Industrial digital factory cooperative work method and cloud platform

Abstract

The invention discloses an industrial digital factory cooperative work method and a cloud platform, and the method comprises the steps: collecting factory multi-mode discrete source data, carrying out the cross-domain feature integration through a dynamic tensor fusion algorithm, and generating a global consistent factory situation fusion tensor; inputting the factory situation fusion tensor into a resource constrained collaborative decision engine, constructing a multi-device collaborative task optimization scheme based on a dynamic resource game model, and outputting a non-competitive multi-node collaborative scheduling instruction; performing abnormal situation recovery processing on the multi-node cooperative scheduling instruction, and outputting a continuous production control flow with enhanced stability; and inputting the continuous production control flow into a distributed edge collaborative architecture, dynamically configuring computing power resources based on an intelligent contract of a trusted execution environment, and generating a globally convergent factory collaborative control strategy set. By utilizing the embodiment of the invention, a high-robustness and self-adaptive cooperative working method can be provided for the industrial digital factory, the production efficiency is improved, and the energy consumption is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of factory collaboration, and particularly relates to a collaborative working method and cloud platform for an industrial digital factory. Background Art

[0002] With the rapid development of intelligent manufacturing, the digital factory, as the core carrier of modern industrial production, is gradually evolving towards a highly collaborative, intelligent, and flexible direction. The industrial digital factory integrates technologies such as the Internet of Things (IoT), edge computing, and artificial intelligence to achieve real-time interconnection and collaborative optimization of multiple elements such as production equipment, material flow, and energy systems, so as to improve production efficiency, reduce energy consumption, and enhance the robustness of the production system. However, the existing technologies still have significant technical bottlenecks in dealing with multi-modal data fusion, dynamic resource scheduling, anomaly recovery, and global optimization in complex industrial scenarios, which is an urgent problem to be solved. Summary of the Invention

[0003] The purpose of the present invention is to provide a collaborative working method and cloud platform for an industrial digital factory to solve the deficiencies in the existing technologies, and be able to provide a highly robust and adaptive collaborative working method for the industrial digital factory, improving production efficiency and reducing energy consumption.

[0004] An embodiment of the present application provides a collaborative working method for an industrial digital factory, and the method includes: Collecting multi-modal discrete source data of the factory, performing cross-domain feature integration through a dynamic tensor fusion algorithm, and using an asynchronous trigger synchronization mechanism to eliminate the communication clock offset and data sampling frequency difference between devices to generate a globally consistent factory situation fusion tensor; Inputting the factory situation fusion tensor into a collaborative decision-making engine with resource constraints, constructing an optimization scheme for multi-device collaborative tasks based on a dynamic resource game model, and adjusting the task execution sequence through a priority dynamic allocation algorithm for conflict resolution to output a non-competitive multi-node collaborative scheduling instruction; Performing anomaly situation recovery processing on the multi-node collaborative scheduling instruction, using a competitive feature refinement network to real-time analyze the device operation state and material flow anomaly features, and reconstructing the damaged production sequence through a parameterized spatio-temporal repair operator to output a continuous production control flow with enhanced stability; Inputting the continuous production control flow into a distributed edge collaboration architecture, dynamically configuring computing power resources based on smart contracts in a trusted execution environment, and synchronously optimizing energy consumption and output efficiency through a multi-objective decision-making algorithm with chaotic optimization to generate a globally convergent set of factory collaborative control strategies to achieve the collaborative working of the industrial digital factory.

[0005] Optionally, the acquisition factory's multi-modal discrete source data is integrated with cross-domain features through a dynamic tensor fusion algorithm, and an asynchronous trigger synchronization mechanism is used to eliminate the communication clock offset and data sampling frequency difference between devices, generating a globally consistent factory situation fusion tensor, including: According to the time-domain waveform of the device vibration spectrum and the spatial coordinates of the production line material displacement trajectory in the factory's multi-modal discrete source data, an event-triggered pulse encoder is used to convert the continuous signal into an asynchronous pulse sequence, generating a multi-modal pulse matrix with spatio-temporal tags; wherein, the encoder suppresses high-frequency noise pulses through a dynamic threshold adjustment mechanism and retains effective operating condition features; Measure the phase difference between the multi-modal pulse matrix and the environmental electromagnetic field intensity data, construct a time-varying weight matrix through an event-driven adaptive phase calibration algorithm, dynamically compensate for the clock offset and sampling frequency difference between devices, and output a cross-modal pulse tensor with phase alignment; Input the cross-modal pulse tensor into the dynamic tensor fusion algorithm, construct a pulse energy density kernel in the time-space-frequency three domains, separate the core feature subspaces of device vibration, material displacement, and electromagnetic field through convolutional tensor decomposition, and eliminate multi-source signal interference; Accumulate the pulse energy of the core feature subspaces and perform spatial topology mapping, generate a globally consistent factory situation fusion tensor based on the dynamic edge weight assignment algorithm; wherein, the factory situation fusion tensor reflects the coupling situation of the device cluster in real time through a pulse-triggered weight update mechanism.

[0006] Optionally, input the factory situation fusion tensor into a resource-constrained collaborative decision-making engine, construct an optimization scheme for multi-device collaborative tasks based on a dynamic resource game model, adjust the task execution sequence through a priority dynamic allocation algorithm for conflict resolution, and output a non-competitive multi-node collaborative scheduling instruction, including: Input the factory situation fusion tensor into a resource-constrained collaborative decision-making engine, construct a multi-device task revenue matrix based on a dynamic resource game model, quantify the task priority weights through the Nash bargaining solution algorithm, and generate an initial game strategy set; Perform a fractal dimension attention mechanism process on the initial game strategy set, calculate the fractal dimension difference between key tasks and regular tasks, and forcefully isolate the resource occupancy thresholds of the two types of tasks through an attention mask, outputting an attention weight vector with fractal constraints; According to the attention weight vector, adjust the task execution sequence through a priority dynamic allocation algorithm for conflict resolution, wherein the priority dynamic allocation algorithm optimizes the stability of the task queue through the Lyapunov drift plus penalty function; Perform a policy evolution algorithm process on the adjusted task execution sequence, and co-optimize the fault tolerance of the scheduling path through Monte Carlo tree search and deep policy network, and output a non-competitive multi-node collaborative scheduling instruction.

[0007] Optionally, perform an abnormal situation recovery process on the multi-node collaborative scheduling instruction, use a competitive feature refinement network to real-time analyze the device operation state and material flow abnormal features, and reconstruct the damaged production sequence through a parameterized spatio-temporal repair operator, and output a continuous production control flow with enhanced stability, including: Perform a virtual abnormal sample generation process on the collaborative scheduling instruction, use a generative adversarial network to simulate extreme working conditions such as device downtime and material blockage, and construct an enhanced training data set containing abnormal labels; Input the data set into the competitive feature refinement network, separate the device operation state and material flow abnormal features through a two-channel adversarial training mechanism, and output a high-resolution abnormal feature vector; wherein, the network enhances the abnormal detection sensitivity through a game theory-driven feature distillation technology; Perform a parameterized spatio-temporal repair operator process on the abnormal feature vector, construct a spatio-temporal manifold completion model of the damaged production sequence, and reconstruct a continuous production sequence through a geodesic-constrained energy minimization algorithm; Use the reconstructed continuous production sequence, and adopt a reverse error compensation mechanism to synchronously correct the operation deviation between the physical device actuator and the digital twin model. Among them, the reverse error compensation mechanism realizes real-time bidirectional calibration of control instructions through implicit gradient propagation, and outputs a continuous production control flow with enhanced stability.

