Self-adaptive cooperative control method and system for industrial equipment

Abstract

The invention relates to a self-adaptive cooperative control method and system for industrial equipment, and the method comprises the steps: collecting and standardizing the operation data of multiple equipment into a time sequence event stream in real time; performing time sequence alignment on the multi-device action sequence, eliminating phase deviation, and extracting statistical features to construct a cross-device feature library; using a multi-task model to decouple cross-device sharing features and device exclusive features; dynamically fusing the shared features and the exclusive features, extending the shared features to newly added equipment in combination with transfer learning, and extracting the exclusive features to adapt to hardware features; deploying a lightweight inference model at an edge end, loading cloud shared base model parameters and local exclusive parameters, and performing real-time fusion to generate a control instruction; the cloud end aggregates the data, optimizes a global sharing model through incremental learning, and encrypts and synchronously updates parameters to the edge end in a sub-coding manner; and integrating the feature library, the model architecture and the edge cloud collaboration mechanism, generating a global collaboration strategy, and realizing multi-device adaptive collaboration control and rapid expansion.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent manufacturing, and particularly relates to an industrial equipment adaptive collaborative control method and system. Background Art

[0002] In modern industrial production, with the continuous expansion of production scale and the increasing complexity of production processes, various types and brands of industrial equipment are often deployed within a factory. These devices are from different manufacturers, and not only have different operating characteristics and functions, such as the diversity of machine operating states, the complexity of robotic arm movement patterns, the differences in peripheral device interaction logics, etc., but also the communication protocols and interface specifications vary greatly. Traditional industrial equipment control methods are usually designed for single devices or a small number of specific device combinations, and it is difficult to meet the requirements of multi-type and multi-brand devices for collaborative operation in complex production scenarios. In complex production processes, there are deviations in the action timings between different devices, resulting in low collaborative operation efficiency. At the same time, in the face of the access of new devices, traditional methods lack effective adaptive mechanisms and cannot quickly integrate them into the existing collaborative system, making it difficult to guarantee the collaborative accuracy between devices, seriously affecting production efficiency and product quality. Therefore, there is an urgent need for a technology that can achieve high-precision and adaptive collaborative execution of multi-type and multi-brand industrial equipment in complex scenarios. Summary of the Invention

[0003] The main object of the present invention is to provide an industrial equipment adaptive collaborative control method and system to achieve the purpose of adaptive collaborative control and rapid expansion of multi-type industrial equipment in a dynamic production environment.

[0004] To achieve the above object, the present invention provides an industrial equipment adaptive collaborative control method, including the following steps:

[0005] Real-time collect multi-device operation data, and convert the multi-device operation data into a standardized time-series event stream;

[0006] Perform time-series alignment of multi-device action sequences on the time-series event stream, and extract key time features to construct a cross-device feature library;

[0007] Extract high-level semantic features from the cross-device feature library, and decouple them into cross-device shared features and device-specific features through a pre-trained multi-task learning model;

[0008] Adaptively fuse the shared features and the specific features through an attention mechanism, optimize the device control accuracy using the specific features, extend the shared features to new devices through transfer learning, and extract new device-specific features from the real-time operation data of the new devices through domain adaptive training;

[0009] Based on the edge-cloud collaborative architecture, a lightweight inference model deployed at the edge of the new device loads the exclusive feature parameters of the new device and the shared feature base model parameters sent from the cloud respectively. The new exclusive feature parameters are used to adapt to the hardware characteristics of the new device and dynamically fuse with the shared features during inference to optimize the accuracy of control instructions. The shared feature base model is generated by compressing the global shared feature model through knowledge distillation, which can adapt to the edge computing power and perform low-latency instruction control for the new device;

[0010] Aggregate the operation data of multiple devices through the cloud, optimize the global shared feature model based on incremental learning, and synchronize the updated shared feature parameters to the edge of each device;

[0011] Integrate the time series feature library, the multi-task model, and the edge-cloud collaborative architecture, generate a collaborative strategy through the global shared features, and drive multi-type industrial devices to cooperate under a unified framework.

[0012] Further, the step of collecting the operation data of multiple devices in real time and converting the operation data of multiple devices into a standardized time series event stream includes:

[0013] Collect the operation data of each device through sensors, including the operation status data of the machine tool, the motion data of the robotic arm, the interaction data of peripheral devices, and the logic index data;

[0014] Convert the collected data into a time series event stream in a unified format according to a preset time window. Each event includes a device identifier, a timestamp, a data value, and a status identifier.

