Robot data intelligent cleaning method and system based on neural network

Abstract

The application discloses a kind of robot data intelligent cleaning method and system based on neural network, it is related to data processing technical field.The method includes: robot data is collected in edge computing node and is managed by adaptive ring buffer queue;Using noise reduction desensitization feature extraction model, denoising and desensitization are completed synchronously at feature level;Using structured reconstruction model, structured standard robot data is generated;Confidence score is calculated to filter low-score samples for local fine-tuning, and local model parameter update amount and auxiliary update data are collected;In the center server, adaptive aggregation weight is calculated based on auxiliary update data, and global model is updated by aggregating local model parameter update amount.The application realizes efficient cleaning of robot data on edge side and robust updating of global model by constructing noise reduction desensitization feature extraction model and structured reconstruction model, combined with adaptive ring buffer queue and adaptive weighted aggregation mechanism.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for intelligent cleaning of robot data based on neural networks. Background Technology

[0002] With the rapid development of the Industrial Internet of Things (IIoT) and intelligent manufacturing, robots are widely used in various complex industrial scenarios. During robot operation, massive amounts of multimodal sensor data (such as vision, force, and pose) are generated. However, due to the complex and ever-changing industrial environment and the physical limitations of the sensors themselves, the raw data collected often contains a large amount of noise, missing information, and privacy-sensitive data. Directly using this unprocessed raw data for robot control, status monitoring, or model training can not only lead to decreased model performance and decision-making errors but may also cause serious data privacy leaks.

[0003] Existing robot data cleaning methods typically employ traditional filtering and denoising algorithms or rule-based data anonymization methods. These methods have the following shortcomings:

[0004] 1) Traditional denoising methods are difficult to effectively handle complex noise in non-stationary, high-dimensional multimodal data;

[0005] 2) Conventional desensitization methods usually directly mask or disturb the original data space, which can easily destroy the original spatiotemporal correlation characteristics of the data, resulting in a significant reduction in the usability of the cleaned data;

[0006] 3) In the architecture of edge computing and cloud collaboration, the data quality and environmental distribution collected by different edge computing nodes vary. Traditional federated learning or data aggregation methods lack adaptive evaluation of node data quality and model parameter stability, resulting in insufficient robustness of global model updates.

[0007] Therefore, there is an urgent need for a robot data intelligent cleaning method that can simultaneously achieve efficient denoising and secure desensitization at the feature level, and combine graph neural networks for structured reconstruction and adaptive federated aggregation. Summary of the Invention

[0008] This invention provides a method and system for intelligent cleaning of robot data based on neural networks. By constructing a noise reduction and desensitization feature extraction model and a structured reconstruction model, combined with an adaptive circular buffer queue and an adaptive weighted aggregation mechanism, this invention achieves efficient cleaning of robot data at the edge and robust updating of the global model.

[0009] In a first aspect, embodiments of the present invention provide a robot data intelligent cleaning method based on neural networks, the method comprising:

[0010] At the edge computing node, robot data generated during robot operation is collected, and the robot data is managed based on an adaptive circular buffer queue management mechanism;

[0011] The robot data is input into a noise reduction and desensitization feature extraction model based on a neural network to generate an intermediate feature tensor that performs both noise reduction and desensitization at the feature level.

[0012] The intermediate feature tensor is input into a structured reconstruction model built on a neural network to generate structured standard robot data with semantic labels and topological structure.

[0013] Calculate the confidence score of the structured standard robot data, perform local model fine-tuning based on the local training sample set whose confidence scores are less than the score threshold, collect the local model parameter update amount and auxiliary update data, and encrypt and upload them to the central server;

[0014] At the central server, based on the adaptive weighted aggregation mechanism, the update amounts of local model parameters uploaded by several edge computing nodes are aggregated according to the auxiliary update data, and the generated updated global model parameters are distributed to all edge computing nodes.

[0015] The technical solution provided in this application has at least the following beneficial effects:

[0016] Through an adaptive circular buffer queue management mechanism combined with timeliness-information entropy scoring, high-value data can be intelligently filtered and retained under conditions of limited edge computing resources and sudden data surges, effectively preventing data backlog and loss of key information. By constructing a noise reduction and desensitization feature extraction model that incorporates adversarial game theory, dynamic adaptive masking operations are performed in the feature space, effectively preventing privacy reconstruction attacks while maximizing the retention of effective information required for cleaning, achieving simultaneous noise reduction and desensitization. By constructing a structured reconstruction model and using graph neural networks combined with time-varying decaying edge weights for dynamic topology reconstruction, the spatiotemporal correlation between robot states in industrial scenarios can be accurately captured, generating structured standard data with rich semantics and topological structure. Abnormal data is filtered and fine-tuned locally based on confidence scores, and Fisher information regularization and topology preservation penalties are introduced into the total loss function, effectively overcoming catastrophic forgetting and ensuring topological consistency before and after desensitization. The central server calculates adaptive aggregation weights based on multi-dimensional auxiliary data, improving the robustness and convergence speed of global model updates.

[0017] In one alternative implementation, robot data generated during robot operation is collected at the edge computing node, and managed based on an adaptive circular buffer queue management mechanism, including:

[0018] At the edge computing node, raw robot data generated during robot operation is collected and preprocessed to obtain preprocessed robot data.

[0019] Based on the PTP protocol and a local temperature-controlled crystal oscillator, the time deviation of the preprocessed robot data is calculated, and nanosecond-level timestamps are marked on the preprocessed robot data according to the time deviation.

[0020] Based on an adaptive circular buffer queue management mechanism, a multidimensional data circular buffer queue is maintained, and the arrival rate of raw robot data is collected.

[0021] If the arrival rate is greater than the processing rate of the edge computing node, the preprocessed robot data with timestamps will be stored in a multidimensional data circular buffer queue, and the corresponding timeliness-information entropy score will be calculated.

[0022] If a queue feedback instruction is received, the data retention priority of the corresponding preprocessed robot data with timestamps will be adjusted to the highest level.

[0023] If the number of data in the multidimensional data circular buffer queue reaches the preset capacity, the preprocessed robot data with the lowest timeliness-information entropy score or the lowest data retention priority with timestamps will be discarded, and the latest preprocessed robot data with timestamps will be stored to realize the management of robot data.

[0024] In one optional implementation, the noise reduction and desensitization feature extraction model includes a noise reduction autoencoder neural network, a feature space dynamic suppression module, a privacy reconstruction attack network, and a decoder.