[0008] Optionally, input the continuous production control flow into a distributed edge collaborative architecture, dynamically configure computing power resources based on the smart contract of the trusted execution environment, and synchronously optimize energy consumption and output efficiency through a chaotic optimization multi-objective decision algorithm, and generate a globally convergent factory collaborative control strategy set, including: Input the continuous production control flow into a distributed edge collaborative architecture, verify the integrity and compliance of the control instruction based on the smart contract of the trusted execution environment, and generate a verification vector encrypted with a digital signature; Perform a chaotic optimization multi-objective decision algorithm process on the verification vector, generate a chaotic sequence to traverse the solution space through Logistic mapping, and use the Pareto front fast convergence technology to synchronously optimize energy consumption and output efficiency, and output a preliminary control strategy set; Perform an implicit state synchronization protocol process between edge nodes on the preliminary control strategy set, eliminate resource contention conflicts through a distributed consensus algorithm for conflict residuals, and generate a conflict-free subset of collaborative control instructions; The time - stamp dynamic interpolation technique is adopted to align the collaborative control instruction subset along the time axis, and the influence of network jitter is eliminated through an instruction buffer queue constrained by a sliding window, and finally a globally convergent and temporally consistent factory collaborative control strategy set is output.

[0009] Another embodiment of the present application provides an industrial digital factory collaborative working cloud platform, and the cloud platform includes: An acquisition module, which is used to acquire multi - modal discrete source data of the factory, perform cross - domain feature integration through a dynamic tensor fusion algorithm, eliminate the communication clock offset between devices and the data sampling frequency difference by using an asynchronous trigger synchronization mechanism, and generate a globally consistent factory situation fusion tensor; An adjustment module, which is used to input the factory situation fusion tensor into a collaborative decision - making engine constrained by resources, construct an optimization scheme for multi - device collaborative tasks based on a dynamic resource game model, adjust the task execution sequence through a priority dynamic allocation algorithm for conflict resolution, and output a non - competitive multi - node collaborative scheduling instruction; An analysis module, which is used to perform abnormal situation recovery processing on the multi - node collaborative scheduling instruction, adopt a competitive feature refinement network to analyze the device operation state and material flow abnormal features in real - time, reconstruct the damaged production sequence through a parameterized spatio - temporal repair operator, and output a continuous production control flow with enhanced stability; An optimization module, which is used to input the continuous production control flow into a distributed edge collaborative architecture, dynamically configure computing power resources based on smart contracts in a trusted execution environment, synchronously optimize energy consumption and output efficiency through a multi - objective decision - making algorithm optimized by chaos, and generate a globally convergent factory collaborative control strategy set to achieve the collaborative work of the industrial digital factory.

[0010] Another embodiment of the present application provides a storage medium, in which a computer program is stored, and the computer program is set to execute the method described in any one of the above when running.

[0011] Another embodiment of the present application provides an electronic device, including a memory and a processor, a computer program is stored in the memory, and the processor is set to run the computer program to execute the method described in any one of the above.

[0012] Compared with the prior art, a collaborative working method for an industrial digital factory provided by the present invention collects multi-modal discrete source data of the factory, performs cross-domain feature integration through a dynamic tensor fusion algorithm, and generates a globally consistent factory situation fusion tensor; inputs the factory situation fusion tensor into a collaborative decision-making engine with resource constraints, constructs an optimization scheme for multi-device collaborative tasks based on a dynamic resource game model, and outputs non-competitive multi-node collaborative scheduling instructions; performs abnormal situation recovery processing on the multi-node collaborative scheduling instructions, and outputs a continuous production control flow with enhanced stability; inputs the continuous production control flow into a distributed edge collaborative architecture, and dynamically configures computing resources based on smart contracts in a trusted execution environment to generate a globally convergent set of factory collaborative control strategies, thereby being able to provide a highly robust and adaptive collaborative working method for the industrial digital factory, improving production efficiency and reducing energy consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 is a hardware structure block diagram of a computer terminal for a collaborative working method of an industrial digital factory provided by an embodiment of the present invention; Figure 2 is a flowchart of a collaborative working method of an industrial digital factory provided by an embodiment of the present invention; Figure 3 is a structural diagram of a collaborative working cloud platform of an industrial digital factory provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0014] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0015] An embodiment of the present invention first provides a collaborative working method for an industrial digital factory, which can be applied to an electronic device, such as a computer terminal, specifically, an ordinary computer, etc.

[0016] The following takes running on a computer terminal as an example to describe it in detail. Figure 1 is a hardware structure block diagram of a computer terminal for a collaborative working method of an industrial digital factory provided by an embodiment of the present invention. As Figure 1 shown, the computer device includes a processor, a memory, and a network interface connected through a cloud platform bus. Among them, the memory may include a non-volatile storage medium and an internal memory.

[0017] The non-volatile storage medium can store an operating cloud platform and a computer program. The computer program includes program instructions, and when the program instructions are executed, the processor can execute any collaborative working method of an industrial digital factory.

[0018] The processor is used to provide computing and control capabilities to support the operation of the entire computer device.

[0019] The internal memory provides an environment for the operation of a computer program in a non-volatile storage medium. When the computer program is executed by a processor, the processor can be caused to execute any one of the industrial digital factory collaborative working methods.

[0020] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art can understand that Figure 1 The structure shown in is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0021] It should be understood that the processor may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0022] See Figure 2 , an embodiment of the present invention provides an industrial digital factory collaborative working method, which may include the following steps: S201, collect factory multi-modal discrete source data, perform cross-domain feature integration through a dynamic tensor fusion algorithm, and use an asynchronous trigger synchronization mechanism to eliminate the communication clock offset and data sampling frequency difference between devices, and generate a globally consistent factory situation fusion tensor; Specifically, according to the time-domain waveform of the device vibration spectrum and the spatial coordinates of the production line material displacement trajectory in the factory multi-modal discrete source data, an event-triggered pulse encoder can be used to convert the continuous signal into an asynchronous pulse sequence, and generate a multi-modal pulse matrix with spatio-temporal tags; among them, the encoder suppresses high-frequency noise pulses through a dynamic threshold adjustment mechanism and retains effective working condition characteristics; The factory's multi-modal discrete source data includes the time-domain waveform of the device vibration acceleration sensor (sampling rate 10 kHz) and the lidar trajectory data of the production line material displacement (sampling rate 100 Hz). The Event-Triggered Pulse Encoder (ETPE) converts the continuous signal into a pulse sequence through a dynamic threshold adjustment mechanism.

[0023] Dynamic threshold setting: The initial threshold is set to 1.5 times the root mean square (RMS) value of the signal. For example, the RMS of a certain device vibration signal is 0.3 g (unit of gravitational acceleration), and the initial threshold is 0.45 g. When 5 consecutive sampling points exceeding the threshold are detected, a pulse is triggered and the threshold is dynamically adjusted to 80% of the current peak value (e.g., when the detected peak value is 0.6 g, the new threshold becomes 0.48 g).

[0024] Pulse generation rule: Each pulse corresponds to an event and carries spatio-temporal tags; Timestamp: Accurate to the microsecond level (e.g., 163045 μs), and the global clock synchronization is ensured through the GPS synchronization module; Spatial coordinates: The XYZ coordinates of the material displacement trajectory (accuracy ±0.1 mm), which are real-time located by the lidar SLAM algorithm.

[0025] Noise suppression: Adopt frequency-domain energy analysis with a sliding window (window width 50 ms) to filter out high-frequency noise above 1 kHz (such as motor electromagnetic interference). For example, when it is detected that the energy ratio of a certain signal segment in the 1.5 kHz frequency band exceeds 30%, it is determined as noise and the corresponding pulse trigger is suppressed.

[0026] The generated multi-modal pulse matrix is a three-dimensional structure (time × space × modality), for example: Time axis: Each row corresponds to a 1 ms time slot; Spatial axis: Each column corresponds to the production line station coordinates (such as station A1 - X: 100.5, Y: 200.3); Modality axis: Store data such as vibration, displacement, and temperature in layers. An example of a record: [Time: 163045 μs, Coordinate: A1, Modality: Vibration, Pulse intensity: 0.6 g].

[0027] Measure the phase difference between the multi-modal pulse matrix and the environmental electromagnetic field intensity data, construct a time-varying weight matrix through an event-driven adaptive phase calibration algorithm, dynamically compensate for the clock offset and sampling frequency difference between devices, and output a cross-modal pulse tensor with phase alignment; Environmental electromagnetic field data is collected by distributed electromagnetic sensors (sampling rate 1 kHz), and there are clock deviations (typical value ±5 ms) and sampling rate differences (vibration 10 kHz vs electromagnetic 1 kHz) from equipment vibration and material displacement data. The phase difference measurement uses the cross-correlation algorithm.