[0015] Further, the step of performing time series alignment on the multi-device action sequences of the time series event stream and extracting key time features to construct a cross-device feature library includes:

[0016] Based on the time series event stream, eliminate the phase deviation of the action time series between devices through the dynamic time warping algorithm, and align the time series data of each device;

[0017] Based on the aligned time series data, extract the statistical features of each action sequence through a sliding window, including mean, variance, and frequency domain features;

[0018] Associate and store the statistical features with the device type, construct a cross-device feature library, and establish a three-level index of device type, action category, and timestamp.

[0019] Further, the step of decoupling into shared features and device-exclusive features through a pre-trained multi-task learning model includes:

[0020] Encoding the time series data of the cross-device feature library through a multi-task model, and extracting the collaborative patterns and abnormal correlation features of each device as high-level semantic features;

[0021] Using a shared feature branch to extract cross-device general parameters, and extracting device-specific parameters of each device through an exclusive feature branch;

[0022] Through a decoupled loss function, constraining the orthogonality of the outputs of the shared feature branch and the exclusive feature branch to obtain the shared features of multiple devices and the exclusive features of each device.

[0023] Further, the step of adaptively fusing the shared features and the exclusive features through an attention mechanism includes:

[0024] Taking the shared features as the reference input and the exclusive features as the conditional input, calculating dynamic fusion weights through a multi-head attention mechanism and fusing them;

[0025] Through domain adversarial training, minimizing the spatial distribution difference between the exclusive features of the new device and the shared features to perform feature space alignment;

[0026] Updating the exclusive feature parameters of the new device based on the real-time operation data of the new device.

[0027] Further, the step of deploying a lightweight inference model at the edge of the new device includes:

[0028] Deploying an inference model framework at the edge of the new device and loading the shared feature base model parameters sent from the cloud;

[0029] Synchronously loading the exclusive feature parameters of the new device stored at the edge, and the exclusive feature parameters of the new device are generated through domain adaptation training and stored locally;

[0030] Parsing the control instructions of the new device through a real-time scheduling algorithm, and dynamically adjusting the motion trajectory of the robotic arm or the force of the end effector of the new device in combination with the loaded exclusive feature parameters of the new device.

[0031] Further, the step of aggregating the operation data of multiple devices through the cloud, optimizing the global shared feature model based on incremental learning, and synchronizing the updated shared feature parameters to the edge of each device includes:

[0032] Aggregating the real-time operation status data and historical log data of multiple devices, and updating the global shared feature model parameters using an online incremental learning algorithm;

[0033] Compressing the updated global shared feature model through knowledge distillation technology, retaining the parameters of the cross-device shared feature layer, and removing the device-specific feature-related nodes and redundant computing units;

[0034] Compress the model through knowledge distillation, retain the parameters of the shared layer and eliminate redundant nodes;

[0035] Differentially encode the change amount of the compressed shared feature parameters, and transmit them to the edge of each device through an encryption protocol. The edge terminal verifies and loads the updated parameters according to the version number.

[0036] Further, the step of aggregating the multi-device operation data through the cloud, optimizing the global shared feature model based on incremental learning, and synchronizing the updated shared feature parameters to the edge of each device further includes:

[0037] During the process of optimizing by incremental learning, calculate the historical update amplitude of the shared feature parameters, and reduce the learning rate for the core parameters that are frequently updated to prevent key parameters from being overwritten;

[0038] The lightweight inference model reduces the memory occupancy at the edge by pruning the neuron connections related to the device-specific features in the shared feature base model and compressing the floating-point parameters into 8-bit integers.

[0039] When a new device joins the collaborative operation, calculate the instruction execution delay based on the device-specific feature parameters of the new device. If the delay of the new device exceeds 20% of the average delay of the current online devices, the priority of the new device in the collaboration time sequence is dynamically reduced.

[0040] Further, the step of driving multi-type industrial devices to perform collaborative operations under a unified framework includes:

[0041] Generate the device collaborative start / stop time sequence and priority allocation rules based on the global shared feature model;

[0042] Combined with the device-specific feature parameters of each device, send a collaborative instruction set including the robotic arm trajectory coordinates, conveyor belt speed adjustment instructions, and machine start / stop thresholds to the edge;

[0043] Real-time monitor the execution status of each device. When detecting conflicts in the robotic arm trajectory coordinates, the conveyor belt speed exceeding the safety threshold, or the machine operation load exceeding the limit, trigger an exception handling protocol to recalculate the collaboration strategy.