[0025] The noise reduction autoencoder neural network includes several feature extraction layers, and a sensitive feature mask layer is set at the output of the feature extraction layer of the last layer of the noise reduction autoencoder neural network.

[0026] The structured reconstruction model includes a spatiotemporal joint attention network, a cross-modal temporal weighted stitching module, and a graph neural network;

[0027] The graph neural network is connected to a pre-built industrial scene data ontology graph.

[0028] In one alternative implementation, robot data is input into a denoising and desensitization feature extraction model built on a neural network to generate an intermediate feature tensor that simultaneously performs denoising and desensitization at the feature level, including:

[0029] The latest preprocessed robot data with timestamps, received by the edge computing node or stored in the multidimensional data circular buffer queue, is input into the noise reduction and desensitization feature extraction model built on a neural network.

[0030] Several feature extraction layers are used to extract features from the preprocessed robot data with timestamps in sequence to generate the final high-dimensional feature tensor.

[0031] During the inference phase, the trained sensitive feature mask layer is used to calculate the attention score of each spatial region in the final high-dimensional feature tensor based on the multi-head cross attention mechanism, and a dynamic adaptive mask map is generated through the Sigmoid function.

[0032] The feature space dynamic suppression module is used to perform Hadamard product operation on the final high-dimensional feature tensor and the dynamic adaptive mask map to generate a desensitized high-dimensional feature tensor for dynamically suppressed sensitive privacy regions.

[0033] The decoder is used to decode the desensitized high-dimensional feature tensor to generate an intermediate feature tensor that simultaneously performs denoising and desensitization at the feature level.

[0034] In one alternative implementation, intermediate feature tensors are input into a structured reconstruction model built on a neural network to generate structured standard robot data with semantic labels and topological structure, including:

[0035] The intermediate feature tensor is input into a structured reconstruction model built on a neural network;

[0036] Using a spatiotemporal joint attention network, based on a learnable spatial transformation matrix, a six-degree-of-freedom coordinate system is established for the sensor disparities of different modalities in the intermediate feature tensor to obtain a unified feature tensor.

[0037] Using the cross-modal temporal weighted concatenation module, based on the preset cross-modal temporal sliding window and the learnable attention weight matrix, the temporal correlation weight of each modality in the unified feature tensor within the same cross-modal temporal sliding window is calculated.

[0038] Based on the temporal correlation weights, the feature components of all modes of the unified feature tensor are concatenated at the feature level to obtain the weighted concatenated feature tensor.

[0039] Using graph neural networks, weighted splicing feature tensors are mapped to a pre-built industrial scene data ontology graph. Based on the time-varying decay edge weights at the current moment, dynamic topology reconstruction and node feature updates are performed to generate structured standard robot data with semantic labels and topological structures.

[0040] In one alternative implementation, a graph neural network is used to map the weighted concatenated feature tensor to a pre-built industrial scene data ontology graph. Based on the time-varying decaying edge weights at the current moment, dynamic topology reconstruction and node feature updates are performed to generate structured standard robot data with semantic labels and topological structure, including:

[0041] The graph structure is constructed by using the robots corresponding to the weighted splicing feature tensors as the graph nodes of the pre-constructed industrial scene data ontology graph, and the physical operation state relationships between robots as the graph edges of the industrial scene data ontology graph.

[0042] Obtain the state update time difference between adjacent graph nodes at the current time, and calculate the time-varying decay edge weight of the graph edge at the current time by combining it with a preset decay coefficient.

[0043] If the weight of the time-varying decay edge is less than the preset topology disconnection threshold, the dynamic disconnection of the corresponding graph edge is triggered, and the message transmission of the graph edge is stopped. If the feature cosine similarity between any two non-adjacent graph nodes calculated based on the unified feature tensor is greater than the preset topology association threshold, dynamic edge construction is triggered to build a graph structure between the two non-adjacent graph nodes, and the reconstructed industrial scene data ontology graph is obtained.

[0044] Based on the time-varying decay edge weights and the message passing mechanism of the graph neural network, the node features of adjacent graph nodes are weighted and aggregated to update the feature representation of the current node, thus obtaining the updated node features.

[0045] The updated node features are decoded and mapped to generate corresponding semantic labels. Combined with the topological connection relationship between the graph nodes and graph edges, structured standard robot data with semantic labels and topological structure is output.

[0046] In one optional implementation, a confidence score for the structured standard robot data is calculated. Based on the local training sample set whose confidence scores are less than a threshold, local model fine-tuning is performed. Local model parameter updates and auxiliary update data are collected and encrypted before being uploaded to the central server, including:

[0047] Structured standard robot data is input into a quality assessment classifier built on a multilayer perceptron. The corresponding high-dimensional feature distribution anomaly probability is output through the Softmax activation function to generate a confidence score for the current data.

[0048] If the confidence score is greater than or equal to the score threshold, the structured standard robot data is deemed qualified, the structured standard robot data is stored in the cloud, and the robot data intelligent cleaning process ends.

[0049] If the confidence score is less than the score threshold, it is determined that there is a sudden change in the distribution of environmental noise in the structured standard robot data. A queue feedback instruction is generated to adjust the data retention priority of the preprocessed robot data corresponding to the structured standard robot data and add it to the new sample set, and then proceed to the next step.

[0050] Until the local training cycle is reached, data augmentation operations are performed on the new sample set to obtain a local training sample set including positive and negative sample pairs.

[0051] Based on the total loss function, the backpropagation algorithm is used to input the local training sample set into the local noise reduction and desensitization feature extraction model and the structured reconstruction model for local fine-tuning, and calculate the local model parameter update amount, including the parameter update amount of the noise reduction and desensitization feature extraction model and the parameter update amount of the structured reconstruction model.

[0052] Collect auxiliary update data for this round of local fine-tuning, mark the local model parameter update amount and auxiliary update data with upload timestamps, and encrypt and upload them to the central server.

[0053] In one alternative implementation, the auxiliary update data includes the sample size of the local training sample set used in this round of local fine-tuning, the average confidence score, the accuracy on the validation set, and the gradient second moment oscillation data calculated based on the gradient vector of this round of local fine-tuning.

[0054] In one optional implementation, at the central server, based on an adaptive weighted aggregation mechanism, the update amounts of local model parameters uploaded by several edge computing nodes are aggregated according to auxiliary update data, and the generated updated global model parameters are distributed to all edge computing nodes, including:

[0055] On the central server, based on a preset sliding time window, the update volume of all local model parameters and auxiliary update data with upload timestamps in the same preset sliding time window are collected, and the corresponding number of uploaded data is recorded.