[0028] Time window alignment: Taking the vibration signal as a reference, after downsampling the electromagnetic signal by a factor of 10:1, calculate the peak offset of the cross-correlation function of the two signals within a 1-second window. For example, if the peak offset within a certain window is 3 ms, it is determined that the electromagnetic signal is delayed by 3 ms relative to the vibration signal.

[0029] Dynamic compensation: Adjust the electromagnetic signal timestamp through linear interpolation. For example, adjust the original sampling point t = 1000 ms to t = 997 ms.

[0030] Core process of the event-driven adaptive phase calibration algorithm: Weight matrix construction: The clock deviation Δti and sampling rate difference ri of each device node are mapped to the weight wij, where i and j represent device pairs. For example, if the Δti between device A and B is 2 ms and ri = 0.98 (sampling rate 98% match), then wij = 0.98×e^(-0.002² / 2σ²) (σ = 1 ms).

[0031] Phase synchronization: Using the distributed consensus protocol, each device periodically broadcasts its local clock status, and updates the global clock reference through weighted averaging. For example, after 5 rounds of iteration (each round interval is 100 ms), the global clock synchronization error is reduced from ±5 ms to ±0.1 ms.

[0032] The output cross-modal pulse tensor realizes data alignment through spatio-temporal interpolation: Time alignment: Resample all signals at a granularity of 10 μs; Space alignment: Map the material displacement coordinates to a unified grid (grid size 10 cm × 10 cm).

[0033] Example tensor slice: [Time: 163045 μs, Grid: G12, Mode: Electromagnetic field, Intensity: 50 μT].

[0034] Input the cross-modal pulse tensor into the dynamic tensor fusion algorithm, construct a pulse energy density kernel in the time-space-frequency three domains, separate the core feature subspaces of equipment vibration, material displacement, and electromagnetic field through convolutional tensor decomposition, and eliminate multi-source signal interference; The dynamic tensor fusion algorithm adopts three-stage processing: Time-space-frequency three-domain decomposition: Time domain: Convert the pulse intensity into a time-frequency matrix through the Short-Time Fourier Transform (STFT, window length 256 ms, overlap 50%); Spatial domain: Extract spatial neighborhood features using a three-dimensional convolutional kernel (size 5×5×5); Frequency domain: Perform wavelet packet decomposition (4 layers, mother wavelet db4) on the time-frequency matrix to extract low-frequency energy below 1 kHz.

[0035] Pulse energy density kernel construction: Perform Hadamard product on the three-domain features to generate the energy density kernel. For example, the energy density of a certain spatio-temporal unit (time slot 163045 μs, grid G12) in the 100 Hz frequency band is vibration 0.6 g² / Hz + displacement 0.1 mm² / Hz + electromagnetic field (50 μT)² / Hz.

[0036] Convolutional tensor decomposition: Use CP decomposition (Canonical Polyadic Decomposition) to decompose the original tensor into three core feature subspaces: equipment vibration, material displacement, and electromagnetic field: Equipment vibration subspace: Extract the first 3 principal components through Higher-Order Singular Value Decomposition (HOSVD), explaining 90% of the variance; Material displacement subspace: Use Non-Negative Matrix Factorization (NMF) to separate the X / Y / Z axial motion patterns; Electromagnetic field subspace: Use Independent Component Analysis (ICA) to eliminate power frequency interference.

[0037] Example of interference cancellation: The 50 Hz power frequency interference at a certain work station is identified as an independent component in the electromagnetic field subspace, and its weight is reduced from 0.8 to 0.05, while the weight of the effective signal (such as 1 kHz motor harmonics) is increased from 0.2 to 0.7.

[0038] Perform pulse energy accumulation and spatial topology mapping on the core feature subspaces, and generate a globally consistent factory situation fusion tensor based on the dynamic edge weight assignment algorithm; among them, the factory situation fusion tensor reflects the coupling situation of the equipment cluster in real time through the weight update mechanism triggered by pulses.

[0039] Pulse energy accumulation uses sliding window integration: Time accumulation: Use 1 second as the window to sum the equipment vibration energy. For example, a certain equipment triggers 120 pulses within 1 second, and the total energy Σ(0.6² + 0.5² +...) = 45 g²·s.

[0040] Spatial accumulation: Aggregate the energy according to the production line work station grid. For example, the total displacement energy of work station G12 is the mean of the sum of the squares of the X / Y / Z axis displacements.

[0041] Spatial topology mapping: Device Nodes: Each device is mapped to a graph node, and the node attributes include the energy accumulation value, fault code, etc.; Connection Edges: The edge weight is jointly determined by the material flow relationship (such as the transmission frequency of device A→B) and the energy coupling degree (such as the vibration transfer coefficient). For example, the edge weight wij from device A to B is 0.7×transmission frequency + 0.3×vibration coupling degree.

[0042] Dynamic Edge Weight Allocation Algorithm: Pulse-triggered Update: When the pulse energy of a certain device exceeds the threshold (such as 30 g²·s), it triggers the update of the weights of neighboring nodes. For example, abnormal vibration of device A causes the edge weight between it and device B to increase from 0.5 to 0.8.

[0043] Decay Mechanism: The weight decays exponentially with time (decay factor λ = 0.9 / s) to prevent historical anomalies from continuously affecting.

[0044] The generated factory situation fusion tensor is a four-dimensional structure (time×space×device×coupling strength), and an example slice: [Time: 10:00:00, Space: G12, Device: CNC-01, Coupling Strength: 0.85] This tensor is transmitted to the MES system in real time through the OPC UA protocol, supporting the evaluation of device health and the optimization of production rhythm. For example, when the coupling strength in a certain area is continuously >0.9 for 5 minutes, it triggers a preventive maintenance alarm.

[0045] ‌S202, input the factory situation fusion tensor into a collaborative decision-making engine with resource constraints, construct an optimization scheme for multi-device collaborative tasks based on a dynamic resource game model, adjust the task execution sequence through a priority dynamic allocation algorithm for conflict resolution, and output a non-competitive multi-node collaborative scheduling instruction; Specifically, the factory situation fusion tensor can be input into a collaborative decision-making engine with resource constraints, construct a multi-device task revenue matrix based on a dynamic resource game model, quantify the task priority weights through the Nash bargaining solution algorithm, and generate an initial game strategy set; The core of the dynamic resource game model is to conduct game modeling on the task requirements of factory equipment (such as robotic arms, AGV cars, numerically controlled machine tools) and limited resources (such as power, computing power, material supply rate). Resource constraints include: Power Resource: The total power limit is 100 kW, and the power demand of the device changes dynamically (such as the peak power of a numerically controlled machine tool is 30 kW); Computing Power Resource: The upper limit of the computing power allocation of the edge server is 80 TOPS (trillion operations per second), which is used for real-time path planning and quality control; Material Resource: The maximum throughput of the conveyor belt is 500 pieces per hour.

[0046] Revenue Matrix Construction: Task Revenue Function: The revenue of each task (such as part processing, material handling) is calculated by weighting the completion time, energy consumption, and quality qualification rate, with weights of 0.5, 0.3, and 0.2 respectively. For example, if a processing task is completed within 30 seconds and the energy consumption is less than 5 kW, the revenue value is 0.85; Resource Conflict Detection: When multiple devices apply for the same resource simultaneously (such as two AGVs competing for the same path), it is marked as a conflict task pair; Nash Bargaining Solution (NBS): Determine the priority by maximizing the joint utility function of the task pair. The utility function is defined as: ;

[0047] where u1 and u2 are the task revenues, and d1 and d2 are the default losses in case of conflict (such as the revenue decline caused by task delay).

[0048] Initial Game Strategy Set Generation: Strategy Encoding: Each strategy is a tuple of device ID, task type, and resource allocation ratio (such as AGV_01, handling task, occupying 30% of the conveyor belt bandwidth); Strategy Screening: Eliminate strategies with resource overlimits (such as total power demand > 100 kW) and retain valid strategies; Example Output: The initial strategy set contains 200 valid strategies, of which 50 are high-revenue strategies (revenue > 0.8).

