportion logic.

8. The method according to claim 1, characterized in that, The adaptive update of the health mutation threshold in S4 includes: Increase the threshold when the frequency of consecutive task triggers is too high; Lower the threshold when no consecutive tasks are triggered; The new threshold will take effect after the system restarts and the change log will be recorded.

9. A system for implementing the method according to any one of claims 1-8, characterized in that, include: Multiple welding robot units, each with a micro edge computing module embedded in its joint actuator; The module is connected to the encoder, temperature sensor and motor driver via a hardware splitter. The module's internal circuitry is configured to execute the entire process from S1 to S4, achieving hardware-level coupling between health status and obstacle avoidance control.

10. The system according to claim 9, characterized in that, The micro edge computing module is integrated with the motor driver housing via a heat-conducting medium, and internally integrates analog-to-digital conversion, hardware filtering, and multi-unit processing circuits.

Images