[0044] The present invention also provides an industrial device adaptive collaboration control system, including:

[0045] A data acquisition unit for real-time collecting multi-device operation data and converting the multi-device operation data into a standardized time series event stream;

[0046] A database construction unit for performing time series alignment of the multi-device action sequences of the time series event stream and extracting key time features to construct a cross-device feature library;

[0047] A feature extraction unit, used to extract high-level semantic features from the cross-device feature library, and decouple them into cross-device shared features and device-specific features through a pre-trained multi-task learning model;

[0048] A feature fusion unit, configured to adaptively fuse shared features with exclusive features through an attention mechanism, optimize device control accuracy using the exclusive features, extend the shared features to newly added devices through transfer learning, and extract exclusive features of the newly added devices from real-time operating data of the newly added devices through domain adaptive training;

[0049] A data optimization unit is used to load the newly added device-specific feature parameters and the shared feature base model parameters issued by the cloud into a lightweight inference model deployed at the edge of the newly added device based on an edge-cloud collaborative architecture;

[0050] A parameter synchronization unit, configured to aggregate the multi-device operation data through the cloud, optimize the global shared feature model based on incremental learning, and synchronize the updated shared feature parameters to the edge of each device;

[0051] An execution unit is used to integrate the timing feature library, the multi-task model and the edge-cloud collaborative architecture, generate a collaborative strategy through global shared features, and drive multiple types of industrial equipment to work collaboratively under a unified framework.

[0052] The industrial equipment adaptive collaborative control method and system provided by the present invention have the following beneficial effects:

[0053] The present invention collects multi-device operation data in real time and converts it into a standardized time series event stream, providing a unified and standardized data basis for subsequent processing, improving the efficiency and accuracy of data processing; the time series event stream is aligned with the time series of multi-device action sequences, and a cross-device feature library is constructed, which is helpful to deeply explore the collaborative mode and abnormal correlation features between devices, and provide strong support for device collaborative control. Afterwards, the pre-trained multi-task learning model is used to decouple the shared features and device-specific features, and adaptively fuse them through the attention mechanism, which not only optimizes the device control accuracy, but also can flexibly extend the shared features to new devices, thereby enhancing the adaptability of the system. Under the edge-cloud collaborative architecture, a lightweight inference model is deployed to achieve low-latency command control of new devices and optimize the control command accuracy. In addition, the cloud optimizes the global shared feature model based on incremental learning and synchronizes parameters to ensure the timeliness and accuracy of the model. In addition, a collaborative strategy is generated to drive multiple types of industrial equipment to work together under a unified framework, which improves the overall efficiency and stability of industrial production, effectively reduces the probability of conflicts and abnormal situations between devices, and comprehensively improves the intelligent level of industrial production. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1It is a schematic flowchart of the industrial equipment adaptive collaborative control method in an embodiment of the present invention;

[0055] Figure 2 It is a structural block diagram of the industrial equipment adaptive collaborative control system in an embodiment of the present invention;

[0056] The realization of the purpose of the present invention, functional characteristics and advantages will be further described with reference to the embodiments and the accompanying drawings. Specific embodiments

[0057] In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0058] Referring to Figure 1 , it is a schematic flowchart of an industrial equipment adaptive collaborative control method proposed by the present invention, including the following steps:

[0059] S1, Real-time collect the operation data of multiple devices, and convert the operation data of the multiple devices into a standardized time series event stream;

[0060] S2, Perform time series alignment of the action sequences of multiple devices on the time series event stream, and extract key time features to construct a cross-device feature library;

[0061] S3, Extract high-level semantic features from the cross-device feature library, and decouple them into cross-device shared features and device-specific features through a pre-trained multi-task learning model;

[0062] S4, Adaptively fuse the shared features and the specific features through the attention mechanism, optimize the device control accuracy using the specific features, extend the shared features to new devices through transfer learning, and extract new device-specific features from the real-time operation data of the new devices through domain adaptation training;

[0063] S5, Based on the edge-cloud collaborative architecture, deploy a lightweight inference model at the edge of the new device, and load the new device-specific feature parameters and the shared feature base model parameters sent from the cloud respectively. The new specific feature parameters are used to adapt to the hardware characteristics of the new device, and are dynamically fused with the shared features during inference to optimize the control instruction accuracy. The shared feature base model is generated by compressing the global shared feature model through knowledge distillation, and can adapt to the edge computing power for low-latency instruction control of new devices;

[0064] S6, Aggregate the operation data of the multiple devices through the cloud, optimize the global shared feature model based on incremental learning, and synchronize the updated shared feature parameters to the edge of each device;

[0065] S7. Integrate the timing feature library, the multi-task model, and the edge-cloud collaborative architecture, generate a collaborative strategy through globally shared features, and drive multi-type industrial devices to collaborate under a unified framework.