[0049] Perform fractal dimension attention mechanism processing on the initial game strategy set, calculate the fractal dimension difference between critical tasks and regular tasks, and forcefully isolate the resource occupancy thresholds of the two types of tasks through an attention mask, and output the attention weight vector with fractal constraints; Fractal Dimension Calculation: Task Feature Extraction: The feature vector of each task includes duration, resource occupancy rate, and quality impact factor (such as machining accuracy requirement ±0.01 mm), with a dimension of 10; Fractal Dimension Estimation: Use the Box-Counting Method to calculate the fractal dimension of the task in the feature space: Critical Tasks (such as precision part processing): High fractal dimension (>1.5), complex feature distribution; Regular Tasks (such as material handling): Low fractal dimension (<1.2), concentrated features.

[0050] Attention Mechanism Design: Fractal Difference Quantification: Calculate the fractal dimension difference ratio between critical tasks and regular tasks (such as 1.8 vs 1.1, difference ratio 0.7); Attention mask generation: Critical task mask: weight coefficient 1.0, resource occupancy upper limit increased by 20% (e.g., the power upper limit of a CNC machine tool is increased from 30kW to 36kW); Regular task mask: weight coefficient 0.5, resource occupancy upper limit decreased by 30% (e.g., the path priority of an AGV is downgraded).

[0051] Dynamic isolation: When there is a resource conflict between critical tasks and regular tasks, force the regular tasks to be delayed (e.g., delayed by 5 seconds).

[0052] Example: Under fractal constraints, the precision machining task obtains 40% of the exclusive resources of the conveyor belt bandwidth, while the bandwidth limit of the AGV handling task is 15%, ensuring zero interruption of critical tasks.

[0053] According to the attention weight vector, adjust the task execution sequence by using a priority dynamic allocation algorithm for conflict resolution, where the priority dynamic allocation algorithm optimizes the stability of the task queue through Lyapunov drift plus penalty function; Lyapunov drift optimization: Queue model: Each device maintains a task queue, and the queue length Q(t) reflects the backlog situation; Drift function: Define the Lyapunov function , measuring the cumulative delay of the system, where is the task queue length of the i-th device at time t, i is the device node index identifier, and t is the discrete time step number, is the Lyapunov function, a scalar index characterizing the overall congestion degree of the system; Penalty function: Introduce a resource overrun penalty term , is the total power consumption, ; Optimization objective: Minimize drift plus penalty : ; where is the drift term, , representing the change in the system congestion degree between adjacent time steps, used to measure the dynamic impact of the control strategy on task backlog: : The system congestion is alleviated (task processing speed > new task rate); : The system congestion is aggravated (task backlog deteriorates); is the trade-off parameter, controlling the balance between resource constraint violation and delay.

[0054] Priority dynamic allocation: Real-time scheduling: Update the task sequence every 500 ms and sort based on DPP values: High-priority tasks: DPP value > threshold (e.g., > 5), execute immediately; Medium-priority tasks: DPP value 2 - 5, allocate resources according to fractal weights; Low-priority tasks: DPP value < 2, enter the buffer queue.

[0055] Conflict resolution: When multiple tasks compete for the same resource, select the task with the highest DPP value to execute first.

[0056] Example: At a certain moment, the DPP of the CNC machine tool task is 8 (high priority), and the DPP of the AGV task is 3 (medium priority). The system preferentially allocates 36 kW of power to the machine tool, and the AGV task is delayed until the next cycle.

[0057] Perform a policy evolution algorithm on the adjusted task execution sequence, and co-optimize the fault tolerance of the scheduling path through Monte Carlo tree search and deep policy network to output a non-competitive multi-node collaborative scheduling instruction.

[0058] Monte Carlo tree search (MCTS): Number of simulations: Conduct 1000 simulations for each task sequence to explore different resource allocation paths; Node expansion: Each node represents a resource allocation state (such as a power allocation plan), and the child nodes are possible actions (such as +5 kW for the machine tool); Reward calculation: Simulation result reward = task revenue - 0.1 × delay time (seconds).

[0059] Deep policy network (DPN): Input: Current factory situation tensor (dimension 256), task queue status (dimension 50); Output: Action probability distribution (e.g., probability of allocating power to the machine tool is 0.7, and to the AGV is 0.3); Training: Perform supervised learning through the simulation results of MCTS, and the loss function is cross-entropy + KL divergence.

[0060] Co-optimization process: MCTS generates candidate paths: For example, path A (guarantee the machine tool first) has a reward of 0.85, and path B (balanced allocation) has a reward of 0.78; DPN selects the optimal path: According to the action probabilities output by the network, select path A; Fault tolerance verification: Inject simulated faults (such as a certain AGV crashing) to verify whether path A can be dynamically adjusted to an alternative path.

[0061] Output example: The final scheduling instruction contains 50 conflict-free task sequences, and the pass rate of the fault tolerance index (such as the single-point fault recovery time < 3 seconds) is 98%.

[0062] S203, perform abnormal situation recovery processing on the multi-node collaborative scheduling instruction, use a competitive feature refinement network to real-time analyze the device operation state and material flow abnormal features, reconstruct the damaged production sequence through a parameterized spatio-temporal repair operator, and output a continuous production control flow with enhanced stability; Specifically, virtual abnormal sample generation processing can be performed on the collaborative scheduling instruction, use a generative adversarial network to simulate extreme working conditions such as device downtime and material blockage, and construct an enhanced training data set containing abnormal labels; The core goal of virtual abnormal sample generation is to simulate rare extreme fault modes in industrial scenarios through a generative adversarial network (GAN) to enhance the robustness of the model. The generator uses a deep convolutional network (DCGAN architecture), with a 100-dimensional Gaussian noise vector as the input and the output being simulated abnormal working condition data. The generator structure contains 4 transposed convolutional layers (layer parameters: kernel size 4×4, stride 2, number of channels 512 / 256 / 128 / 64), and the last layer uses the Tanh activation function to normalize the output to [-1,1]. The discriminator consists of 4 convolutional layers (kernel size 4×4, stride 2, number of channels 64 / 128 / 256 / 512), and finally outputs the probability of data authenticity through the Sigmoid function.

[0063] Abnormal mode definition and generation strategy: Device downtime: Simulate the motor current suddenly dropping to 0A and the temperature curve rising sharply (such as rising from 60°C to 120°C within 5 seconds); Material blockage: Generate the conveyor belt speed dropping from 1.5 m / s to 0.2 m / s and the pressure sensor reading exceeding the limit (such as soaring from 10 kPa to 50 kPa); Sensor failure: Construct signal drift (such as continuous linear offset ±20%) or pulse noise (random spike amplitude ±30%).

[0064] Training process: Data preparation: Use normal working condition data (10,000 samples) as real data, and the initial generator randomly generates abnormal samples; Adversarial training: Use the Wasserstein GAN (WGAN) framework, with a gradient penalty coefficient λ = 10, batch size 64, and the generator and discriminator are alternately trained, and the learning rates are both set to 0.0002; Label enhancement: Label the generated abnormal samples with fault types (such as the device downtime label being [1,0,0], and material blockage being [0,1,0]) and severity levels (continuous values from 0 to 1).

[0065] Example output: The generated dataset contains 20,000 samples (50% normal and 50% abnormal). The features of one abnormal sample include: Current signal: Drops from 15A to 0A within 2 seconds; Temperature curve: Rises from 65°C to 115°C within 5 seconds; Label: [Equipment downtime, severity 0.9].

[0066] Input the dataset into the Competitive Feature Refinement Network. Through a dual-channel adversarial training mechanism, separate the device operating state and the abnormal features of the material flow, and output a high-resolution abnormal feature vector. Among them, the network enhances the anomaly detection sensitivity through game theory-driven feature distillation technology; Competitive Feature Refinement Network (CFRN) architecture: The network adopts a dual-channel design to process device state features and material flow features respectively: Device state channel: The input is time series data such as vibration spectrum, current, and temperature (dimension 100×6). Extract features through a 1D convolutional layer (kernel size 5, number of channels 32 / 64 / 128), and output the device health score (0~1); Material flow channel: The input is conveyor belt speed, pressure, and position coordinates (dimension 100×4). Use an LSTM network (128 hidden units) to capture spatio-temporal dependencies and output the probability of material flow anomaly (0~1).

