Equipment maintenance robot group with maintenance experience mutual learning function and working method of equipment maintenance robot group

Abstract

The invention discloses an equipment maintenance robot group for mutual learning of maintenance experience and a working method thereof. The system comprises a cloud server, at least three maintenance robots, an edge computing node and a distributed sensor network. The cloud server stores a global maintenance experience library, and a block chain-database hybrid architecture is adopted to ensure that data cannot be tampered and is efficiently stored. The maintenance robot integrates a multi-modal data acquisition module, a federated learning module, a heterogeneous communication module, a maintenance execution module and an energy management module. The edge computing node is responsible for local experience verification, conflict resolution and model aggregation, and cloud load is reduced. And the distributed sensor network feeds back the equipment state in real time to assist in fault pre-judgment. Through the steps of multi-modal data acquisition and preprocessing, local experience generation, encryption and uploading, edge verification and aggregation, global collaborative optimization, dynamic task allocation and execution and the like, robot group cross-regional maintenance experience mutual learning is realized, and the intelligent level and overall efficiency of equipment maintenance are improved.

Description

Technical Field

[0001] This invention relates to the field of equipment maintenance robot technology, specifically to a group of equipment maintenance robots that learn from each other's maintenance experience and their working method. Background Technology

[0002] Against the backdrop of accelerated industrial intelligent transformation, equipment maintenance, as a core link to ensure production continuity, directly impacts enterprise operating costs and production safety in terms of efficiency and accuracy. Traditional equipment maintenance models, which mainly rely on manual inspections or single-machine automated robot operations, are gradually becoming unable to meet the high dynamic and high reliability requirements of complex industrial scenarios, exposing numerous technical bottlenecks.

[0003] Existing maintenance robots mostly adopt a standalone, independent learning model, and their experience accumulation is limited to the maintenance tasks they perform, lacking a mechanism for sharing experience across devices and regions. When a robot encounters a new type of fault, it must start exploring solutions from scratch, which not only prolongs the fault handling cycle but may also cause secondary damage to the equipment due to repeated trial and error. At the same time, different robots are equipped with different sensor models and brands, and the formats and accuracy of the fault characteristic data they collect are inconsistent, making it difficult to reuse experience across machines and creating a "data silo" phenomenon.

[0004] Regarding data security and trustworthiness, traditional maintenance experience databases often employ centralized storage, which presents challenges such as data tampering and difficulties in tracing the source. The authenticity of maintenance cases lacks effective verification mechanisms, and erroneous experiences may spread through sharing, leading to collective maintenance errors. Furthermore, communication between robots and the cloud, and between robots themselves, often relies on fixed encryption methods with delayed key updates, making them vulnerable to security threats such as man-in-the-middle attacks and data theft. Especially in industrial IoT environments, the leakage of sensitive equipment operating data can pose production safety risks.

[0005] Insufficient communication efficiency and adaptability are also key factors restricting maintenance effectiveness. Existing systems mostly rely on a single communication protocol, which is prone to network congestion under high-load scenarios (such as large-scale experience synchronization), while consuming too much energy under low load. At the same time, the cloud-based centralized processing of all data leads to increased communication latency, making it difficult to meet the real-time requirements of fault prediction and emergency handling. The computing resources of edge nodes are not effectively utilized, resulting in concentrated cloud load and wasted resources.

[0006] The lack of experience iteration and reuse mechanisms further reduces the adaptability of maintenance systems. Traditional experience bases lack dynamic optimization capabilities, outdated or inefficient maintenance strategies cannot be eliminated in a timely manner, and the generation of emergency solutions for new fault types relies on manual intervention, resulting in slow response speeds. There are technical barriers to cross-scenario experience transfer, and maintenance experience for different equipment types (such as motors and hydraulic systems) is difficult to adapt quickly, leading to insufficient robot handling capabilities when faced with unfamiliar equipment faults.

[0007] In addition, the intelligence level of task scheduling and energy management is insufficient. Existing systems mostly adopt a fixed allocation mode and do not dynamically adjust tasks based on the robot's real-time status (location, power, historical success rate), which can easily lead to resource misallocation. Energy management lacks accurate energy consumption prediction and charging scheduling mechanisms, resulting in frequent robot shutdowns due to low power, which affects the continuity of maintenance tasks.

[0008] Therefore, there is an urgent need to build a robot group maintenance system with cross-machine experience learning, secure and reliable communication, edge intelligent collaboration, and dynamic optimization and iteration capabilities, in order to break through the technical limitations of the traditional model and improve the intelligence level and overall efficiency of industrial equipment maintenance. Summary of the Invention

[0009] The technical problem to be solved by this invention is to provide a group of equipment maintenance robots that learn from each other's maintenance experience and their working method, which solves problems such as "data silos" formed by differences in sensors between different robots, lack of assurance of the reliability of maintenance experience, insufficient communication efficiency and security, ineffective utilization of edge computing power, and low level of intelligence in experience iteration and task scheduling.

[0010] To address the aforementioned technical problems, embodiments of the present invention provide the following technical solution: a group of equipment maintenance robots that learn from each other's maintenance experience and a method for operating the same, comprising a cloud server, at least three maintenance robots, an edge computing node, and a distributed sensor network; The cloud server is used to store a global maintenance experience base, robot identity authentication information and task scheduling data, and realizes data interaction with edge computing nodes through a hybrid communication protocol; The maintenance robot includes a multimodal data acquisition module, a federated learning module, a heterogeneous communication module, a maintenance execution module, and an energy management module; wherein: The multimodal data acquisition module integrates lidar, infrared sensors, and high-definition cameras to acquire vibration spectrum, temperature field distribution, surface defect images, and operating parameters of the target equipment. The federated learning module adopts a hierarchical training mechanism. Local nodes perform initial fusion of maintenance experience based on the federated averaging algorithm, and then cross-regional experience aggregation is achieved through edge computing nodes. The heterogeneous communication module supports the IPv6 / 5G converged protocol and adopts a security mechanism that combines dynamic key encryption and identity authentication to achieve bidirectional data transmission with cloud servers, edge nodes and other robots; The maintenance execution module includes a six-degree-of-freedom robotic arm and a tool library, which performs operations, including but not limited to disassembly, inspection, and replacement, based on the optimized maintenance strategy. The energy management module dynamically adjusts the allocation of communication and computing resources through an energy consumption prediction model, and automatically requests charging scheduling when a low power warning is triggered. The edge computing nodes are deployed at the edge of the subnet in the robot's activity area to process the validity verification of local maintenance experience and resolve conflicts in real time, thereby reducing the communication load on the cloud server. The distributed sensor network consists of IoT sensors deployed on the target device, used to provide real-time feedback on the device's operating status and assist the maintenance robot in predicting faults.