[0066] As described in step S1 above, collect heterogeneous data. Through built-in device sensors (such as encoders and torque sensors) and external IoT modules, collect heterogeneous data such as the operating state of the machine tool (such as temperature and vibration), the motion data of the robotic arm (joint angles and end coordinates), the interaction data of peripheral devices (such as conveyor belt speed and material arrival signal), and logical indicators (PLC instruction status). Slice the data according to a preset time window (such as 100 ms) and encapsulate it into a structured event unit of [device ID, timestamp, data value, status code]. For example, the format of the robotic arm event stream is [Robot_A, t = 162000.5 ms, joint_angle = [30°, 45°], status = 0x01]. For different device protocols (such as Modbus and OPC UA), map the raw data to standard fields through a protocol conversion middleware (such as an industrial gateway) to ensure data semantic consistency.

[0067] As described in step S2 above, use the dynamic time warping (DTW) algorithm to align the action sequences of multiple devices. For example, align the grasping action of robotic arm A with the material arrival signal of conveyor belt B to eliminate the phase deviation caused by mechanical delay or communication jitter. Calculate statistical features (mean, variance, kurtosis) and frequency domain features (FFT main frequency, energy spectral density) within a sliding window (such as a 5-second window), and store them in association with the device type (such as robotic arm and CNC machine tool) to construct a cross-device feature library with a three-level index (device type → action category → timestamp), which supports quickly retrieving the features of different devices at specific action stages (such as start, processing, and stop). Through S2, eliminate the asynchrony of the action timings between devices and extract common features that can be analyzed across devices.

[0068] As described in step S3 above, the pre-trained multi-task learning model uses a Transformer encoder. The aligned time-series data is input, and two types of features are output: shared features, which are cross-device general patterns (such as coordinated start-stop time series and energy consumption patterns); and exclusive features, which are single-device hardware characteristics (such as the friction coefficient of robotic arm joints and motor response delay). The decoupled loss function (such as orthogonal constraint loss) is used to ensure the orthogonality of the shared features and exclusive features in the vector space, avoiding feature mixing. The high-level semantic features include the collaborative patterns of each device and the abnormal associations between devices (such as a device overheating causing another device to slow down). The long-range dependencies are captured through the self-attention mechanism. Step S3 improves the generalization ability of the model by separating the cross-device general rules and single-device specific parameters. Among them, the multi-task learning model is pre-trained on the historical data of the device network (such as delay logs, device status, protocol type) to learn the general representations of device behaviors (such as delay fluctuation patterns and abnormal signal correlations), and is jointly fine-tuned on multiple related tasks (delay overrun detection, device health prediction, network topology optimization). The multi-task learning model uses a shared underlying encoder (such as a time-series feature extraction module) to capture the common rules across devices (such as the time-sensitivity features of the IEEE 802.1AS protocol stack), and at the same time distinguishes different targets through task-specific head networks (classification tasks to judge whether it exceeds the threshold, regression tasks to predict the delay trend). It not only inherits the pre-trained model's macro understanding of the device group dynamics (such as the periodic changes in the regional network load), but also realizes knowledge transfer through multi-task collaboration (for example, the low-delay pattern learned from normal devices can assist in the diagnosis of abnormal devices). Finally, the intelligent determination and root cause analysis of complex conditions such as "the delay exceeds 20% of the average value of devices in the same cluster" are realized in a single model, while reducing the repeated calculations and inconsistent standards caused by independent task modeling.

[0069] As described in step S4 above, for adaptive feature fusion, the multi-head attention mechanism is used. The shared features are used as Query, and the exclusive features are used as Key-Value to calculate the dynamic weight matrix. For example, in robotic arm trajectory planning, the shared features provide a general kinematic model, and the exclusive features compensate for the gear clearance error. In domain adaptation, for newly added devices, through domain adversarial training (DANN), the spatial distribution difference (such as MMD distance) between their exclusive features and shared features is minimized to achieve feature space alignment. In transfer learning, the shared features are used as prior knowledge to initialize the model of the newly added device, and the exclusive features are fine-tuned online through a small amount of real-time data (such as 10 minutes of operation logs). Step S4 adapts to the hardware differences of newly added devices by dynamically adjusting the weights of the shared and exclusive features.