[0056] If the amount of uploaded data exceeds the threshold, proceed to the next step; otherwise, continue waiting for the edge computing node to upload data.

[0057] Based on the adaptive weighted aggregation mechanism, the weight influence factor of each edge computing node is calculated according to the auxiliary update data. The weight influence factor includes the sample size factor, the data quality and reliability factor, and the parameter stability penalty factor.

[0058] The adaptive aggregation weights of the adaptive weighted aggregation are obtained by multiplying the sample size factor, data quality and reliability factor, and parameter stability penalty factor and normalizing them.

[0059] Based on the adaptive aggregation weight, the local model parameter update amounts uploaded by several edge computing nodes are aggregated to generate updated global model parameters;

[0060] The generated updated global model parameters are distributed to all edge computing nodes.

[0061] Secondly, embodiments of the present invention provide a robot data intelligent cleaning system based on neural networks, used to implement a robot data intelligent cleaning method, the system comprising:

[0062] The robot data acquisition unit is used to collect robot data generated during robot operation at the edge computing node, and manage the robot data based on an adaptive circular buffer queue management mechanism.

[0063] The noise reduction and desensitization feature extraction unit is used to input robot data into the noise reduction and desensitization feature extraction model based on neural network to generate an intermediate feature tensor that completes noise reduction and desensitization at the feature level.

[0064] The structured reconstruction unit is used to input intermediate feature tensors into a structured reconstruction model built on a neural network to generate structured standard robot data with semantic labels and topological structures.

[0065] The local model fine-tuning unit is used to calculate the confidence score of structured standard robot data. Based on the local training sample set with confidence scores below the score threshold, it performs local model fine-tuning, collects local model parameter updates and auxiliary update data, and encrypts and uploads them to the central server.

[0066] The adaptive weighted aggregation unit is used on the central server to aggregate the local model parameter update amounts uploaded by several edge computing nodes based on the adaptive weighted aggregation mechanism, and then distribute the generated updated global model parameters to all edge computing nodes.

[0067] A third aspect of this invention provides an electronic device, which includes:

[0068] At least one processor; and a memory communicatively connected to the at least one processor; wherein,

[0069] The memory stores instructions that can be executed by at least one processor, such that the at least one processor can perform the method proposed in the first aspect of the present invention.

[0070] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described in the first aspect of the present invention. Attached Figure Description

[0071] Figure 1 This is a schematic diagram of the electronic device structure of the hardware operating environment involved in the embodiments of the present invention;

[0072] Figure 2This is a flowchart illustrating the steps of a robot data intelligent cleaning method based on neural networks, as provided in an embodiment of the present invention.

[0073] Figure 3 This is a functional unit diagram of a robot data intelligent cleaning system based on neural networks provided in an embodiment of the present invention. Detailed Implementation

[0074] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0075] The present invention will be further described below with reference to the accompanying drawings.

[0076] Reference Figure 1 , Figure 1 This is a schematic diagram of the electronic device structure of the hardware operating environment involved in the embodiments of the present invention.

[0077] like Figure 1 As shown, the electronic device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be a high-speed random access memory (RAM) or a stable non-volatile memory (NVM), such as a disk drive. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0078] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on the electronic device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0079] like Figure 1As shown, the memory 1005, which serves as a storage medium, may include an operating system, a data storage module, a network communication module, a user interface module, and an electronic program for a neural network-based intelligent robot data cleaning system.

[0080] exist Figure 1 In the electronic device shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in the electronic device of the present invention can be set in the electronic device. The electronic device calls the electronic program of the robot data intelligent cleaning system based on neural network stored in the memory 1005 through the processor 1001, and executes the robot data intelligent cleaning method based on neural network provided in the embodiment of the present invention.

[0081] Reference Figure 2 The present invention provides a method for intelligent cleaning of robot data based on neural networks, the method comprising:

[0082] S201: At the edge computing node, robot data generated during robot operation is collected, and the robot data is managed based on an adaptive circular buffer queue management mechanism;

[0083] S202: Input robot data into a noise reduction and desensitization feature extraction model based on a neural network to generate an intermediate feature tensor that simultaneously performs noise reduction and desensitization at the feature level;

[0084] S203: Input the intermediate feature tensor into a structured reconstruction model based on a neural network to generate structured standard robot data with semantic labels and topological structure;

[0085] S204: Calculate the confidence score of the structured standard robot data, perform local model fine-tuning based on the local training sample set whose confidence scores are less than the score threshold, collect the local model parameter update amount and auxiliary update data, and encrypt and upload them to the central server.

[0086] S205: On the central server, based on the adaptive weighted aggregation mechanism, the update amounts of local model parameters uploaded by several edge computing nodes are aggregated according to the auxiliary update data, and the generated updated global model parameters are distributed to all edge computing nodes.

[0087] The technical solution provided in this application has at least the following beneficial effects:

[0088] Through an adaptive circular buffer queue management mechanism combined with timeliness-information entropy scoring, high-value data can be intelligently filtered and retained under conditions of limited edge computing resources and sudden data surges, effectively preventing data backlog and loss of key information. By constructing a noise reduction and desensitization feature extraction model that incorporates adversarial game theory, dynamic adaptive masking operations are performed in the feature space, effectively preventing privacy reconstruction attacks while maximizing the retention of effective information required for cleaning, achieving simultaneous noise reduction and desensitization. By constructing a structured reconstruction model and using graph neural networks combined with time-varying decaying edge weights for dynamic topology reconstruction, the spatiotemporal correlation between robot states in industrial scenarios can be accurately captured, generating structured standard data with rich semantics and topological structure. Abnormal data is filtered and fine-tuned locally based on confidence scores, and Fisher information regularization and topology preservation penalties are introduced into the total loss function, effectively overcoming catastrophic forgetting and ensuring topological consistency before and after desensitization. The central server calculates adaptive aggregation weights based on multi-dimensional auxiliary data, improving the robustness and convergence speed of global model updates.

[0089] In one alternative implementation, robot data generated during robot operation is collected at the edge computing node, and managed based on an adaptive circular buffer queue management mechanism, including:

[0090] S2011: At the edge computing node, the raw robot data generated during the robot's operation is collected, and the raw robot data is preprocessed to obtain preprocessed robot data.

[0091] In this embodiment, at the robot's work site, the edge computing node collects the original multimodal data of the robot's operation process through multiple sensors (including RGB-D cameras, LiDAR, six-axis torque sensors, joint encoders, etc.) at a fixed frequency (e.g., 20Hz for vision, 100Hz for torque).