[0011] Furthermore, the global maintenance experience base adopts a blockchain-database hybrid architecture, including: The blockchain layer stores the hash values ​​of maintenance cases that have been verified through consensus, ensuring that the experience data is immutable. Each block contains a case ID, timestamp, robot signature, and the hash of the previous block. The database layer adopts a sharded storage strategy, dividing the data into shards according to the device type. Each shard contains a fault feature vector, a maintenance step sequence, an effect evaluation matrix, and a related case index.

[0012] Furthermore, the security mechanism of the heterogeneous communication module includes: Based on device fingerprint and blockchain identity authentication, the robot hardware serial number and IPv6 address generate a unique device fingerprint, which is then verified by the cloud server and written into the blockchain for storage. Dynamic key generation uses the SM4 national cryptographic algorithm and combines the current timestamp, device fingerprints of both communicating parties, and physical location information to generate a session key. The communication range is adaptively adjusted, and the communication mode is dynamically switched according to the task load. 5G slicing technology is enabled when the load is high, and LoRaWAN protocol is switched when the load is low.

[0013] Furthermore, the federated learning module includes: The feature alignment unit uses a contrastive learning algorithm to standardize the fault features collected by different robots, eliminating feature offsets caused by differences in sensor models. The local training unit, based on the improved deep reinforcement learning (DRL) algorithm, uses the reward function of maintenance operations to generate local maintenance policies. The model aggregation unit performs weighted aggregation of the local models of each robot through edge nodes, and the weights are dynamically adjusted according to the robot's historical maintenance success rate. The strategy migration unit uses a meta-learning algorithm to quickly adapt maintenance experience from other scenarios to local device types and generate emergency handling plans for new faults.

[0014] A method for operating a group of equipment maintenance robots that learn from each other's maintenance experience using wireless communication technology, applied to any of the systems described above, includes the following steps: Step S1: Multimodal data acquisition and preprocessing The maintenance robot obtains three-dimensional point cloud data by scanning the surface of the target equipment with LiDAR, collects temperature distribution heat maps with infrared sensors, and captures defect images with high-definition cameras. After removing noise from the data through wavelet transform, fault feature parameters, including but not limited to vibration frequency peak, temperature gradient, and defect area, are extracted and encapsulated into feature vectors in JSON format. Step S2: Local experience generation and encrypted upload The local training unit of the federated learning module matches historical cases based on feature vectors. If the matching degree is greater than the threshold, the historical policy is directly called; otherwise, DRL training is started to generate a new maintenance policy and calculate the effect evaluation value. The heterogeneous communication module adopts a dynamic key encryption strategy to upload the new policy to the edge computing node and receive the shared experience summary of robots in the same area. Step S3: Experience verification and aggregation of edge nodes Edge nodes perform three-layer verification of the received maintenance experience: Format validation: Check if the required fields for fault characteristics, operation steps, and effect evaluation are included; Consistency verification: Verify the logical consistency between the new strategy and similar cases by querying historical cases through the blockchain; Effect verification: If the estimated resolution rate of the new strategy is less than the threshold, it will be pushed to the senior maintenance robot for review; After verification, the edge nodes use a federated averaging algorithm to aggregate the maintenance models of local robots, generate regional maintenance strategies, and synchronize them to the cloud server. Step S4: Global Experience Collaborative Learning and Strategy Optimization The cloud server integrates regional-level strategies with a global experience base and uses a similarity algorithm to match cross-regional related experiences for each robot. After the maintenance robot receives the global experience, the policy migration unit generates an optimized strategy adapted to the local scenario based on meta-learning, updates the local model, and feeds back the optimization effect to the cloud server. Step S5: Dynamic Task Allocation and Maintenance Execution When the distributed sensor network detects a device malfunction, the cloud server allocates tasks based on the robot's real-time location, battery level, and historical success rate. Using an improved auction algorithm, robots bid based on their own load, and the winning bidder obtains the maintenance rights. The maintenance robot invokes the optimized strategy and performs operations through the maintenance execution module. Process data is uploaded to the cloud server in real time via edge nodes. After a task is completed, the system updates the robot's reputation value based on the actual resolution effect. Robots with reputation values ​​below the threshold are restricted from receiving high-priority tasks.

[0015] Furthermore, the process of extracting the fault characteristic parameters in step S1 includes: S11: Vibration feature extraction: Perform Fourier transform on the original vibration signal and extract the peak frequency and corresponding amplitude; S12: Temperature Feature Extraction: Extract the highest temperature, temperature standard deviation, hot spot area, and temperature change rate from the heat map; S13: Visual feature extraction: Use the YOLOv8 algorithm to identify defect types and calculate the defect area, perimeter, and circumscribed rectangle dimensions; S14: Normalize the above features to the [0,1] interval to form a feature vector, in which vibration features account for 40% of the weight, temperature features account for 30%, and visual features account for 30%.

[0016] Furthermore, the implementation process of dynamic key encryption in step S2 includes: S21: The sending robot generates a device fingerprint: the hardware serial number + IPv6 address are hashed using SHA-256, and the first 64 bits are taken as the device fingerprint; S22: Generate an initial key by combining the current timestamp and the physical location coordinates of the receiver; S23: Encrypt the empirical data using the SM4 algorithm, and introduce random numbers during the encryption process to prevent replay attacks; S24: When using 5G communication, network slicing is enabled to allocate dedicated channels for experience transmission. When the channel quality index is lower than the set threshold, it automatically switches to the TCP retransmission mechanism.