[0070] As described in step S5 above, the shared feature base model is compressed through knowledge distillation. For example, a large cloud-based Transformer model (with 100 million parameters) is distilled into a lightweight LSTM model (with 100,000 parameters) at the edge. The shared feature layer is retained and the parameters are quantized to 8-bit integers. During edge inference, the exclusive feature parameters stored locally (such as the robotic arm motor response curve) are loaded in real time and dynamically weighted with the general instructions (such as target coordinates) output by the shared feature base model to generate control instructions adapted to the current device's hardware characteristics. A real-time scheduling algorithm (such as Earliest Deadline First) is used to prioritize high-priority instructions (such as emergency stop signals) to ensure that the latency of newly added devices does not exceed 20% of the average latency of the collaborative system. Step S5 adopts an edge-cloud collaboration architecture to achieve low-latency control instruction generation at the edge while leveraging global optimization in the cloud.

[0071] As described in step S6 above, the online EWC (Elastic Weight Consolidation) algorithm is used to update the shared feature model to prevent forgetting of old knowledge. For frequently updated core parameters (such as the collaborative start-stop timing weights), the learning rate is reduced (such as from 0.01 to 0.001) to avoid key parameters being overwritten. The compressed shared feature parameters are sent to the edge through differential coding (only transmitting the changes) and an encryption protocol (such as TLS1.3). The edge verifies and hot-updates the model based on the version number without requiring downtime. If the latency of a newly added device exceeds the standard, its priority in the collaborative task is dynamically reduced (such as from priority 1 to 3) to avoid slowing down the overall system. Step S6 continuously optimizes the global shared feature model and synchronizes it to the edge to avoid model aging.

[0072] As described in step S7 above, the device collaboration timing is generated based on the global shared feature model. For example, robotic arm A starts grasping when the speed of conveyor belt B reaches 2 m / s, and CNC machine C starts processing 5 seconds after the grasping is completed. The edge of each device receives the collaborative instruction set sent from the cloud, including the robotic arm trajectory coordinates (Cartesian space path), conveyor belt speed adjustment instructions (PID parameters), and machine start-stop thresholds (such as temperature ≤ 80°C). The device status is monitored in real time. If a robotic arm trajectory conflict (such as the distance between two arms < 0.5 m) or machine load overrun (CPU utilization > 90%) is detected, an exception protocol is triggered to recalculate the collaboration strategy (such as switching to an obstacle avoidance path or reducing the frequency). Step S7 achieves the globally optimal production rhythm by coordinating the actions of heterogeneous devices in a unified manner.

[0073] In one embodiment, on an automotive welding production line, a six-axis robotic arm, a laser welding machine, and an intelligent conveyor belt deployed on-site upload operation data in real time through an edge gateway. The robotic arm joint encoder collects the angles and torques of each axis at a frequency of 100 Hz. The welding machine monitors the current fluctuations and weld appearance data. The conveyor belt feeds back the material position information through photoelectric sensors. After all the data is accessed through the OPC UA protocol, it is uniformly converted into a JSON-formatted event stream with an NTP-calibrated timestamp. Each event contains the device ID, millisecond-level timestamp, normalized value (e.g., torque 0.82 corresponds to an actual 12.5 N·m), and status label (normal / overload alarm).

[0074] When the production line switches to welding a new model of car door, the system detects a 300-ms timing deviation between the robotic arm's grasping action and the conveyor belt's in-position signal. After aligning the action sequences through the dynamic time warping algorithm, 18-dimensional features such as the variance of the robotic arm's acceleration (0.15 g2) and the fundamental frequency of the welding current (50 Hz ± 2 Hz) are extracted within a 2-second sliding window and associated with the device type to establish an index in the feature library. The feature library adopts a three-level retrieval structure and can quickly retrieve historical operating condition data according to "welding equipment → positioning action → 2024-05-20 14:00:00".

[0075] The pre-trained Transformer multi-task model in the cloud analyzes the feature library, separating cross-device shared features (such as the collaborative timing pattern of starting welding 500 ms after the material arrives) and exclusive features (the 0.12° backlash compensation parameter of the 4th joint of the KR20 robotic arm). The model ensures that the angle between the two types of feature vectors is greater than 85° through an orthogonal constraint loss function, avoiding the contamination of the general collaborative logic by the dedicated compensation parameters. When a new Yaskawa MH24 robotic arm is added, the system reuses the shared features to initialize its control model and extracts exclusive features (the 2.3-ms response delay parameter of the servo motor) based on the first 30 minutes of operation data, making the maximum mean discrepancy (MMD) distance between the distribution of the exclusive features and the shared feature space less than 0.05 through domain adversarial training.