[0092] The raw robot data is preprocessed as follows: for visual data, lens distortion correction and histogram equalization are performed; for torque data, moving average filtering is used to remove high-frequency glitches; for joint angle data, missing value interpolation is performed to obtain the preprocessed robot data.

[0093] S2012: Based on the Precision Time Protocol (PTP) and a local temperature-controlled crystal oscillator, calculate the time deviation of the preprocessed robot data, and mark the preprocessed robot data with nanosecond-level timestamps according to the time deviation;

[0094] In this embodiment, in order to ensure strict time synchronization of multimodal data, the edge computing node is configured with a network card that supports the IEEE1588 PTP protocol and combined with a local temperature-controlled crystal oscillator to provide a high-precision clock source; the master clock periodically sends Sync messages, and the edge computing node, as a slave clock, calculates the round-trip time of the message, obtains the time deviation, and dynamically adjusts the local clock, thereby giving each frame of preprocessed robot data a unified timestamp with nanosecond-level precision.

[0095] S2013: Based on an adaptive circular buffer queue management mechanism, maintain a multi-dimensional data circular buffer queue and collect the arrival rate of raw robot data;

[0096] S2014: If the arrival rate is greater than the processing rate of the edge computing node, the preprocessed robot data with timestamps will be stored in a multidimensional data circular buffer queue, and the corresponding timeliness-information entropy score will be calculated using the following formula:

[0097]

[0098] In the formula, The timeliness-information entropy score for the i-th data in the multidimensional data circular buffer queue; i is the data index; This refers to the scoring weighting coefficient; It is a natural constant; The current system time; The timestamp of the i-th data; This is the time-related decay coefficient; The information entropy function reflects the amount of information in the data. For image data, the entropy of its grayscale histogram is calculated; for one-dimensional time series data, the entropy of its probability distribution is calculated. The larger the information entropy, the more abnormal or critical operating condition information the data contains. This refers to the i-th data item in a multidimensional data circular buffer queue.

[0099] In this embodiment, due to network fluctuations and computing resource limitations in the industrial field, sudden data arrivals can easily cause memory overflow in edge computing nodes. Therefore, a multi-dimensional data circular buffer queue is maintained in the memory of the edge computing node, with a queue length of 1024. The arrival rate of the raw data and the processing rate of the edge computing node are monitored in real time. If the arrival rate is less than or equal to the processing rate of the edge computing node, the data is directly entered into the subsequent model inference.

[0100] S2015: If a queue feedback instruction is received, the data retention priority of the corresponding preprocessed robot data with timestamps will be adjusted to the highest level;

[0101] In this embodiment, the retention priority is 0 by default. If a queue feedback instruction is received, the retention priority of the original data pointed to by the instruction is adjusted to 1 (highest) to ensure that key abnormal mutation samples can still be retained for subsequent model fine-tuning when data is overloaded.

[0102] S2016: If the number of data in the multidimensional data circular buffer queue reaches the preset capacity, discard the preprocessed robot data with the lowest timeliness-information entropy score or the lowest data retention priority, and store the latest preprocessed robot data with the latest time stamp, thereby realizing the management of robot data.

[0103] In one optional implementation, the noise reduction and desensitization feature extraction model includes a noise reduction autoencoder neural network, a feature space dynamic suppression module, a privacy reconstruction attack network, and a decoder.

[0104] In this embodiment, the encoder of the noise reduction autoencoder neural network consists of 4 one-dimensional convolutional layers and 2 Transformer encoder layers. During the forward propagation process, Gaussian noise is first added to the input data, and then local features are extracted through the convolutional layers, global dependencies are extracted through the Transformer layers, and finally the final high-dimensional feature tensor is output in the last feature extraction layer.

[0105] The noise reduction autoencoder neural network includes several feature extraction layers, and a sensitive feature mask layer is set at the output of the feature extraction layer of the last layer of the noise reduction autoencoder neural network.

[0106] The privacy reconstruction attack network is used to attempt to reconstruct the original sensitive region from the initial sensitivity map output by the sensitive feature mask layer, and to form an adversarial game with the feature space dynamic suppression module to dynamically adjust the mask strength.

[0107] In the adversarial game phase, a sensitive feature mask layer is used. Based on the multi-head cross-attention mechanism, the attention scores of each spatial region in the final high-dimensional feature tensor of the adversarial example are calculated, and the initial sensitivity map is generated by the Sigmoid function.

[0108] The feature space dynamic suppression module is used to perform Hadamard product operation on the final high-dimensional feature tensor of the adversarial example and the initial sensitivity map to generate the initial desensitized high-dimensional feature tensor of the dynamically suppressed sensitive privacy region.

[0109] The privacy-reconstruction attack network is used to reconstruct the initial desensitized feature tensor to obtain the reconstructed sensitive feature tensor.

[0110] Calculate the reconstruction loss of the sensitive regions corresponding to the reconstructed sensitive feature tensor and the final high-dimensional feature tensor, and based on the reconstruction loss, dynamically adjust the mask intensity of the initial sensitivity map to generate a dynamic adaptive mask map.

[0111] For regions where the reconstruction loss is greater than a preset reconstruction threshold, the mask is enhanced; for regions where the reconstruction loss is less than the preset reconstruction threshold, the mask is weakened, so as to retain the information required for cleaning and generate a dynamic adaptive mask map of adversarial examples.

[0112] The structured reconstruction model includes a spatiotemporal joint attention network, a cross-modal temporal weighted stitching module, and a graph neural network;

[0113] The graph neural network is connected to a pre-built industrial scene data ontology graph.

[0114] In one alternative implementation, robot data is input into a denoising and desensitization feature extraction model built on a neural network to generate an intermediate feature tensor that simultaneously performs denoising and desensitization at the feature level, including:

[0115] S2021: Input the latest preprocessed robot data with timestamps received by the edge computing node or stored in the multidimensional data circular buffer queue into the noise reduction and desensitization feature extraction model built on the neural network.

[0116] S2022: Use several feature extraction layers to extract features from the preprocessed robot data with timestamps in sequence to generate the final high-dimensional feature tensor;

[0117] S2023: During the inference phase, the trained sensitive feature mask layer is used to calculate the attention score of each spatial region in the final high-dimensional feature tensor based on the multi-head cross attention mechanism, and a dynamic adaptive mask map is generated through the Sigmoid function.