[0017] Furthermore, the aggregation process of the federated averaging algorithm described in step S3 includes: S31: Edge nodes collect local model parameters and reputation values ​​for each robot and calculate weighting coefficients. It is a quantitative assessment of the overall reliability, accuracy, and compliance capabilities demonstrated by the i-th robot during task execution; This refers to the total amount of data related to model training collected and processed by the i-th robot within a certain period; S32: Calculate a weighted average of the model parameters to generate a regional model. in, For local model parameters, These are parameters for the regional model. S33: Differential privacy technology is used to add Gaussian noise to protect the local data privacy of each robot; S34: Compare the aggregated model with the global experience base. If the difference exceeds the threshold, trigger a second aggregation verification.

[0018] Furthermore, the meta-learning adaptation process described in step S4 includes: S41: Select cases similar to local device types from global experience and build a support set; S42: Based on the MAML algorithm, the model is trained on the support set to initialize parameters, enabling the model to converge quickly with a small amount of local data; S43: Use local fault data to fine-tune the initialization model and generate a maintenance strategy adapted to the local scenario. An early stop strategy is adopted during the fine-tuning process. S44: The new strategy is simulated and verified through Monte Carlo tree search. If the simulation success rate is greater than the threshold, it is put into actual use.

[0019] Furthermore, it also includes experience iteration and self-healing steps: S51: The cloud server iterates and optimizes the global experience database weekly, calculating the comprehensive score for each experience: Score = 0.6 × resolution rate + 0.2 × (1 - time spent / average time spent) + 0.2 × (1 - material consumption / average consumption) S52: For experiences with a score <60, a group revision mechanism is triggered, consisting of a revision group of 3 or more robots with a reputation score ≥90, which generates an iterative version through federated learning; S53: When the robot detects a fault in its own sensor, it automatically sends a self-healing request to the edge node. The edge node schedules redundant data from neighboring robots to compensate for the fault and generates a sensor maintenance reminder.

[0020] The beneficial effects of the above-described technical solution of the present invention are as follows: 1. This invention overcomes the limitations of isolated experience in traditional single-machine maintenance robots by constructing a multi-layered collaborative architecture of "cloud-edge-device-network". Utilizing a layered training mechanism of federated learning, robots can generate local maintenance strategies locally and then aggregate cross-regional experience through edge nodes, achieving mutual learning of experience without the need for centralized sharing of raw data. This model effectively eliminates "data silos" caused by differences in sensors among different robots, allowing high-quality maintenance experience to quickly circulate and be reused within the group, significantly improving the overall response and handling capabilities of the robot swarm for complex equipment failures.

[0021] 2. This invention provides reliable protection through a multi-dimensional security mechanism. Device fingerprint and blockchain-based identity authentication ensures the robot's identity cannot be forged, while dynamic key encryption technology guarantees the confidentiality of data transmission, effectively resisting various network attacks. Simultaneously, the global experience base adopts a blockchain-database hybrid architecture, ensuring both the immutability of maintenance cases and efficient retrieval and management of experience, providing a reliable source of experience for robot groups and preventing maintenance errors caused by the spread of erroneous experience.

[0022] 3. This invention significantly improves the efficiency and continuity of maintenance tasks through intelligent scheduling and dynamic optimization mechanisms. A distributed sensor network monitors equipment status in real time, and a cloud server dynamically allocates tasks based on the robot's real-time status, ensuring that the appropriate robot performs the corresponding task. The energy management module dynamically regulates energy consumption and automatically requests charging scheduling to avoid work interruptions due to power issues. An experience iteration mechanism regularly optimizes the global experience library; low-quality experiences are revised and upgraded by the group, allowing the system to continuously adapt to new scenarios and ensuring a stable improvement in long-term maintenance efficiency. Attached Figure Description

[0023] Figure 1 This is a block diagram illustrating the principle of a group of equipment maintenance robots that learn from each other's maintenance experience, as per the present invention. Figure 2 This is a diagram showing the composition of the equipment maintenance robot group robot module for mutual learning and maintenance experience according to the present invention. Figure 3 This is a flowchart illustrating the working method of a group of equipment maintenance robots that learn from each other's maintenance experience, as described in this invention. Detailed Implementation

[0024] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments. Example 1

[0025] like Figure 1As shown, the present invention proposes a group of equipment maintenance robots that learn maintenance experience from each other. The group includes a cloud server 101, at least three maintenance robots 102, an edge computing node 103, and a distributed sensor network 104. The cloud server 101 stores a global maintenance experience database, robot identity authentication information, and task scheduling data, and interacts with edge computing nodes through a hybrid communication protocol. The stored global maintenance experience database adopts a blockchain-database hybrid architecture. The blockchain layer stores consensus-verified maintenance case hash values ​​(including case ID, timestamp, robot signature, and the hash of the previous block), ensuring data immutability. The database layer stores data in shards according to device type, with each shard containing a fault feature vector, maintenance step sequence, effect evaluation matrix, and associated case index. Interacting with edge computing nodes through the hybrid communication protocol supports high-speed data transmission using 5G slicing technology in high-load scenarios, while also being compatible with the LoRaWAN protocol to reduce energy consumption under low load conditions. It is also responsible for verifying robot identity authentication information and storing it on the blockchain, as well as dynamic task scheduling based on the robot's real-time status (location, battery level, historical success rate).

[0026] like Figure 2 As shown, the maintenance robot 102 includes a multimodal data acquisition module 1021, a federated learning module 1022, a heterogeneous communication module 1023, a maintenance execution module 1024, and an energy management module 1025; wherein: The 1021 multimodal data acquisition module integrates a lidar, an infrared sensor, and a high-definition camera to collect vibration spectrum, temperature field distribution, surface defect images, and its own operating parameters from the target device. The lidar acquires three-dimensional point cloud data of the target device's surface, the infrared sensor collects thermal maps of the temperature distribution, and the high-definition camera captures images of surface defects. Simultaneously, it collects its own operating parameters, such as battery level and robotic arm operating status.