[0076] The lightweight LSTM model deployed on the edge side (compressed to 3 MB through knowledge distillation) receives the shared base model parameters issued by the cloud and the exclusive feature parameters stored locally. During welding trajectory planning, the model dynamically fuses the shared general kinematic model (Cartesian space acceleration limit of 2 m / s2) with the exclusive joint compensation parameters to generate actual control instructions suitable for the MH24 robotic arm. When welding spatter is detected approaching the safety distance, the real-time scheduling algorithm preferentially processes the emergency stop instruction, ensuring that the instruction delay is controlled within 8 ms, which is less than 15% of the average delay of the production line.

[0077] The cloud aggregates the operation data of 120,000 devices every day and performs incremental learning using the Elastic Weight Consolidation (EWC) algorithm. For the welding timing coordination parameters with high-frequency updates (updated 12 times a day on average), the learning rate is reduced from 0.01 to 0.002 to maintain stability. After the updated shared feature model is distilled and compressed, the parameter change amount (average 2.7 KB per time) is encrypted and transmitted to the edge side through differential encoding, and the device completes seamless thermal updates within 0.5 seconds. When the instruction delay of the newly added robotic arm exceeds the standard due to network jitter, the system automatically reduces its priority in the welding timing from P1 to P3 to avoid affecting the overall production rhythm.

[0078] The finally generated coordination strategy guides the three robotic arms to complete the actions of door grabbing, positioning, and welding within a 0.8-second window, and dynamically adjusts the conveyor belt speed to increase the production rhythm to 43 JPH. When the vision system detects that the distance between the ends of the two robotic arms is less than 0.3 meters, the obstacle avoidance protocol is triggered in real time. While re-planning the trajectory based on the shared features, the exclusive parameters of each robotic arm are applied to compensate for the trajectory error, ensuring the safety and accuracy of collaborative operations (positioning error ±0.05 mm). The entire system can complete the adaptation process of the newly added device from access to full-capacity collaboration within 72 hours, reducing the commissioning time by 83% compared with the traditional PLC centralized control method.

[0079] Refer to Figure 2 , which is the structural block diagram of the industrial equipment adaptive collaborative control system in an embodiment of the present invention, including:

[0080] A data acquisition unit for real-time collecting the operation data of multiple devices and converting the operation data of the multiple devices into a standardized time-series event stream;

[0081] A database construction unit for performing time-series alignment of the action sequences of multiple devices on the time-series event stream and extracting key time features to construct a cross-device feature library;

[0082] A feature extraction unit for extracting high-level semantic features from the cross-device feature library and decoupling them into cross-device shared features and device-exclusive features through a pre-trained multi-task learning model;

[0083] A feature fusion unit for adaptively fusing the shared features and exclusive features through an attention mechanism, optimizing the device control accuracy using the exclusive features, extending the shared features to newly added devices through transfer learning, and extracting newly added device-exclusive features from the real-time operation data of the newly added devices through domain adaptation training;

[0084] A data optimization unit for deploying a lightweight inference model at the edge of the newly added device based on an edge-cloud collaborative architecture, and respectively loading the newly added device-exclusive feature parameters and shared feature base model parameters sent from the cloud;

[0085] A parameter synchronization unit, configured to aggregate the multi-device operation data through the cloud, optimize the global shared feature model based on incremental learning, and synchronize the updated shared feature parameters to the edge of each device;

[0086] An execution unit, configured to integrate the timing feature library, the multi-task model, and the edge-cloud collaboration architecture, generate a collaboration strategy through global shared features, and drive multi-type industrial devices to collaborate under a unified framework.

[0087] For the specific implementation of each unit in the above device instance, please refer to that described in the above method embodiment, and details are not described herein again.

[0088] In summary, the present invention collects multi-device operation data in real time, converts the multi-device operation data into a standardized timing event stream; performs timing alignment of multi-device action sequences on the timing event stream, extracts key time features to construct a cross-device feature library; extracts high-level semantic features from the cross-device feature library, and decouples them into cross-device shared features and device-specific features through a pre-trained multi-task learning model; adaptively fuses the shared features and the specific features through an attention mechanism, uses the specific features to optimize the device control accuracy, extends the shared features to new devices through transfer learning, and extracts new device-specific features from the real-time operation data of the new devices through domain adaptation training; based on the edge-cloud collaboration architecture, a lightweight inference model deployed at the edge of the new device loads the new device-specific feature parameters and the shared feature base model parameters sent from the cloud respectively. The new specific feature parameters are used to adapt to the hardware characteristics of the new device and are dynamically fused with the shared features during inference to optimize the control instruction accuracy. The shared feature base model is generated by compressing the global shared feature model through knowledge distillation, can adapt to the edge computing power, and performs low-latency instruction control for new devices; aggregates the multi-device operation data through the cloud, optimizes the global shared feature model based on incremental learning, and synchronizes the updated shared feature parameters to the edge of each device; integrates the timing feature library, the multi-task model, and the edge-cloud collaboration architecture, generates a collaboration strategy through global shared features, and drives multi-type industrial devices to collaborate under a unified framework, so as to achieve the purpose of adaptive collaboration control and rapid expansion of multi-type industrial devices in a dynamic production environment.