[0118] S2024: Using the feature space dynamic suppression module, a Hadamard product operation is performed on the final high-dimensional feature tensor and the dynamic adaptive mask map to generate a desensitized high-dimensional feature tensor for dynamically suppressed sensitive privacy regions. The formula is as follows:

[0119]

[0120] In the formula, This is the final desensitized high-dimensional feature tensor; For the final layer high-dimensional feature tensor; For dynamic adaptive masking patterns; This is the Hadamard product operator;

[0121] S2025: Use a decoder to decode the desensitized high-dimensional feature tensor and generate an intermediate feature tensor that simultaneously performs denoising and desensitization at the feature level.

[0122] In one alternative implementation, intermediate feature tensors are input into a structured reconstruction model built on a neural network to generate structured standard robot data with semantic labels and topological structure, including:

[0123] S2031: Input the intermediate feature tensor into the structured reconstruction model built on a neural network;

[0124] S2032: Using a spatiotemporal joint attention network, based on a learnable spatial transformation matrix, a six-degree-of-freedom coordinate system is established for the sensor disparities of different modalities in the intermediate feature tensor to obtain a unified feature tensor;

[0125] In this embodiment, since different modal sensors (such as left-eye camera, right-eye camera, and lidar) are installed in different positions, there is parallax. The network is based on a learnable spatial transformation matrix (6 degrees of freedom parameters: 3 translations and 3 rotations). Through spatial transformation, it performs affine transformation on the feature maps of different modalities and unifies them into a reference coordinate system centered on the robot to obtain a unified feature tensor.

[0126] S2033: Using a cross-modal temporal weighted concatenation module, based on a preset cross-modal temporal sliding window and a learnable attention weight matrix, calculate the temporal correlation weight of each modality in the unified feature tensor within the same cross-modal temporal sliding window;

[0127] S2034: Based on the temporal correlation weight, perform feature-level weighted concatenation on the feature components of all modes of the unified feature tensor to obtain a weighted concatenated feature tensor.

[0128] S2035: Using a graph neural network, weighted splicing feature tensors are mapped to a pre-built industrial scene data ontology graph. Based on the time-varying decay edge weights at the current moment, dynamic topology reconstruction and node feature updates are performed to generate structured standard robot data with semantic labels and topological structures.

[0129] In one alternative implementation, a graph neural network is used to map the weighted concatenated feature tensor to a pre-built industrial scene data ontology graph. Based on the time-varying decaying edge weights at the current moment, dynamic topology reconstruction and node feature updates are performed to generate structured standard robot data with semantic labels and topological structure, including:

[0130] S20351: Using the robots corresponding to the weighted splicing feature tensors as the graph nodes of the pre-constructed industrial scene data ontology graph, and using the physical running state relationships between robots as the graph edges of the industrial scene data ontology graph, a graph structure is constructed.

[0131] S20352: Obtain the state update time difference between adjacent graph nodes at the current time, and calculate the time-varying decay edge weight of the graph edge at the current time in combination with the preset decay coefficient. The formula is as follows:

[0132]

[0133] In the formula, Let be the time-varying decay edge weights of the adjacent u-th and v-th graph nodes at time t; t is the time index. It is a natural constant; The attenuation coefficient; represents the time difference between the state updates of adjacent graph nodes u and v; u and v are graph node indices.

[0134] S20353: If the weight of the time-varying decay edge is less than the preset topology disconnection threshold, the dynamic disconnection of the corresponding graph edge is triggered, and the message transmission of the graph edge is stopped. If the feature cosine similarity between any two non-adjacent graph nodes calculated based on the unified feature tensor is greater than the preset topology association threshold, dynamic edge construction is triggered to build a graph structure between the two non-adjacent graph nodes and obtain the reconstructed industrial scene data ontology graph.

[0135] S20354: Based on the time-varying decay edge weights and the message passing mechanism of the graph neural network, the node features of adjacent graph nodes are weighted and aggregated to update the feature representation of the current node, resulting in the updated node features, as shown in the formula:

[0136]

[0137] In the formula, The updated node features of the v-th graph node in the (l+1)-th iteration of the graph convolutional network in the graph neural network; Non-linear activation functions (such as ReLU, Sigmoid, etc.) are used to enhance the expressive power of the model and maintain the stability of the feature vector values; Let v be the set of neighboring nodes of the v-th graph node; For the set of adjacent graph nodes The updated node features of the u-th graph node in the l-th iteration of the graph neural network; It is a learnable weight matrix based on relationship type (such as cooperative transport, obstacle avoidance); Let be the time-varying decay edge weights of the adjacent u-th and v-th graph nodes at time t; is a normalization constant (usually taken as the degree of the v-th graph node), used to scale the summed features to prevent the feature values ​​of nodes with too many neighbors from being too large; u, v are graph node indices;

[0138] S20355: Decode and map the updated node features to generate corresponding semantic labels (such as "normal handling", "abnormal shaking", "standby"), and combine the topological connection relationship between the graph nodes and graph edges to output structured standard robot data with semantic labels and topological structure.

[0139] In one optional implementation, a confidence score for the structured standard robot data is calculated. Based on the local training sample set whose confidence scores are less than a threshold, local model fine-tuning is performed. Local model parameter updates and auxiliary update data are collected and encrypted before being uploaded to the central server, including:

[0140] S2041: Input structured standard robot data into a quality assessment classifier built on a multilayer perceptron, output the corresponding high-dimensional feature distribution anomaly probability through the Softmax activation function, and generate the confidence score of the current data;

[0141] S2042: If the confidence score is greater than or equal to the score threshold, the structured standard robot data is deemed qualified, the structured standard robot data is stored in the cloud, and the robot data intelligent cleaning process ends.

[0142] S2043: If the confidence score is less than the score threshold, it is determined that the structured standard robot data has a sudden change in environmental noise distribution. This is most likely due to a sudden change in the distribution of environmental noise in the industrial site (such as a sudden change in light intensity or sudden electromagnetic interference), which leads to a decrease in the model's generalization ability. A queue feedback instruction is generated to adjust the data retention priority of the preprocessed robot data corresponding to the structured standard robot data and add it to the new sample set before proceeding to the next step.

[0143] The new sample set is extracted from a multidimensional data circular buffer queue;

[0144] S2044: Until the local training cycle is reached (e.g., every 24 hours or when the new sample set accumulates to 500 samples), perform data augmentation on the new sample set to obtain a local training sample set that includes positive and negative sample pairs.