[0027] The Federated Learning Module 1022 employs a hierarchical training mechanism. Local nodes initially fuse maintenance experience based on a federated averaging algorithm, and then edge computing nodes achieve cross-regional experience aggregation. At the local node level, the maintenance experience collected by the robot itself is initially fused based on the federated averaging algorithm. Feature alignment units eliminate feature offsets caused by sensor differences, and local training units generate local policies using maintenance benefits as the reward function. At the edge node level, the local models of each robot are dynamically weighted and aggregated according to their historical maintenance success rates to complete cross-regional experience aggregation, reduce cloud load, and achieve efficient experience collaboration.

[0028] The heterogeneous communication module 1023 supports the IPv6 / 5G converged protocol and employs a security mechanism combining dynamic key encryption and identity authentication to achieve bidirectional data transmission with cloud servers, edge nodes, and other robots. For security, a unique device fingerprint is generated using the hardware serial number and IPv6 address, verified by the cloud server, and then written to the blockchain for authentication. It uses the SM4 national cryptographic algorithm, combined with timestamps, device fingerprints, and physical location to generate a dynamic session key, and incorporates random numbers for encryption to prevent replay attacks. In terms of communication, it supports the IPv6 / 5G converged protocol, enabling 5G slicing under high load and switching to LoRaWAN under low load, achieving secure bidirectional transmission with the cloud, edge nodes, and other robots.

[0029] The maintenance execution module 1024 includes a six-degree-of-freedom robotic arm and a tool library. Based on an optimized maintenance strategy, it performs operations not limited to disassembly, inspection, and replacement. In terms of hardware configuration, it integrates a six-degree-of-freedom robotic arm and a tool library. The robotic arm supports high-precision operation and adapts to complex equipment structures, while the tool library covers the necessary tools for disassembly, inspection, and replacement, encompassing various equipment maintenance scenarios. In terms of operational logic, it strictly follows the maintenance strategy optimized through federated learning, including component disassembly sequence and inspection parameter settings. Process data is fed back to edge nodes in real time for performance evaluation.

[0030] The energy management module 1025 dynamically adjusts the allocation of communication and computing resources through an energy consumption prediction model, and automatically requests charging scheduling when a low battery warning is triggered. Based on the energy consumption prediction model, it analyzes communication and computing resource demands in real time and dynamically allocates energy to reduce waste. When the battery level falls below a threshold, triggering a low battery warning, it automatically sends a charging scheduling request to the cloud server or edge node through the heterogeneous communication module.

[0031] The edge computing node 103 is deployed at the edge of the subnet in the robot's activity area. It serves as an industrial edge computing gateway or edge computing server, used for real-time verification of the validity of local maintenance experience and conflict resolution, reducing the communication load on the cloud server. Deployed at the edge of the subnet in the robot's activity area, its core responsibility is the real-time processing of local maintenance experience: verifying the validity of experience through a three-layer mechanism: format verification (checking required fields such as fault characteristics), consistency verification (blockchain-based historical case lookup logic), and effect verification (manual review of low-resolution strategies).

[0032] The distributed sensor network 104 consists of IoT sensors deployed on the target device, used to provide real-time feedback on the device's operating status and assist the maintenance robot in fault prediction. Composed of IoT sensors deployed on the target device, it collects real-time data on the device's operating status, such as vibration frequency, temperature distribution, and pressure fluctuations. This data serves as key input to the maintenance robot, assisting it in combining multimodal data collection to perform fault characteristic analysis and prediction.

[0033] In this embodiment, the global maintenance experience base adopts a blockchain-database hybrid architecture, including: The blockchain layer stores the hash values ​​of maintenance cases that have been verified through consensus, ensuring that the experience data is immutable. Each block contains a case ID, timestamp, robot signature, and the hash of the previous block. Specifically, it adopts a consortium blockchain architecture, and the consensus mechanism is the Practical Byzantine Fault Tolerance (PBFT) algorithm.

[0034] The database layer employs a sharded storage strategy, dividing data into shards based on equipment type. Each shard contains a fault feature vector, maintenance step sequence, effectiveness evaluation matrix, and related case index. Distributed storage is implemented based on MySQLCluster, dividing maintenance data for different types of equipment, such as motors, hydraulic equipment, and electrical systems, into independent primary shards. Inverted index technology links fault feature vectors with historical cases, supporting fast retrieval by "equipment type + fault feature." The effectiveness evaluation matrix covers quantitative indicators such as resolution rate, operation time, and consumable consumption, while the related case index records handling solutions for similar faults.

[0035] In this embodiment, the security mechanism of the heterogeneous communication module includes: Based on device fingerprinting and blockchain-based identity authentication, a unique device fingerprint is generated from the robot's hardware serial number and IPv6 address. After verification by the cloud server, this fingerprint is written to the blockchain for evidence storage. Specifically, the robot's hardware serial number (64-bit unique identifier) ​​and IPv6 address (128-bit network identifier) ​​are fused using the SHA-256 hash algorithm, and the first 64 bits are extracted as the unique device fingerprint. After the cloud server completes fingerprint verification through a blockchain smart contract, the fingerprint information, verification time, and robot ID are packaged and written to the blockchain for evidence storage.

[0036] Dynamic key generation employs the SM4 national cryptographic algorithm, combining the current timestamp, device fingerprints of both communicating parties, and physical location information to generate a session key. During key generation, the timestamp is accurate to the millisecond level to ensure timeliness. Physical location information is obtained through the robot's GPS module and, along with the device fingerprint, serves as the key seed input to the SM4 algorithm to generate the session key. A random number factor is introduced during the encryption process.

[0037] The communication range is adaptively adjusted, and the communication mode is dynamically switched according to the task load. 5G slicing technology is activated under high load, and LoRaWAN protocol is switched under low load. The task load is determined by the data transmission volume. When the data transmission volume of a single robot is greater than 10MB per minute, it is determined to be a high load, and 5G slicing technology is immediately activated to allocate a dedicated communication channel; when the transmission volume is less than 1MB / minute, it automatically switches to LoRaWAN protocol.