[0089] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium provided by the present invention and used in the embodiments can include non-volatile and / or volatile memories. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM, etc.

[0090] It should be noted that in this article, the term "including", "comprising", or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, device, article, or method including a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, device, article, or method. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of another identical element in the process, device, article, or method including the element.

[0091] The above are only the preferred embodiments of the present invention, and do not limit the patent scope of the present invention accordingly. Any equivalent structure or equivalent process transformation made by using the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. An adaptive collaborative control method for industrial equipment, characterized in that, It includes the following steps: Collect the operation data of multiple devices in real time and convert the operation data of the multiple devices into a standardized time series event stream; Perform time series alignment of the action sequences of multiple devices on the time series event stream, extract key time features, and construct a cross-device feature library; Extract high-level semantic features from the cross-device feature library and decouple them into cross-device shared features and device-specific features through a pre-trained multi-task learning model; Adaptively fuse the shared features and the specific features through an attention mechanism, optimize the device control accuracy using the specific features, extend the shared features to new devices through transfer learning, and extract new device-specific features from the real-time operation data of the new devices through domain adaptation training; Based on an edge-cloud collaborative architecture, deploy a lightweight inference model at the edge of the new device, and load the new device-specific feature parameters and the shared feature base model parameters sent from the cloud respectively. The new specific feature parameters are used to adapt to the hardware characteristics of the new device and dynamically fuse with the shared features during inference to optimize the control instruction accuracy. The shared feature base model is generated by compressing the global shared feature model through knowledge distillation, can adapt to the edge computing power, and perform low-latency instruction control for the new device; Aggregate the operation data of the multiple devices through the cloud, optimize the global shared feature model based on incremental learning, and synchronize the updated shared feature parameters to the edge of each device; Integrate the time series feature library, the multi-task model, and the edge-cloud collaborative architecture, generate a collaborative strategy through the global shared features, and drive multiple types of industrial devices to cooperate under a unified framework; 2. The industrial equipment adaptive cooperative control method according to claim 1, characterized in that The step of collecting the operation data of multiple devices in real time and converting the operation data of the multiple devices into a standardized time series event stream includes: Collect the operation data of each device through sensors, including the operation status data of the machine tool, the motion data of the robotic arm, the interaction data of the peripheral devices, and the logic index data; Convert the collected data into a time series event stream in a unified format according to a preset time window. Each event includes a device identifier, a timestamp, a data value, and a status identifier; 3. The industrial equipment adaptive cooperative control method according to claim 1, characterized in that, The step of performing time series alignment of the action sequences of multiple devices on the time series event stream, extracting key time features, and constructing a cross-device feature library includes: Based on the time series event stream, eliminate the phase deviation of the action time series between devices through the dynamic time warping algorithm, and align the time series data of each device; Based on the aligned time series data, extract the statistical features of each action sequence through a sliding window, including the mean, variance, and frequency domain features; Associate and store the statistical features with the device types, construct a cross-device feature library, and establish a three-level index of device type, action category, and timestamp; 4. The industrial equipment adaptive collaborative control method according to claim 1, characterized in that The step of decoupling into shared features and device-specific features through a pre-trained multi-task learning model includes: Encode the time series data of the cross-device feature library through a multi-task model, and extract the collaborative mode and abnormal correlation features of each device as high-level semantic features; Extract cross-device general parameters through the shared feature branch and extract device-specific parameters of each device through the specific feature branch; By decoupling the loss function, the orthogonality of the outputs of the shared feature branch and the exclusive feature branch is constrained to obtain the shared features of multiple devices and the exclusive features of each device.

5. The industrial equipment adaptive collaborative control method according to claim 1, wherein The step of adaptively fusing the shared features and the exclusive features through the attention mechanism includes: Taking the shared features as the reference input and the exclusive features as the conditional input, calculating the dynamic fusion weights through the multi-head attention mechanism and fusing them; Through domain adversarial training, minimize the spatial distribution difference between the exclusive features of the new device and the shared features to align the feature space; Update the exclusive feature parameters of the new device based on the real-time operation data of the new device.