[0145] In this embodiment, each sample is subjected to operations such as random cropping, color jittering, and Gaussian blurring to construct positive sample pairs; other samples are randomly selected as negative sample pairs to obtain a local training sample set.

[0146] S2045: Based on the total loss function, using the backpropagation algorithm, the local training sample set is sequentially input into the local denoising and desensitization feature extraction model and the structured reconstruction model for local fine-tuning. The local model parameter update amount, including the parameter update amounts of the denoising and desensitization feature extraction model and the structured reconstruction model, is calculated using the following formula:

[0147]

[0148] In the formula, This represents the total loss value. The contrastive learning loss value is obtained by using a contrastive learning algorithm to calculate the similarity between samples in the local training sample set in the feature space; This is a regularization hyperparameter used to balance the weights between the model's ability to adapt to new environments and its ability to retain old knowledge. For local model parameter indexing; The diagonal element of the Fisher information matrix represents the first element. The "importance" of each historical local model parameter to the task under normal operating conditions in the past; For the first One historical local model parameter; After local fine-tuning, the first Update the local model parameters; The topology preservation penalty coefficient is used to control the strength of topology distribution constraints; represents the total number of graph nodes in the industrial scene data ontology graph; u, v are the graph node indices; Before performing the desensitization operation for the feature space dynamic suppression module, the topological distribution similarity between the u-th and v-th graph nodes is calculated based on the mapping of the final high-dimensional feature tensor to the graph nodes. This is the topological distribution similarity between the u-th and v-th graph nodes calculated after the desensitization operation (Hadamard product mask) is performed in the feature space dynamic suppression module, based on the mapping of the desensitized high-dimensional feature tensor to the graph nodes;

[0149] In this embodiment, the parameters of the noise reduction and desensitization feature extraction model include:

[0150] Parameters of the noise reduction autoencoder neural network:

[0151] Convolutional layer parameters: weight matrices and bias terms of the one-dimensional / two-dimensional convolutional kernels of each layer, used to extract local features of multimodal data;

[0152] Transformer encoder layer parameters: including the weight matrix in the multi-head self-attention mechanism, as well as the layer normalization parameters and the weights and biases of the feedforward neural network, used to capture global feature dependencies;

[0153] Sensitive feature mask layer parameters:

[0154] Multi-head cross-attention parameters: The weight matrix and bias of cross-attention calculation using the final high-dimensional features as the query and the preset sensitive template as the key and value, are used to generate the initial sensitivity map and dynamically adjust the mask strength.

[0155] Feature space dynamic suppression module parameters:

[0156] If the module contains a differentiable mask generation network, it contains linear mapping weights in the mask generation process; if only Hadamard products are performed, the module has no independently learnable parameters (it depends on the output of the preceding mask layer).

[0157] Privacy Reconstruction Attack Network Parameters:

[0158] Deconvolutional layer parameters: The weights and biases of the deconvolution kernel used to attempt to reconstruct sensitive region features from desensitized features;

[0159] Decoder parameters:

[0160] Deconvolution and fully connected layer parameters: deconvolution weights, biases, and the final fully connected layer weight matrix used to decode and reduce the desensitized high-dimensional feature tensor into an intermediate feature tensor;

[0161] The parameters of the structured reconstruction model include:

[0162] Spatiotemporal joint attention network parameters:

[0163] Learnable spatial transformation matrix parameters: six-DOF affine transformation parameters (3 translation components, 3 rotation components), used to eliminate multi-sensor parallax;

[0164] Localization network parameters: weights and biases of convolutional and fully connected layers used to predict spatial transformation parameters;

[0165] Cross-modal time-weighted stitching module parameters:

[0166] Temporal attention weight matrix: A matrix used to calculate the temporal correlation weights within the sliding window and its bias;

[0167] Modality fusion weights: linear layer weights and biases used to weight and fuse the various modalities of the unified features;

[0168] Graph neural network parameters:

[0169] Learnable weight matrices based on relation types: graph convolution weight matrices corresponding to different physical operation state relations (such as cooperative transport, obstacle avoidance, etc.), used to perform linear transformation on the features of neighbor nodes in the message passing mechanism;

[0170] Graph convolutional layer parameters: including linear combination weights before nonlinear activation, bias terms, and normalization parameters;

[0171] S2046: Collect auxiliary update data for this round of local fine-tuning, mark the local model parameter update amount and auxiliary update data with upload timestamps, and encrypt and upload them to the central server.

[0172] In one optional implementation, the auxiliary update data includes the sample size of the local training sample set used in this round of local fine-tuning, the average confidence score, the accuracy on the validation set, and the gradient second moment oscillation data calculated based on the gradient vector of this round of local fine-tuning, as shown in the formula:

[0173]

[0174] In the formula, The data represents the second-order moment oscillation of the gradient. It is a mathematical expectation (mean) calculation function; This refers to the gradient vectors of the noise reduction and desensitization feature extraction model and the structured reconstruction model during this round of local fine-tuning.

[0175] In one optional implementation, at the central server, based on an adaptive weighted aggregation mechanism, the update amounts of local model parameters uploaded by several edge computing nodes are aggregated according to auxiliary update data, and the generated updated global model parameters are distributed to all edge computing nodes, including:

[0176] S2051: On the central server, based on a preset sliding time window, collect all local model parameter update amounts and auxiliary update data with upload timestamps within the same preset sliding time window, and record the corresponding number of uploaded data.

[0177] S2052: If the amount of uploaded data exceeds the threshold, proceed to the next step; otherwise, continue waiting for the edge computing node to upload data.

[0178] S2053: Based on an adaptive weighted aggregation mechanism, the weight influence factor of each edge computing node is calculated according to the auxiliary update data. The weight influence factor includes a sample size factor, a data quality and reliability factor, and a parameter stability penalty factor. The formula is as follows:

[0179]

[0180] In the formula, Let be the sample size factor, data quality and reliability factor, and parameter stability penalty factor for the k-th edge computing node; k,j is the index of the edge computing node. The sample size, average confidence score, accuracy, and gradient second moment oscillation data of the k-th edge node are calculated. The sample size, average confidence score, accuracy, and gradient second moment oscillation data of the j-th edge node are calculated. The total number of nodes at the edge is calculated; It is a very small number;

[0181] S2054: Multiply the sample size factor, data quality and reliability factor, and parameter stability penalty factor, and normalize the result to obtain the adaptive aggregation weights for adaptive weighted aggregation. The formula is as follows:

[0182]

[0183] In the formula, For the k-th edge computing node, there are the sample size factor, data quality and reliability factor, and parameter stability penalty factor. For the j-th edge computing node, there are the sample size factor, data quality and reliability factor, and parameter stability penalty factor. The adaptive aggregation weight is calculated for the k-th edge node;

[0184] S2055: Based on the adaptive aggregation weight, aggregate the local model parameter updates uploaded by several edge computing nodes to generate updated global model parameters. The formula is as follows:

[0185]

[0186] In the formula, Update the global model parameters obtained from this round of aggregation; Update the global model parameters obtained from the previous round of aggregation; The adaptive aggregation weight is calculated for the k-th edge node; This represents the amount of local model parameter updates for the k-th edge computing node;

[0187] S2056: The generated updated global model parameters are sent to all edge computing nodes. After receiving them, the edge computing nodes overwrite the old local model parameters and enter the next round of intelligent data cleaning and periodic fine-tuning cycle, thereby realizing the continuous evolution and adaptive optimization of the cleaning model of the entire robot cluster.