[0038] In this embodiment, the federated learning module includes: The feature alignment unit employs a contrastive learning algorithm to standardize fault features collected by different robots, eliminating feature offsets caused by differences in sensor models. Specifically, it utilizes the SimCLR contrastive learning framework. First, it performs data augmentation (including random rotation, scaling, and noise injection) on the raw fault data collected by multimodal sensors (LiDAR, infrared sensors, and high-definition cameras). Then, it uses a Siamese neural network to map the fault features of different robots to a unified high-dimensional feature space. By calculating a feature similarity loss function, it optimizes the model parameters, ultimately eliminating feature offsets caused by different sensor models.

[0039] The local training unit, based on an improved Deep Reinforcement Learning (DRL) algorithm, uses the reward function of maintenance operations to generate local maintenance policies. The reward function is designed as a combination of multi-dimensional quantitative indicators. The reward function is Reward = 0.7 * Fault resolution success rate + 0.2 * (1 - Actual consumable consumption / Theoretical consumable requirement) + 0.1 * (1 - Actual operation time), where the standard operation time is taken from the best case data of similar faults in the global experience base. Through continuous interaction with the equipment maintenance environment, the model continuously adjusts the maintenance step sequence (such as disassembly order and detection parameter settings) until it generates a local maintenance strategy that maximizes benefits.

[0040] The model aggregation unit performs weighted aggregation of the local models of each robot through edge nodes. The weights are dynamically adjusted based on the robot's historical maintenance success rate. The aggregation weight calculation formula is as follows: in Set the average success rate of maintenance tasks within a frequency range (value range 0-1) for the i-th robot. The normalized value (range 0-1) of the effective fault data contributed by the robot. After the edge nodes collect the local model parameters of each robot, they are weighted and averaged according to this weight to generate a regional maintenance model.

[0041] The strategy transfer unit employs a meta-learning algorithm to quickly adapt maintenance experience from other scenarios to local equipment types, generating emergency response plans for new faults. Specifically, it is implemented based on the MAML (Model Independent Meta-Learning) algorithm. For example, it first selects cross-regional maintenance cases with a similarity ≥80% to local equipment types (such as motors and hydraulic systems) from the global experience base to construct an experience support set. The model is then trained on this support set to initialize parameters, enabling rapid adaptation. When a new fault is encountered, only a small amount of local fault data is needed to fine-tune the model. This fine-tuning process uses an early-stop strategy. Finally, the new strategy is verified through Monte Carlo tree search simulations. For instance, only solutions with a simulation success rate ≥85% can be deployed as emergency response plans. Example 2

[0042] like Figure 3 As shown, a method for a group of equipment maintenance robots to learn from each other's maintenance experience, applied to the system described in Example 1, includes the following steps: Step S1: Multimodal data acquisition and preprocessing The maintenance robot obtains three-dimensional point cloud data by scanning the surface of the target equipment with LiDAR, collects temperature distribution heat maps with infrared sensors, and captures defect images with high-definition cameras. After removing noise from the data through wavelet transform, fault feature parameters, including but not limited to vibration frequency peak, temperature gradient, and defect area, are extracted and encapsulated into feature vectors in JSON format. Step S2: Local experience generation and encrypted upload The local training unit of the federated learning module matches historical cases based on feature vectors. If the matching degree is greater than the threshold, the historical policy is directly called; otherwise, DRL training is started to generate a new maintenance policy and calculate the effect evaluation value. The heterogeneous communication module adopts a dynamic key encryption strategy to upload the new policy to the edge computing node and receive the shared experience summary of robots in the same area. Step S3: Experience verification and aggregation of edge nodes Edge nodes perform three-layer verification of the received maintenance experience: Format validation: Check if the required fields for fault characteristics, operation steps, and effect evaluation are included; Consistency verification: Verify the logical consistency between the new strategy and similar cases by querying historical cases through the blockchain; Effect verification: If the estimated resolution rate of the new strategy is less than the threshold, it will be pushed to the senior maintenance robot for review; After verification, the edge nodes use a federated averaging algorithm to aggregate the maintenance models of local robots, generate regional maintenance strategies, and synchronize them to the cloud server. Step S4: Global Experience Collaborative Learning and Strategy Optimization The cloud server integrates regional-level strategies with a global experience base and uses a similarity algorithm to match cross-regional related experiences for each robot. After the maintenance robot receives the global experience, the policy migration unit generates an optimized strategy adapted to the local scenario based on meta-learning, updates the local model, and feeds back the optimization effect to the cloud server. Step S5: Dynamic Task Allocation and Maintenance Execution When the distributed sensor network detects a device malfunction, the cloud server allocates tasks based on the robot's real-time location, battery level, and historical success rate. Using an improved auction algorithm, robots bid based on their own load, and the winning bidder obtains the maintenance rights. The maintenance robot invokes the optimized strategy and performs operations through the maintenance execution module. Process data is uploaded to the cloud server in real time via edge nodes. After a task is completed, the system updates the robot's reputation value based on the actual resolution effect. Robots with reputation values ​​below the threshold are restricted from receiving high-priority tasks.

[0043] In this embodiment, the process of extracting the fault feature parameters in step S1 includes: S11: Vibration feature extraction: Perform Fourier transform on the original vibration signal and extract the peak frequency and corresponding amplitude; S12: Temperature Feature Extraction: Extract the highest temperature, temperature standard deviation, hot spot area, and temperature change rate from the heat map; S13: Visual feature extraction: Use the YOLOv8 algorithm to identify defect types and calculate the defect area, perimeter, and circumscribed rectangle dimensions; S14: Normalize the above features to the [0,1] interval to form a feature vector, in which vibration features account for 40% of the weight, temperature features account for 30%, and visual features account for 30%.

[0044] In this embodiment, the implementation process of dynamic key encryption in step S2 includes: S21: The sending robot generates a device fingerprint: the hardware serial number + IPv6 address are hashed using SHA-256, and the first 64 bits are taken as the device fingerprint; S22: Generate an initial key by combining the current timestamp and the physical location coordinates of the receiver; S23: Encrypt the empirical data using the SM4 algorithm, and introduce random numbers during the encryption process to prevent replay attacks; S24: When using 5G communication, network slicing is enabled to allocate dedicated channels for experience transmission. When the channel quality index is lower than the set threshold, it automatically switches to the TCP retransmission mechanism.