[0067] Dual-channel adversarial training mechanism: Feature distillation: Constrain the difference in feature distributions between the two channels through Kullback-Leibler Divergence (KL divergence), and force the separation of abnormal patterns of the device and the material flow. For example, when the device channel detects an abnormal current, the material channel needs to suppress the response to pressure mutations; Game theory-driven optimization: Design a zero-sum game framework. The device channel and the material channel are used as adversarial parties, and the objective function is to minimize the detection accuracy of the other channel. The loss function is:

[0068] Among them, is the total loss (the global loss of dual-channel adversarial training), is the loss of the device state channel, which is used to measure the anomaly detection error of the device operating state (such as vibration, current, temperature); is the loss of the material flow channel, which is used to evaluate the detection accuracy of material flow anomalies (such as blockage, interruption of flow). α = 0.5 is the balance coefficient, and adversarial training is achieved through a Gradient Reversal Layer (GRL); Sensitivity Enhancement: Introduce Focal Loss (γ = 2) to handle class imbalance and improve the detection ability for small-sample anomalies (such as faults with severity > 0.8).

[0069] Training Parameters and Output: Batch Size: 32; Optimizer: Adam (β1 = 0.9, β2 = 0.999); Learning Rate: 0.001, decaying by 50% every 10 epochs; Output Abnormal Feature Vector: 64-dimensional, including device anomaly score (0.92), material anomaly score (0.15), spatio-temporal correlation weight (0.78), etc.

[0070] Example Application: When the conveyor belt pressure suddenly increases to 45 kPa, the output abnormal probability of the material channel is 0.88, and the device channel outputs 0.12 due to stable current. The system determines it as material blockage rather than device failure.

[0071] Perform parametric spatio-temporal repair operator processing on the abnormal feature vector, construct a spatio-temporal manifold completion model for the damaged production sequence, and reconstruct the continuous production sequence through the geodesic-constrained energy minimization algorithm; Parametric Spatio-Temporal Repair Operator Design: Manifold Modeling: Map the spatio-temporal state of the production sequence to the Riemannian manifold space, and each state point consists of a 5-dimensional vector composed of position coordinates (x, y, z), velocity v, and time t; Geodesic Constraint: Define the geodesic distance (shortest path) on the manifold to measure the deviation between the normal production sequence and the damaged sequence. For example, the geodesic distance for the material from A to B under normal conditions is 10 units, and it increases to 15 units when damaged; Energy Function Construction: Data Fidelity Term: Minimize the mean square error between the repaired sequence and the measured data (weight 0.7); Smoothing Term: Constrain the state change rate of adjacent time steps (weight 0.2); Geodesic Term: Force the repair path to approximate the normal manifold (weight 0.1).

[0072] Energy Minimization Algorithm Process: Initialization: Start with the damaged sequence, set the maximum number of iterations to 100, and the convergence threshold Δ < 0.001; Gradient Descent: Update the sequence parameters along the negative gradient direction of the energy function with a step size of 0.01; Manifold Projection: Project the current sequence onto the normal manifold every 5 iterations to ensure physical feasibility. For example, if a certain update causes the manipulator to move beyond the limit (joint angle > 120°), project it to the nearest feasible point (120°).

[0073] Example Reconstruction: A damaged production sequence has a 30 - second material delay due to a conveyor belt jam. After repair, the sequence compresses the total delay to 5 seconds by inserting an acceleration section (speed increased from 1.0 m / s to 1.5 m / s for 10 seconds), and the geodesic distance drops from 18.7 to 2.3.

[0074] Using the reconstructed continuous production sequence, adopt a reverse error compensation mechanism to synchronously correct the operation deviation between the physical device actuator and the digital twin model. Among them, the reverse error compensation mechanism realizes real - time two - way calibration of control instructions through implicit gradient propagation, and outputs a continuous production control flow with enhanced stability.

[0075] Implementation of the Reverse Error Compensation Mechanism: Two - way Calibration Architecture: Physical → Digital: Real - time collect the actual position of the device actuator (such as encoder readings) and force sensor data, compare with the predicted values of the twin model, and calculate the residual Δ; Digital → Physical: Send the repaired control instructions (such as speed set values) to the actuator, and dynamically adjust the twin model parameters according to Δ.

[0076] Implicit Gradient Propagation: Construct an implicit function relationship between the residual Δ and the model parameter θ, and calculate the gradient ∂Δ / ∂θ through automatic differentiation (Autograd); Update θ using momentum gradient descent (Momentum = 0.9), with a learning rate of 0.001, and update once every 100 ms.

[0077] Real - time Guarantee: Deploy a lightweight inference engine (TensorRT) on the edge computing node to achieve microsecond - level gradient calculation; The control instruction publishing frequency is 1 kHz, ensuring that the actuator response delay < 1 ms.

[0078] Stability Enhancement Strategy: Dead - zone Compensation: When |Δ| < 0.5% (such as position error < 0.1 mm), pause calibration to avoid oscillation; Historical Memory: Maintain a residual sequence of a sliding window (length 10), and suppress noise through weighted average (weight of the latest data is 0.3).

[0079] Example Output: The positioning error of a certain robotic arm accumulated up to 2 mm due to gear clearance. After compensation, the error was reduced to 0.05 mm. At the same time, the reverse calibration of the digital twin model improved the prediction accuracy by 40%. The finally output continuous production control flow achieved a device synchronization error < 0.1% and a material flow continuity score > 98%.

[0080] S204, input the continuous production control flow into the distributed edge collaborative architecture, dynamically configure computing power resources based on the smart contract in the trusted execution environment, and synchronously optimize energy consumption and output efficiency through the multi-objective decision-making algorithm of chaotic optimization to generate a globally convergent set of factory collaborative control strategies to achieve the collaborative operation of the industrial digital factory.

[0081] Specifically, the continuous production control flow can be input into the distributed edge collaborative architecture. The integrity and compliance of the control instructions are verified based on the smart contract in the trusted execution environment to generate a verification vector encrypted with a digital signature. The distributed edge collaborative architecture consists of multiple edge nodes (such as the intelligent gateways in the factory workshop), and a trusted execution environment (TEE, such as Intel SGX) is deployed on each node. When the continuous production control flow (including device control instructions, material scheduling paths, etc.) arrives at the edge node, compliance verification is first performed through the smart contract. The rules of the smart contract include: Instruction integrity check: Calculate the instruction hash value using the SHA-3 algorithm and compare it with the hash tree pre-stored in the cloud to ensure that the transmission is not tampered with. Compliance check: Verify whether the instruction parameters exceed the device safety threshold (such as the movement speed of the robotic arm ≤ 1.5 m / s), and directly intercept the illegal instructions. Digital signature encryption: Sign the verified instructions using ECDSA (Elliptic Curve Digital Signature Algorithm, secp256k1 curve), and the private key is securely stored by the TEE.

[0082] Verification vector generation: Field 1: Instruction hash value (256 bits); Field 2: Compliance label (0 / 1, 0 indicates violation); Field 3: Timestamp (UTC time, accuracy 1 ms); Field 4: Digital signature (512 bits).

[0083] For example, a certain control instruction "Robotic arm A moves to the coordinates (2.5, 3.1)" generates a vector after verification: [0x3a7d..., 1, 1630000000123, 0x5b9e...]. This vector is broadcast to other edge nodes through 5G URLLC (Ultra-Reliable Low-Latency Communication), and the latency < 10 ms.