[0188] This invention also provides a robot data intelligent cleaning system 300 based on neural networks, see reference. Figure 3 The system may include the following units:

[0189] The robot data acquisition unit 301 is used to acquire robot data generated during robot operation at the edge computing node, and manage the robot data based on an adaptive circular buffer queue management mechanism.

[0190] The noise reduction and desensitization feature extraction unit 302 is used to input robot data into the noise reduction and desensitization feature extraction model based on neural network to generate an intermediate feature tensor that completes noise reduction and desensitization at the feature level.

[0191] The structured reconstruction unit 303 is used to input intermediate feature tensors into a structured reconstruction model built on a neural network to generate structured standard robot data with semantic labels and topological structures.

[0192] The local model fine-tuning unit 304 is used to calculate the confidence score of structured standard robot data. Based on the local training sample set whose confidence scores are less than the score threshold, it performs local model fine-tuning, collects local model parameter update amounts and auxiliary update data, and encrypts and uploads them to the central server.

[0193] The adaptive weighted aggregation unit 305 is used on the central server to aggregate the local model parameter update amounts uploaded by several edge computing nodes based on the adaptive weighted aggregation mechanism and auxiliary update data, and then distribute the generated updated global model parameters to all edge computing nodes.

[0194] Based on the same inventive concept, another embodiment of the present invention provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus.

[0195] Memory, used to store computer programs;

[0196] The processor, when executing the program stored in the memory, implements the neural network-based intelligent cleaning method for robot data of the present invention.

[0197] The communication bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used in the diagram, but this does not indicate that there is only one bus or one type of bus. The communication interface is used for communication between the aforementioned terminal and other devices. The memory can include Random Access Memory (RAM) or Non-Volatile Memory (NVM), such as at least one disk storage device. Optionally, the memory can also be at least one storage device located remotely from the aforementioned processor.

[0198] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be 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, or discrete hardware components.

[0199] Furthermore, to achieve the above objectives, embodiments of the present invention also propose a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the neural network-based intelligent cleaning method for robot data according to embodiments of the present invention.

[0200] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, embodiments of the present invention can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of the present invention can take the form of computer program products implemented on one or more computer-usable hardware devices (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0201] The embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (apparatus), and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0202] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0203] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0204] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. "And / or" indicates that either one or both can be chosen. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes the element.

[0205] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for intelligent cleaning of robot data based on neural networks, characterized in that, The method includes: At the edge computing node, robot data generated during robot operation is collected, and the robot data is managed based on an adaptive circular buffer queue management mechanism; The robot data is input into a noise reduction and desensitization feature extraction model based on a neural network to generate an intermediate feature tensor that performs both noise reduction and desensitization at the feature level. The intermediate feature tensor is input into a structured reconstruction model built on a neural network to generate structured standard robot data with semantic labels and topological structure. Calculate the confidence score of the structured standard robot data, perform local model fine-tuning based on the local training sample set whose confidence scores are less than the score threshold, collect the local model parameter update amount and auxiliary update data, and encrypt and upload them to the central server; At the central server, based on the adaptive weighted aggregation mechanism, the update amounts of local model parameters uploaded by several edge computing nodes are aggregated according to the auxiliary update data, and the generated updated global model parameters are distributed to all edge computing nodes.

2. The intelligent robot data cleaning method based on neural networks according to claim 1, characterized in that, At the edge computing node, robot data generated during robot operation is collected and managed based on an adaptive circular buffer queue management mechanism, including: At the edge computing node, raw robot data generated during robot operation is collected and preprocessed to obtain preprocessed robot data. Based on the PTP protocol and a local temperature-controlled crystal oscillator, the time deviation of the preprocessed robot data is calculated, and nanosecond-level timestamps are marked on the preprocessed robot data according to the time deviation. Based on an adaptive circular buffer queue management mechanism, a multidimensional data circular buffer queue is maintained, and the arrival rate of raw robot data is collected. If the arrival rate is greater than the processing rate of the edge computing node, the preprocessed robot data with timestamps will be stored in a multidimensional data circular buffer queue, and the corresponding timeliness-information entropy score will be calculated. If a queue feedback instruction is received, the data retention priority of the corresponding preprocessed robot data with timestamps will be adjusted to the highest level. If the number of data in the multidimensional data circular buffer queue reaches the preset capacity, the preprocessed robot data with the lowest timeliness-information entropy score or the lowest data retention priority with timestamps will be discarded, and the latest preprocessed robot data with timestamps will be stored to realize the management of robot data.

3. The intelligent robot data cleaning method based on neural networks according to claim 2, characterized in that, The noise reduction and desensitization feature extraction model includes a noise reduction autoencoder neural network, a feature space dynamic suppression module, a privacy reconstruction attack network, and a decoder. The noise reduction autoencoder neural network includes several feature extraction layers, and a sensitive feature mask layer is set at the output of the feature extraction layer of the last layer of the noise reduction autoencoder neural network. The structured reconstruction model includes a spatiotemporal joint attention network, a cross-modal temporal weighted stitching module, and a graph neural network; The graph neural network is connected to a pre-built industrial scene data ontology graph.