[0045] In this embodiment, the aggregation process of the federated average algorithm in step S3 includes: S31: Edge nodes collect local model parameters and reputation values ​​for each robot and calculate weighting coefficients. It is a quantitative assessment of the overall reliability, accuracy, and compliance capabilities demonstrated by the i-th robot during task execution; This refers to the total amount of data related to model training collected and processed by the i-th robot within a certain period; S32: Calculate a weighted average of the model parameters to generate a regional model. in, For local model parameters, These are parameters for the regional model. S33: Differential privacy technology is used to add Gaussian noise to protect the local data privacy of each robot; S34: Compare the aggregated model with the global experience base. If the difference exceeds the threshold, trigger a second aggregation verification.

[0046] In this embodiment, the meta-learning adaptation process in step S4 includes: S41: Select cases similar to local device types from global experience and build a support set; S42: Based on the MAML algorithm, the model is trained on the support set to initialize parameters, enabling the model to converge quickly with a small amount of local data; S43: Use local fault data to fine-tune the initialization model and generate a maintenance strategy adapted to the local scenario. An early stop strategy is adopted during the fine-tuning process. S44: The new strategy is simulated and verified through Monte Carlo tree search. If the simulation success rate is greater than the threshold, it is put into actual use.

[0047] This embodiment also includes an experience iteration and self-healing step: S51: The cloud server iterates and optimizes the global experience database weekly, calculating the comprehensive score for each experience: Score = 0.6 × resolution rate + 0.2 × (1 - time spent / average time spent) + 0.2 × (1 - material consumption / average consumption) S52: For experiences with a score <60, a group revision mechanism is triggered, consisting of a revision group of 3 or more robots with a reputation score ≥90, which generates an iterative version through federated learning; S53: When the robot detects a fault in its own sensor, it automatically sends a self-healing request to the edge node. The edge node schedules redundant data from neighboring robots to compensate for the fault and generates a sensor maintenance reminder. Example 3

[0048] Taking the preventive maintenance scenario of a cluster of motor equipment in a smart manufacturing plant as an example, the workflow of the robot swarm is explained in detail based on the principles of this invention: I. Equipment Anomaly Detection and Data Acquisition A distributed sensor network (vibration sensors, infrared temperature sensors) deployed in the factory workshop monitors the operating status of three motors in real time. When the sensors detect that the vibration frequency of motor No. 1 suddenly increases to 50Hz (normal range 20-30Hz) and the casing temperature reaches 75℃ (threshold 65℃), the abnormal signal is immediately pushed to the cloud server.

[0049] After the cloud server triggers a task alert, three nearby maintenance robots (A, B, and C) respond to the command and conduct detailed inspections through a multimodal data acquisition module: LiDAR scans the surface of the motor to obtain three-dimensional point cloud data and identify signs of loose screws on the casing. Infrared sensors generate temperature thermal maps, identifying the high-temperature hotspots at the positioning bearing area; A high-definition camera captured an image of the bearing end cover, and the YOLOv8 algorithm identified a minor oil leakage defect.

[0050] After the data is denoised by wavelet transform, features such as vibration peak frequency (50Hz), temperature gradient (10℃ / min), and defect size are extracted and normalized to form a feature vector (vibration 40% + temperature 30% + visual 30%).

[0051] II. Local Experience Generation and Secure Upload The federated learning module of the three robots initiates the local policy generation process: Robot A queries local historical cases and finds that it matches the case of "abnormal noise and high temperature of motor bearing" by 85% (threshold 80%), and directly calls the historical strategy (replace bearing grease + tighten screws). Robots B and C did not find highly similar cases, so improved DRL training was initiated: a new strategy was generated that includes "bearing disassembly sequence, grease type selection, and torque parameter settings" with "fault resolution rate (70%) + consumable consumption (20%) + time consumption (10%)" as the reward function.

[0052] After the strategy is generated, the heterogeneous communication module starts encrypted transmission: the robot hardware serial number + IPv6 address are hashed using SHA-256 to generate a device fingerprint, and an initial key is generated by combining the timestamp and edge node location; the strategy data is encrypted using the SM4 national cryptographic algorithm, and random numbers are introduced to prevent replay attacks. Due to the transmission volume reaching 12MB / minute (high load), 5G slicing technology is used to allocate a dedicated channel for uploading to the edge node.

[0053] III. Edge Validation and Region Model Aggregation The edge computing node performs three-layer verification on the three received policy documents: Format validation: All checks include required fields such as fault characteristics, operation steps, and effect evaluation; Consistency verification: Historical cases are queried through the blockchain to confirm that the strategy is consistent with the logic of the "Standard Procedure for Motor Bearing Maintenance"; Results verification: The estimated resolution rate of robot strategies B and C reached 92% (threshold 85%), requiring no further review.

[0054] After verification, the edge nodes aggregate the model using the federated average algorithm: dynamically allocate weights (A: 40%, B: 35%, C: 25%) based on the robot's historical success rate (A: 90%, B: 85%, C: 80%), generate regional maintenance strategies (optimize bearing disassembly sequence, shorten time by 15%), add Gaussian noise to protect privacy, and then synchronize to the cloud server.

[0055] IV. Global Optimization and Task Execution The cloud server integrates regional strategies with a global experience base, and uses a similarity algorithm to match cross-factory cases of "high-temperature failure of the same model motor," discovering that the key step of "bearing clearance adjustment" was not covered. After receiving the global experience, the robot's strategy transfer unit uses the MAML algorithm: it selects a similar support set from global cases, trains and initializes parameters, fine-tunes the model using local motor data, generates an optimized strategy that includes "clearance adjustment," and updates the local model after verification by Monte Carlo tree simulation.