6. The industrial equipment adaptive collaborative control method according to claim 1, wherein, The step of deploying the lightweight inference model at the edge of the new device includes: Deploy the inference model framework at the edge of the new device and load the shared feature base model parameters sent from the cloud; Synchronously load the exclusive feature parameters of the new device stored at the edge. The exclusive feature parameters of the new device are generated through domain adaptation training and stored locally; Parse the control instructions of the new device through the real-time scheduling algorithm, and dynamically adjust the robotic arm movement trajectory or the force of the end effector of the new device in combination with the loaded exclusive feature parameters of the new device.

7. The industrial equipment adaptive collaborative control method according to claim 1, characterized in that The step of aggregating the operation data of the multiple devices by the cloud, optimizing the global shared feature model based on incremental learning, and synchronizing the updated shared feature parameters to the edge of each device includes: Aggregate the real-time operation status data and historical log data of multiple devices, and update the global shared feature model parameters using the online incremental learning algorithm; Compress the updated global shared feature model through knowledge distillation technology, retain the parameters of the cross-device shared feature layer, and remove the device-exclusive feature-related nodes and redundant computing units; Compress the model through knowledge distillation, retain the parameters of the shared layer and remove redundant nodes; Differentially encode the change amount of the compressed shared feature parameters, and transmit them to the edge of each device through an encryption protocol. The edge end verifies and loads the updated parameters according to the version number.

8. The adaptive collaborative control method for industrial equipment according to claim 1, characterized in that, The step of aggregating the operation data of the multiple devices by the cloud, optimizing the global shared feature model based on incremental learning, and synchronizing the updated shared feature parameters to the edge of each device further includes: During the process of incremental learning optimization, calculate the historical update amplitude of the shared feature parameters, and reduce the learning rate for the core parameters that are frequently updated to prevent key parameters from being overwritten; The lightweight inference model reduces the memory occupancy at the edge by pruning the neuron connections related to the device-exclusive features in the shared feature base model and compressing the floating-point parameters into 8-bit integers; When a new device joins the collaborative operation, calculate the instruction execution delay based on the exclusive feature parameters of the new device. If the delay of the new device exceeds 20% of the average delay of the current online devices, the priority of the new device in the collaborative time sequence is dynamically reduced.

9. The industrial equipment adaptive collaborative control method according to claim 1, characterized in that The step of driving multiple types of industrial devices to collaborate under a unified framework includes: Generate the device collaborative start / stop time sequence and priority allocation rules based on the global shared feature model; Combined with the exclusive feature parameters of each device, send a collaborative instruction set including robotic arm trajectory coordinates, conveyor belt speed adjustment instructions, and machine start / stop thresholds to the edge end; Monitor the execution status of each device in real time. When detecting conflicts in the robotic arm trajectory coordinates, the conveyor belt speed exceeding the safety threshold, or the machine operation load exceeding the limit, trigger an exception handling protocol to recalculate the collaboration strategy.

10. An industrial equipment adaptive collaborative control system, characterized in that, It includes: A data acquisition unit, which is used to collect the operation data of multiple devices in real time and convert the operation data of the multiple devices into a standardized time series event stream; A database construction unit, which is used to perform time series alignment of the action sequences of multiple devices on the time series event stream, and extract key time features to construct a cross-device feature library; A feature extraction unit, which is used to extract high-level semantic features from the cross-device feature library and decouple them into cross-device shared features and device-specific features through a pre-trained multi-task learning model; A feature fusion unit, which is used to adaptively fuse the shared features and the specific features through an attention mechanism, optimize the device control accuracy using the specific features, extend the shared features to new devices through transfer learning, and extract new device-specific features from the real-time operation data of the new devices through domain adaptation training; A data optimization unit, which is used to deploy a lightweight inference model at the edge of the new device based on an edge-cloud collaboration architecture, and load the new device-specific feature parameters and the shared feature base model parameters sent from the cloud respectively; A parameter synchronization unit, which is used to aggregate the operation data of the multiple devices through the cloud, optimize the global shared feature model based on incremental learning, and synchronize the updated shared feature parameters to the edge of each device; An execution unit, which is used to integrate the time series feature library, the multi-task model, and the edge-cloud collaboration architecture, generate a collaboration strategy through the global shared features, and drive multiple types of industrial devices to collaborate under a unified framework.

Images