[0084] Perform multi-objective decision-making algorithm processing on the verification vector, generate a chaotic sequence to traverse the solution space through Logistic mapping, use the Pareto front fast convergence technology to synchronously optimize energy consumption and output efficiency, and output a preliminary control strategy set; The chaotic optimization algorithm generates a pseudo-random sequence through Logistic mapping for efficient search of the solution space: Definition of the solution space: Energy consumption dimension: Equipment power (0~100% rated value, step size 1%); Efficiency dimension: Output rate (pieces / minute, depending on the equipment linkage timing).

[0085] Initialization of the chaotic sequence: Randomly generate 100 initial solutions (x0∈(0,1)) and map them to the solution space. For example, x0 = 0.37 corresponds to a power of 53% and an output rate of 24 pieces / minute.

[0086] Pareto front screening: Use the NSGA-III algorithm (population size 100, crossover probability 0.8) to evaluate the quality of solutions. The objective functions are: Minimize energy consumption (weight 0.6); Maximize output efficiency (weight 0.4).

[0087] Fast convergence technology: Local search: Perform neighborhood perturbation (±5% parameter fine-tuning) on high-quality solutions in the chaotic sequence (such as energy consumption < 60% and efficiency > 20 pieces / minute); Elite retention: Retain the top 10% of non-dominated solutions in each generation to avoid convergence stagnation.

[0088] The finally output preliminary control strategy set contains 50 groups of solutions. For example: Strategy A: Power 70%, efficiency 28 pieces / minute (high-output mode); Strategy B: Power 45%, efficiency 18 pieces / minute (energy-saving mode).

[0089] Perform implicit state synchronization protocol processing on the preliminary control strategy set, eliminate resource preemption conflicts through the distributed consensus algorithm of conflict residuals, and generate a conflict-free subset of cooperative control instructions; The implicit state synchronization protocol is improved based on the Raft algorithm and includes the following steps: State broadcast: Each edge node encapsulates its local strategy set (such as Strategy A, B) as a state log and broadcasts it to other nodes through the gRPC protocol; Conflict detection: Use the resource occupancy matrix to detect conflicts. For example, if Strategy A of Node 1 requires the robotic arm X to work during the time period t1~t2, while Strategy B of Node 2 needs to occupy the same robotic arm during the same time period, a conflict is determined; Conflict residual calculation: Define the residual Δ = difference in policy benefits (e.g., efficiency of A is 28 vs 18 of B, Δ = 10); Distributed consensus: If Δ > threshold (e.g., 5), the high-benefit policy is executed first; If Δ ≤ threshold, use the PBFT (Practical Byzantine Fault Tolerance) algorithm for voting, and 2 / 3 of the nodes need to reach an agreement.

[0090] Conflict-free instruction subset generation: Round-robin time slicing: For unavoidable conflicts (such as shared devices), allocate time slices according to policy priorities (e.g., policy A occupies 60ms and policy B occupies 40ms); Resource reservation: Reserve 20% of the resource buffer for critical policies (such as instructions involving high-risk devices).

[0091] For example, the finally executed subset after negotiation is: policy A (robot arm X), policy B (robot arm Y), and the conflict is completely eliminated.

[0092] ‌Use the timestamp dynamic interpolation technology to align the time axis of the collaborative control instruction subset, and eliminate the influence of network jitter through the instruction buffer queue constrained by the sliding window, and finally output a globally convergent and time-sequentially consistent factory collaborative control policy set.

[0093] The timestamp dynamic interpolation technology is implemented based on IEEE 1588v2 (PTP Precision Time Protocol): Clock synchronization: The master node (such as the central controller) periodically sends synchronization messages, and the slave nodes calibrate their local clocks, with an error < ±100μs; Interpolation compensation: Perform linear interpolation on the instruction timestamps. For example, the timestamp of instruction 1 is t1 = 1000ms, and that of instruction 2 is 1005ms. If the network delay causes instruction 2 to arrive early, it will be buffered locally until 1005ms before execution; Sliding window mechanism: Window size: 50ms, accommodating 10 instructions (each 5ms); Jitter suppression: If network jitter is detected (such as delay fluctuation > 2ms), the window automatically shrinks to 30ms; Redundancy elimination: Discard expired instructions (such as delay > 50ms).

[0094] Final policy set output: Time-sequential consistency: The time deviation of all devices executing instructions < 1ms; Global convergence: Verify the system stability through the Lyapunov function to ensure that the policy has no oscillation during long-term execution.

[0095] For example, a certain collaborative policy set includes: t = 1000 ms: Robot arm A grabs the material; t = 1003 ms: Conveyor belt B accelerates to 1.2 m / s; t = 1005 ms: Robot arm C assembles parts.

[0096] After all instructions are aligned by timestamps, the actual execution time error is controlled within ±0.8 ms, meeting the industrial real-time requirements.

[0097] It can be seen that by collecting multi-modal discrete source data of the factory, integrating cross-domain features through the dynamic tensor fusion algorithm, generating a globally consistent factory situation fusion tensor; inputting the factory situation fusion tensor into a resource-constrained collaborative decision-making engine, constructing an optimization plan for multi-device collaborative tasks based on the dynamic resource game model, outputting non-competitive multi-node collaborative scheduling instructions; performing abnormal situation recovery processing on the multi-node collaborative scheduling instructions, outputting a continuous production control flow with enhanced stability; inputting the continuous production control flow into a distributed edge collaborative architecture, dynamically configuring computing power resources based on smart contracts in a trusted execution environment, generating a globally convergent set of factory collaborative control strategies, so as to provide a highly robust and adaptive collaborative working method for industrial digital factories, improving production efficiency and reducing energy consumption.

[0098] Another embodiment of the present invention provides an industrial digital factory collaborative working cloud platform. Refer to Figure 3 , the cloud platform may include: A collection module 301, configured to collect multi-modal discrete source data of the factory, integrate cross-domain features through the dynamic tensor fusion algorithm, eliminate the communication clock offset and data sampling frequency difference between devices by using an asynchronous trigger synchronization mechanism, and generate a globally consistent factory situation fusion tensor; An adjustment module 302, configured to input the factory situation fusion tensor into a resource-constrained collaborative decision-making engine, construct an optimization plan for multi-device collaborative tasks based on the dynamic resource game model, adjust the task execution sequence through a priority dynamic allocation algorithm for conflict resolution, and output non-competitive multi-node collaborative scheduling instructions; An analysis module 303, configured to perform abnormal situation recovery processing on the multi-node collaborative scheduling instructions, use a competitive feature refinement network to real-time analyze the device operation status and material flow abnormal features, reconstruct the damaged production sequence through a parameterized spatio-temporal repair operator, and output a continuous production control flow with enhanced stability; An optimization module 304, configured to input the continuous production control flow into a distributed edge collaborative architecture, dynamically configure computing power resources based on smart contracts in a trusted execution environment, synchronously optimize energy consumption and output efficiency through a multi-objective decision-making algorithm with chaotic optimization, and generate a globally convergent set of factory collaborative control strategies to achieve the collaborative work of industrial digital factories.

[0099] It can be seen that the multi-modal discrete source data of the acquisition factory is collected, cross-domain feature integration is carried out through the dynamic tensor fusion algorithm, and a globally consistent factory situation fusion tensor is generated; the factory situation fusion tensor is input into the resource-constrained collaborative decision-making engine, and a multi-device collaborative task optimization scheme is constructed based on the dynamic resource game model, and a non-competitive multi-node collaborative scheduling instruction is output; the abnormal situation recovery process is performed on the multi-node collaborative scheduling instruction, and a continuously enhanced production control flow with enhanced stability is output; the continuously enhanced production control flow is input into the distributed edge collaborative architecture, and the computing power resources are dynamically configured based on the smart contract in the trusted execution environment, and a globally convergent factory collaborative control strategy set is generated, so as to provide a highly robust and adaptive collaborative working method for the industrial digital factory, improve production efficiency and reduce energy consumption.

[0100] An embodiment of the present invention also provides a storage medium, in which a computer program is stored, and the computer program is configured to execute the steps in any one of the above method embodiments when running.