[0056] The cloud server allocates tasks based on the robots' real-time status (A: 80% battery, 10m distance, 90% success rate; B: 60% battery, 15m distance; C: 70% battery, 20m distance) using an improved auction algorithm, and robot A wins the bid. A invokes the optimized strategy and performs the following operations using its six-DOF robotic arm: disassembling the end cap, checking clearance, replacing grease, and tightening screws. Process data is uploaded in real-time via edge nodes. After the task is completed, the motor vibration drops to 25Hz and the temperature reaches 60℃, and the system updates A's reputation value to 92%.

[0057] V. Dynamic Iteration and Self-Healing Guarantee Weekly cloud server performance evaluation: This strategy achieved a 100% resolution rate, took 20% less time than the average, and had no material waste, resulting in a comprehensive score of 96 points, which was then stored in the global database. One month later, robot C's sensor malfunctioned, automatically requesting self-healing from the edge node.

[0058] In summary, this invention constructs a collaborative "cloud-edge-device-network" system consisting of a cloud server, multiple maintenance robots, edge computing nodes, and a distributed sensor network. This system enables cross-regional learning of maintenance experience among robot groups, significantly improving the intelligence level and overall efficiency of industrial equipment maintenance. The cloud server, acting as the global hub, stores a global maintenance experience database using a blockchain-database hybrid architecture. The blockchain layer ensures data immutability, while the database layer enables efficient retrieval by sharding data according to equipment type. It also handles robot authentication and task scheduling. The maintenance robots integrate core modules such as multimodal data acquisition, federated learning, and heterogeneous communication, forming a complete closed loop from data perception to maintenance execution. Edge computing nodes, deployed at the subnet edge, handle local experience processing and load optimization. The distributed sensor network provides real-time feedback on equipment status, offering data support for fault prediction.

[0059] The workflow of this invention includes key steps such as data acquisition and preprocessing, local experience generation and encrypted uploading, edge verification and aggregation, global collaborative optimization, and dynamic task allocation and execution. After denoising and feature extraction, multimodal data is used by the federated learning module to generate local policies and uploaded in encrypted form. Edge node verification and aggregation generate regional policies that are synchronized to the cloud. The cloud then integrates global experience and adapts to the local scenario through meta-learning. Task allocation is combined with dynamic scheduling based on the robot's real-time status, and execution process data is fed back in real time. The system also possesses experience iteration and self-healing capabilities, regularly optimizing the global experience library and revising low-quality experiences, and compensating for redundant data during sensor failures. Through intelligent design across the entire process, it effectively overcomes the limitations of traditional maintenance modes, significantly improving fault response speed, maintenance accuracy, and task continuity.

[0060] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A group of equipment maintenance robots that learn from each other's maintenance experience, characterized in that, Includes cloud servers, at least three maintenance robots, edge computing nodes, and a distributed sensor network; The cloud server is used to store a global maintenance experience base, robot identity authentication information and task scheduling data, and realizes data interaction with edge computing nodes through a hybrid communication protocol; The maintenance robot includes a multimodal data acquisition module, a federated learning module, a heterogeneous communication module, a maintenance execution module, and an energy management module; wherein: The multimodal data acquisition module integrates lidar, infrared sensors, and high-definition cameras to acquire vibration spectrum, temperature field distribution, surface defect images, and operating parameters of the target equipment. The federated learning module adopts a hierarchical training mechanism. Local nodes perform initial fusion of maintenance experience based on the federated averaging algorithm, and then cross-regional experience aggregation is achieved through edge computing nodes. The heterogeneous communication module supports the IPv6 / 5G converged protocol and adopts a security mechanism that combines dynamic key encryption and identity authentication to achieve bidirectional data transmission with cloud servers, edge nodes and other robots; The maintenance execution module includes a six-degree-of-freedom robotic arm and a tool library, which performs operations, including but not limited to disassembly, inspection, and replacement, based on the optimized maintenance strategy. The energy management module dynamically adjusts the allocation of communication and computing resources through an energy consumption prediction model, and automatically requests charging scheduling when a low power warning is triggered. The edge computing nodes are deployed at the edge of the subnet in the robot's activity area to process the validity verification of local maintenance experience and resolve conflicts in real time, thereby reducing the communication load on the cloud server. The distributed sensor network consists of IoT sensors deployed on the target device, used to provide real-time feedback on the device's operating status and assist the maintenance robot in predicting faults.

2. The equipment maintenance robot group that learns maintenance experience from each other according to claim 1, characterized in that, The global maintenance experience base adopts a blockchain-database hybrid architecture, including: The blockchain layer stores the hash values ​​of maintenance cases that have been verified through consensus, ensuring that the experience data is immutable. Each block contains a case ID, timestamp, robot signature, and the hash of the previous block. The database layer adopts a sharded storage strategy, dividing the data into shards according to the device type. Each shard contains a fault feature vector, a maintenance step sequence, an effect evaluation matrix, and a related case index.

3. The equipment maintenance robot group that learns maintenance experience from each other according to claim 1, characterized in that, The security mechanism of the heterogeneous communication module includes: Based on device fingerprint and blockchain identity authentication, the robot hardware serial number and IPv6 address generate a unique device fingerprint, which is then verified by the cloud server and written into the blockchain for storage. Dynamic key generation uses the SM4 national cryptographic algorithm and combines the current timestamp, device fingerprints of both communicating parties, and physical location information to generate a session key. The communication range is adaptively adjusted, and the communication mode is dynamically switched according to the task load. 5G slicing technology is enabled when the load is high, and LoRaWAN protocol is switched when the load is low.

4. The equipment maintenance robot group that learns maintenance experience from each other according to claim 1, characterized in that, The federated learning module includes: The feature alignment unit uses a contrastive learning algorithm to standardize the fault features collected by different robots, eliminating feature offsets caused by differences in sensor models. The local training unit, based on the improved deep reinforcement learning (DRL) algorithm, uses the reward function of maintenance operations to generate local maintenance policies. The model aggregation unit performs weighted aggregation of the local models of each robot through edge nodes, and the weights are dynamically adjusted according to the robot's historical maintenance success rate. The strategy migration unit uses a meta-learning algorithm to quickly adapt maintenance experience from other scenarios to local device types and generate emergency handling plans for new faults.

5. A method for operating a group of equipment maintenance robots that learn maintenance experience from each other as described in any one of claims 1-4, characterized in that, Includes the following steps: Step S1: Multimodal data acquisition and preprocessing The maintenance robot obtains three-dimensional point cloud data by scanning the surface of the target equipment with LiDAR, collects temperature distribution heat maps with infrared sensors, and captures defect images with high-definition cameras. After removing noise from the data through wavelet transform, fault feature parameters, including but not limited to vibration frequency peak, temperature gradient, and defect area, are extracted and encapsulated into feature vectors in JSON format. Step S2: Local experience generation and encrypted upload The local training unit of the federated learning module matches historical cases based on feature vectors. If the matching degree is greater than the threshold, the historical policy is directly called; otherwise, DRL training is started to generate a new maintenance policy and calculate the effect evaluation value. The heterogeneous communication module adopts a dynamic key encryption strategy to upload the new policy to the edge computing node and receive the shared experience summary of robots in the same area. Step S3: Experience verification and aggregation of edge nodes Edge nodes perform three-layer verification of the received maintenance experience: Format validation: Check if the required fields for fault characteristics, operation steps, and effect evaluation are included; Consistency verification: Verify the logical consistency between the new strategy and similar cases by querying historical cases through the blockchain; Effect verification: If the estimated resolution rate of the new strategy is less than the threshold, it will be pushed to the senior maintenance robot for review; After verification, the edge nodes use a federated averaging algorithm to aggregate the maintenance models of local robots, generate regional maintenance strategies, and synchronize them to the cloud server. Step S4: Global Experience Collaborative Learning and Strategy Optimization The cloud server integrates regional-level strategies with a global experience base and uses a similarity algorithm to match cross-regional related experiences for each robot. After the maintenance robot receives the global experience, the policy migration unit generates an optimized strategy adapted to the local scenario based on meta-learning, updates the local model, and feeds back the optimization effect to the cloud server. Step S5: Dynamic Task Allocation and Maintenance Execution When the distributed sensor network detects a device malfunction, the cloud server allocates tasks based on the robot's real-time location, battery level, and historical success rate. Using an improved auction algorithm, robots bid based on their own load, and the winning bidder obtains the maintenance rights. The maintenance robot invokes the optimized strategy and performs operations through the maintenance execution module. Process data is uploaded to the cloud server in real time via edge nodes. After a task is completed, the system updates the robot's reputation value based on the actual resolution effect. Robots with reputation values ​​below the threshold are restricted from receiving high-priority tasks.

6. The working method of the equipment maintenance robot group that learns maintenance experience from each other according to claim 5, characterized in that, The process of extracting fault feature parameters in step S1 includes: S11: Vibration feature extraction: Perform Fourier transform on the original vibration signal and extract the peak frequency and corresponding amplitude; S12: Temperature Feature Extraction: Extract the highest temperature, temperature standard deviation, hot spot area, and temperature change rate from the heat map; S13: Visual feature extraction: Use the YOLOv8 algorithm to identify defect types and calculate the defect area, perimeter, and circumscribed rectangle dimensions; S14: Normalize the above features to the [0,1] interval to form a feature vector, in which vibration features account for 40% of the weight, temperature features account for 30%, and visual features account for 30%.

7. The working method of the equipment maintenance robot group that learns maintenance experience from each other according to claim 1, characterized in that, The implementation process of dynamic key encryption in step S2 includes: S21: The sending robot generates a device fingerprint: the hardware serial number + IPv6 address are hashed using SHA-256, and the first 64 bits are taken as the device fingerprint; S22: Generate an initial key by combining the current timestamp and the physical location coordinates of the receiver; S23: Encrypt the empirical data using the SM4 algorithm, and introduce random numbers during the encryption process to prevent replay attacks; S24: When using 5G communication, network slicing is enabled to allocate dedicated channels for experience transmission. When the channel quality index is lower than the set threshold, it automatically switches to the TCP retransmission mechanism.

8. The working method of the equipment maintenance robot group that learns maintenance experience from each other according to claim 1, characterized in that, The aggregation process of the federated average algorithm described in step S3 includes: S31: Edge nodes collect local model parameters and reputation values ​​for each robot and calculate weighting coefficients. It is a quantitative assessment of the overall reliability, accuracy, and compliance capabilities demonstrated by the i-th robot during task execution; This refers to the total amount of data related to model training collected and processed by the i-th robot within a certain period; S32: Calculate a weighted average of the model parameters to generate a regional model. in, For local model parameters, These are parameters for the regional model. S33: Differential privacy technology is used to add Gaussian noise to protect the local data privacy of each robot; S34: Compare the aggregated model with the global experience base. If the difference exceeds the threshold, trigger a second aggregation verification.

9. The working method of the equipment maintenance robot group that learns maintenance experience from each other according to claim 1, characterized in that, The meta-learning adaptation process described in step S4 includes: S41: Select cases similar to local device types from global experience and build a support set; S42: Based on the MAML algorithm, the model is trained on the support set to initialize parameters, enabling the model to converge quickly with a small amount of local data; S43: Use local fault data to fine-tune the initialization model and generate a maintenance strategy adapted to the local scenario. An early stop strategy is adopted during the fine-tuning process. S44: The new strategy is simulated and verified through Monte Carlo tree search. If the simulation success rate is greater than the threshold, it is put into actual use.

10. The working method of the equipment maintenance robot group that learns maintenance experience from each other according to claim 1, characterized in that, It also includes experience iteration and self-healing steps: S51: The cloud server iterates and optimizes the global experience database weekly, calculating the comprehensive score for each experience: Score = 0.6 × resolution rate + 0.2 × (1 - time spent / average time spent) + 0.2 × (1 - material consumption / average consumption) S52: For experiences with a score <60, a group revision mechanism is triggered, consisting of a revision group of 3 or more robots with a reputation score ≥90, which generates an iterative version through federated learning; S53: When the robot detects a fault in its own sensor, it automatically sends a self-healing request to the edge node. The edge node schedules redundant data from neighboring robots to compensate for the fault and generates a sensor maintenance reminder.

Images