[0101] Specifically, in this embodiment, the above storage medium may be configured to store a computer program for executing the following steps: S201, collect the multi-modal discrete source data of the factory, perform cross-domain feature integration through the dynamic tensor fusion algorithm, use the asynchronous trigger synchronization mechanism to eliminate the communication clock offset and data sampling frequency difference between devices, and generate a globally consistent factory situation fusion tensor; ‌S202, input the factory situation fusion tensor into the resource-constrained collaborative decision-making engine, construct a multi-device collaborative task optimization scheme based on the dynamic resource game model, adjust the task execution sequence through the priority dynamic allocation algorithm for conflict resolution, and output a non-competitive multi-node collaborative scheduling instruction; S203, perform abnormal situation recovery processing on the multi-node collaborative scheduling instruction, use the competitive feature refinement network to real-time analyze the device operation state and material flow abnormal features, reconstruct the damaged production sequence through the parameterized space-time repair operator, and output a continuously enhanced production control flow with enhanced stability; S204, input the continuously enhanced production control flow into the distributed edge collaborative architecture, dynamically configure the computing power resources based on the smart contract in the trusted execution environment, synchronously optimize the energy consumption and output efficiency through the multi-objective decision-making algorithm with chaotic optimization, and generate a globally convergent factory collaborative control strategy set to realize the collaborative work of the industrial digital factory.

[0102] It can be seen that the multi-modal discrete source data of the acquisition factory is collected, cross-domain feature integration is performed through the dynamic tensor fusion algorithm, and a globally consistent factory situation fusion tensor is generated by using the asynchronous trigger synchronization mechanism to eliminate the communication clock offset and data sampling frequency difference between devices; the factory situation fusion tensor is input into the resource-constrained collaborative decision-making engine, a multi-device collaborative task optimization scheme is constructed based on the dynamic resource game model, and the task execution sequence is adjusted through the priority dynamic allocation algorithm of conflict resolution, and a non-competitive multi-node collaborative scheduling instruction is output; the abnormal situation recovery process is performed on the multi-node collaborative scheduling instruction, the running state of the device and the abnormal characteristics of the material flow are analyzed in real time by using the competitive feature refinement network, and the damaged production sequence is reconstructed through the parameterized spatio-temporal repair operator, and a continuous production control flow with enhanced stability is output; the continuous production control flow is input into the distributed edge collaborative architecture, the computing power resources are dynamically configured based on the smart contract in the trusted execution environment, and the energy consumption and output efficiency are synchronously optimized through the multi-objective decision-making algorithm of chaotic optimization, and a globally convergent factory collaborative control strategy set is generated, so as to provide a highly robust and adaptive collaborative working method for the industrial digital factory, improve production efficiency and reduce energy consumption.

[0103] An embodiment of the present invention further provides an electronic device, including a memory and a processor, where a computer program is stored in the memory, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0104] Specifically, the above electronic device may further include a transmission device and an input / output device, where the transmission device is connected to the above processor, and the input / output device is connected to the above processor.

[0105] Specifically, in this embodiment, the above processor may be configured to execute the following steps through a computer program: S201, collect the multi-modal discrete source data of the factory, perform cross-domain feature integration through the dynamic tensor fusion algorithm, and use the asynchronous trigger synchronization mechanism to eliminate the communication clock offset and data sampling frequency difference between devices, and generate a globally consistent factory situation fusion tensor; ‌S202, input the factory situation fusion tensor into the resource-constrained collaborative decision-making engine, construct a multi-device collaborative task optimization scheme based on the dynamic resource game model, and adjust the task execution sequence through the priority dynamic allocation algorithm of conflict resolution, and output a non-competitive multi-node collaborative scheduling instruction; S203, perform an abnormal situation recovery process on the multi-node collaborative scheduling instruction, use the competitive feature refinement network to analyze the running state of the device and the abnormal characteristics of the material flow in real time, and reconstruct the damaged production sequence through the parameterized spatio-temporal repair operator, and output a continuous production control flow with enhanced stability; S204, input the continuous production control flow into the distributed edge collaborative architecture, dynamically configure the computing power resources based on the smart contract in the trusted execution environment, and synchronously optimize the energy consumption and output efficiency through the multi-objective decision-making algorithm of chaotic optimization, and generate a globally convergent factory collaborative control strategy set to realize the collaborative work of the industrial digital factory.

[0106] It can be seen that the multi-modal discrete source data of the acquisition factory is collected, and cross-domain feature integration is carried out through the dynamic tensor fusion algorithm to generate a globally consistent factory situation fusion tensor; the factory situation fusion tensor is input into the resource-constrained collaborative decision-making engine, and a multi-device collaborative task optimization scheme is constructed based on the dynamic resource game model, and a multi-node collaborative scheduling instruction without competition is output; the multi-node collaborative scheduling instruction is processed for abnormal situation recovery, and a continuous production control flow with enhanced stability is output; the continuous production control flow is input into the distributed edge collaborative architecture, and the computing power resources are dynamically configured based on the smart contract in the trusted execution environment to generate a globally convergent factory collaborative control strategy set, so as to provide a highly robust and adaptive collaborative working method for the industrial digital factory, improve production efficiency and reduce energy consumption.

[0107] The structure, features and action effects of the present invention have been described in detail based on the embodiments shown in the drawings. The above is only the preferred embodiment of the present invention, but the present invention is not limited to the scope of implementation shown in the drawings. Any changes made according to the concept of the present invention, or equivalent embodiments modified into equivalent changes, should still be within the protection scope of the present invention when they do not exceed the spirit covered by the specification and the drawings.

Claims

1. An industrial digital factory collaborative working method, characterized in that, The method includes: Collecting multi-modal discrete source data of the factory, integrating cross-domain features through a dynamic tensor fusion algorithm, eliminating communication clock offset and data sampling frequency differences between devices by using an asynchronous trigger synchronization mechanism, and generating a globally consistent factory situation fusion tensor; Inputting the factory situation fusion tensor into a resource-constrained collaborative decision-making engine, constructing an optimization scheme for multi-device collaborative tasks based on a dynamic resource game model, adjusting the task execution sequence through a priority dynamic allocation algorithm for conflict resolution, and outputting a non-competitive multi-node collaborative scheduling instruction; Performing abnormal situation recovery processing on the multi-node collaborative scheduling instruction, using a competitive feature refinement network to real-time analyze the abnormal features of device operation states and material flow, reconstructing the damaged production sequence through a parameterized spatio-temporal repair operator, and outputting a continuously production control flow with enhanced stability; Inputting the continuously production control flow into a distributed edge collaborative architecture, dynamically configuring computing power resources based on smart contracts in a trusted execution environment, synchronously optimizing energy consumption and output efficiency through a multi-objective decision-making algorithm with chaotic optimization, and generating a globally convergent set of factory collaborative control strategies to achieve the collaborative operation of an industrial digital factory.

2. The method according to claim 1, characterized in that The collecting multi-modal discrete source data of the factory, integrating cross-domain features through a dynamic tensor fusion algorithm, eliminating communication clock offset and data sampling frequency differences between devices by using an asynchronous trigger synchronization mechanism, and generating a globally consistent factory situation fusion tensor includes: According to the time-domain waveform of the device vibration spectrum and the spatial coordinates of the material displacement trajectory on the production line in the multi-modal discrete source data of the factory, using an event-triggered pulse encoder to convert the continuous signal into an asynchronous pulse sequence, and generating a multi-modal pulse matrix with spatio-temporal markers; wherein, the encoder suppresses high-frequency noise pulses through a dynamic threshold adjustment mechanism and retains effective operating condition features; Measuring the phase difference between the multi-modal pulse matrix and the environmental electromagnetic field intensity data, constructing a time-varying weight matrix through an event-driven adaptive phase calibration algorithm, dynamically compensating for the clock offset and sampling frequency differences between devices, and outputting a phase-aligned cross-modal pulse tensor; Inputting the cross-modal pulse tensor into a dynamic tensor fusion algorithm, constructing a pulse energy density kernel in the time-space-frequency three domains, separating the core feature sub-spaces of device vibration, material displacement, and electromagnetic field through convolutional tensor decomposition, and eliminating multi-source signal interference; Accumulating pulse energy and performing spatial topology mapping on the core feature sub-spaces, generating a globally consistent factory situation fusion tensor based on a dynamic edge weight allocation algorithm; wherein, the factory situation fusion tensor reflects the coupling situation of the device cluster in real time through a weight update mechanism triggered by pulses.

3. The method according to claim 2, characterized in that, The inputting the factory situation fusion tensor into a resource-constrained collaborative decision-making engine, constructing an optimization scheme for multi-device collaborative tasks based on a dynamic resource game model, adjusting the task execution sequence through a priority dynamic allocation algorithm for conflict resolution, and outputting a non-competitive multi-node collaborative scheduling instruction includes: Input the factory situation fusion tensor into the collaborative decision-making engine with resource constraints, construct a multi-device task revenue matrix based on the dynamic resource game model, quantify the task priority weights through the Nash bargaining solution algorithm, and generate an initial game strategy set; Perform fractal dimension attention mechanism processing on the initial game strategy set, calculate the fractal dimension difference between critical tasks and regular tasks, and forcefully isolate the resource occupancy thresholds of the two types of tasks through an attention mask, and output an attention weight vector with fractal constraints; According to the attention weight vector, adopt a priority dynamic allocation algorithm for conflict resolution to adjust the task execution sequence. Among them, the priority dynamic allocation algorithm optimizes the stability of the task queue through the Lyapunov drift plus penalty function; Perform strategy evolution algorithm processing on the adjusted task execution sequence, and co-optimize the fault tolerance of the scheduling path through Monte Carlo tree search and deep policy network, and output a multi-node collaborative scheduling instruction without competition.

4. The method according to claim 3, wherein Perform abnormal situation recovery processing on the multi-node collaborative scheduling instruction, use a competitive feature refinement network to real-time analyze the device operation state and material flow abnormal features, and reconstruct the damaged production sequence through a parameterized spatio-temporal repair operator, and output a continuous production control flow with enhanced stability, including: Perform virtual abnormal sample generation processing on the collaborative scheduling instruction, use a generative adversarial network to simulate extreme working conditions such as equipment downtime and material blockage, and construct an enhanced training data set containing abnormal labels; Input the data set into the competitive feature refinement network, separate the device operation state and material flow abnormal features through a dual-channel adversarial training mechanism, and output a high-resolution abnormal feature vector; among them, the network enhances the abnormal detection sensitivity through game theory-driven feature distillation technology; Perform parameterized spatio-temporal repair operator processing on the abnormal feature vector, construct a spatio-temporal manifold completion model for the damaged production sequence, and reconstruct the continuous production sequence through a geodesic constraint-based energy minimization algorithm; Use the reconstructed continuous production sequence, adopt a reverse error compensation mechanism to synchronously correct the operation deviation between the physical device actuator and the digital twin model. Among them, the reverse error compensation mechanism realizes real-time bidirectional calibration of control instructions through implicit gradient propagation, and outputs a continuous production control flow with enhanced stability.

5. The method according to claim 4, wherein Input the continuous production control flow into a distributed edge collaborative architecture, dynamically configure computing power resources based on smart contracts in a trusted execution environment, and synchronously optimize energy consumption and output efficiency through a chaotic optimization multi-objective decision-making algorithm, and generate a globally convergent factory collaborative control strategy set, including: Input the continuous production control flow into a distributed edge collaborative architecture, verify the integrity and compliance of control instructions based on smart contracts in a trusted execution environment, and generate a verification vector encrypted with a digital signature; Perform chaotic optimization multi-objective decision-making algorithm processing on the verification vector, generate a chaotic sequence to traverse the solution space through Logistic mapping, and use Pareto front fast convergence technology to synchronously optimize energy consumption and output efficiency, and output a preliminary control strategy set; Perform implicit state synchronization protocol processing among edge nodes on the preliminary control strategy set, eliminate resource preemption conflicts through a distributed consensus algorithm for conflict residuals, and generate a conflict-free subset of collaborative control instructions; Use timestamp dynamic interpolation technology to align the subset of collaborative control instructions on the time axis, eliminate the impact of network jitter through an instruction buffer queue with sliding window constraints, and finally output a globally convergent and temporally consistent factory collaborative control strategy set.

6. An industrial digital factory collaborative working cloud platform, characterized in that, The cloud platform includes: A collection module for collecting factory multi-modal discrete source data, integrating cross-domain features through a dynamic tensor fusion algorithm, eliminating communication clock offsets and data sampling frequency differences between devices using an asynchronous trigger synchronization mechanism, and generating a globally consistent factory situation fusion tensor; An adjustment module for inputting the factory situation fusion tensor into a collaborative decision-making engine with resource constraints, constructing an optimization plan for multi-device collaborative tasks based on a dynamic resource game model, adjusting the task execution sequence through a priority dynamic allocation algorithm for conflict resolution, and outputting non-competitive multi-node collaborative scheduling instructions; An analysis module for performing abnormal situation recovery processing on the multi-node collaborative scheduling instructions, using a competitive feature refinement network to real-time analyze device operation states and abnormal features of material flows, reconstructing damaged production sequences through parametric spatio-temporal repair operators, and outputting a continuous production control flow with enhanced stability; An optimization module for inputting the continuous production control flow into a distributed edge collaboration architecture, dynamically configuring computing power resources based on smart contracts in a trusted execution environment, synchronously optimizing energy consumption and output efficiency through a multi-objective decision-making algorithm with chaotic optimization, and generating a globally convergent factory collaborative control strategy set to achieve collaborative work in an industrial digital factory.

7. The cloud platform according to claim 6, wherein The collection module is specifically used for: According to the time-domain waveform of the device vibration spectrum and the spatial coordinates of the material displacement trajectory on the production line in the factory multi-modal discrete source data, use an event-triggered pulse encoder to convert the continuous signal into an asynchronous pulse sequence, and generate a multi-modal pulse matrix with spatio-temporal tags; among them, the encoder suppresses high-frequency noise pulses through a dynamic threshold adjustment mechanism and retains effective operating condition features; Measure the phase difference between the multi-modal pulse matrix and the environmental electromagnetic field intensity data, construct a time-varying weight matrix through an event-driven adaptive phase calibration algorithm, dynamically compensate for clock offsets and sampling frequency differences between devices, and output a phase-aligned cross-modal pulse tensor; Input the cross-modal pulse tensor into a dynamic tensor fusion algorithm, construct a pulse energy density kernel in the time-space-frequency three domains, separate the core feature subspaces of device vibration, material displacement, and electromagnetic field through convolutional tensor decomposition, and eliminate multi-source signal interference; Perform pulse energy accumulation and spatial topology mapping on the core feature subspaces, and generate a globally consistent factory situation fusion tensor based on a dynamic edge weight allocation algorithm; among them, the factory situation fusion tensor reflects the coupling situation of the device cluster in real time through a weight update mechanism triggered by pulses.

8. The cloud platform according to claim 7, characterized in that, The adjustment module is specifically used for: The factory situation fusion tensor is input into the collaborative decision-making engine of resource constraints, a multi-device task benefit matrix is ​​constructed based on a dynamic resource game model, task priority weights are quantified through a Nash bargaining solution algorithm, and an initial game strategy set is generated; ‌ The initial game strategy set is processed by a fractal dimension attention mechanism, the fractal dimension difference between the key task and the conventional task is calculated, the resource occupation thresholds of the two types of tasks are forcibly isolated by an attention mask, and the attention weight vector of the fractal constraint is output; ‌According to the attention weight vector, a priority dynamic allocation algorithm for conflict resolution is used to adjust the task execution sequence, wherein the priority dynamic allocation algorithm optimizes the stability of the task queue through a Lyapunov drift plus penalty function; The adjusted task execution sequence is processed by the strategy evolution algorithm, and the fault tolerance of the scheduling path is optimized through the Monte Carlo tree search and the deep strategy network, and non-competitive multi-node collaborative scheduling instructions are output.

9. A storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method according to any one of claims 1 to 5 when executed.

10. An electronic device, comprising a memory and a processor, characterized in that, A computer program is stored in the memory, and the processor is configured to run the computer program to perform the method according to any one of claims 1 to 5.

Images