4. The intelligent robot data cleaning method based on neural networks according to claim 3, characterized in that, Robot data is input into a denoising and desensitization feature extraction model built on a neural network, generating an intermediate feature tensor that simultaneously performs denoising and desensitization at the feature level, including: The latest preprocessed robot data with timestamps, received by the edge computing node or stored in the multidimensional data circular buffer queue, is input into the noise reduction and desensitization feature extraction model built on a neural network. Several feature extraction layers are used to extract features from the preprocessed robot data with timestamps in sequence to generate the final high-dimensional feature tensor. During the inference phase, the trained sensitive feature mask layer is used to calculate the attention score of each spatial region in the final high-dimensional feature tensor based on the multi-head cross attention mechanism, and a dynamic adaptive mask map is generated through the Sigmoid function. The feature space dynamic suppression module is used to perform Hadamard product operation on the final high-dimensional feature tensor and the dynamic adaptive mask map to generate a desensitized high-dimensional feature tensor for dynamically suppressed sensitive privacy regions. The decoder is used to decode the desensitized high-dimensional feature tensor to generate an intermediate feature tensor that simultaneously performs denoising and desensitization at the feature level.

5. The intelligent robot data cleaning method based on neural networks according to claim 4, characterized in that, The intermediate feature tensors are input into a structured reconstruction model built on a neural network to generate structured standard robot data with semantic labels and topological structure, including: The intermediate feature tensor is input into a structured reconstruction model built on a neural network; Using a spatiotemporal joint attention network, based on a learnable spatial transformation matrix, a six-degree-of-freedom coordinate system is established for the sensor disparities of different modalities in the intermediate feature tensor to obtain a unified feature tensor. Using the cross-modal temporal weighted concatenation module, based on the preset cross-modal temporal sliding window and the learnable attention weight matrix, the temporal correlation weight of each modality in the unified feature tensor within the same cross-modal temporal sliding window is calculated. Based on the temporal correlation weights, the feature components of all modes of the unified feature tensor are concatenated at the feature level to obtain the weighted concatenated feature tensor. Using graph neural networks, weighted splicing feature tensors are mapped to a pre-built industrial scene data ontology graph. Based on the time-varying decay edge weights at the current moment, dynamic topology reconstruction and node feature updates are performed to generate structured standard robot data with semantic labels and topological structures.

6. The intelligent robot data cleaning method based on neural networks according to claim 5, characterized in that, Using graph neural networks, weighted concatenated feature tensors are mapped to a pre-constructed industrial scene data ontology graph. Based on the time-varying decaying edge weights at the current moment, dynamic topology reconstruction and node feature updates are performed to generate structured standard robot data with semantic labels and topological structures, including: The graph structure is constructed by using the robots corresponding to the weighted splicing feature tensors as the graph nodes of the pre-constructed industrial scene data ontology graph, and the physical operation state relationships between robots as the graph edges of the industrial scene data ontology graph. Obtain the state update time difference between adjacent graph nodes at the current time, and calculate the time-varying decay edge weight of the graph edge at the current time by combining it with a preset decay coefficient. If the weight of the time-varying decay edge is less than the preset topology disconnection threshold, the dynamic disconnection of the corresponding graph edge is triggered, and the message transmission of the graph edge is stopped. If the feature cosine similarity between any two non-adjacent graph nodes calculated based on the unified feature tensor is greater than the preset topology association threshold, dynamic edge construction is triggered to build a graph structure between the two non-adjacent graph nodes, and the reconstructed industrial scene data ontology graph is obtained. Based on the time-varying decay edge weights and the message passing mechanism of the graph neural network, the node features of adjacent graph nodes are weighted and aggregated to update the feature representation of the current node, thus obtaining the updated node features. The updated node features are decoded and mapped to generate corresponding semantic labels. Combined with the topological connection relationship between the graph nodes and graph edges, structured standard robot data with semantic labels and topological structure is output.

7. The intelligent robot data cleaning method based on neural networks according to claim 6, characterized in that, Calculate the confidence score of the structured standard robot data. Based on the local training sample set with confidence scores below the score threshold, perform local model fine-tuning, collect local model parameter update amounts and auxiliary update data, and encrypt and upload them to the central server, including: Structured standard robot data is input into a quality assessment classifier built on a multilayer perceptron. The corresponding high-dimensional feature distribution anomaly probability is output through the Softmax activation function to generate a confidence score for the current data. If the confidence score is greater than or equal to the score threshold, the structured standard robot data is deemed qualified, the structured standard robot data is stored in the cloud, and the robot data intelligent cleaning process ends. If the confidence score is less than the score threshold, it is determined that there is a sudden change in the distribution of environmental noise in the structured standard robot data. A queue feedback instruction is generated to adjust the data retention priority of the preprocessed robot data corresponding to the structured standard robot data and add it to the new sample set, and then proceed to the next step. Until the local training cycle is reached, data augmentation operations are performed on the new sample set to obtain a local training sample set including positive and negative sample pairs. Based on the total loss function, the backpropagation algorithm is used to input the local training sample set into the local noise reduction and desensitization feature extraction model and the structured reconstruction model for local fine-tuning, and calculate the local model parameter update amount, including the parameter update amount of the noise reduction and desensitization feature extraction model and the parameter update amount of the structured reconstruction model. Collect auxiliary update data for this round of local fine-tuning, mark the local model parameter update amount and auxiliary update data with upload timestamps, and encrypt and upload them to the central server.

8. The intelligent robot data cleaning method based on neural networks according to claim 7, characterized in that, The auxiliary update data includes the sample size of the local training sample set used in this round of local fine-tuning, the average confidence score, the accuracy on the validation set, and the gradient second moment oscillation data calculated based on the gradient vector of this round of local fine-tuning.

9. The intelligent robot data cleaning method based on neural networks according to claim 8, characterized in that, At the central server, based on an adaptive weighted aggregation mechanism, the update amounts of local model parameters uploaded by several edge computing nodes are aggregated according to auxiliary update data, and the generated updated global model parameters are distributed to all edge computing nodes, including: On the central server, based on a preset sliding time window, the update volume of all local model parameters and auxiliary update data with upload timestamps in the same preset sliding time window are collected, and the corresponding number of uploaded data is recorded. If the amount of uploaded data exceeds the threshold, proceed to the next step; otherwise, continue waiting for the edge computing node to upload data. Based on the adaptive weighted aggregation mechanism, the weight influence factor of each edge computing node is calculated according to the auxiliary update data. The weight influence factor includes the sample size factor, the data quality and reliability factor, and the parameter stability penalty factor. The adaptive aggregation weights of the adaptive weighted aggregation are obtained by multiplying the sample size factor, data quality and reliability factor, and parameter stability penalty factor and normalizing them. Based on the adaptive aggregation weight, the local model parameter update amounts uploaded by several edge computing nodes are aggregated to generate updated global model parameters; The generated updated global model parameters are distributed to all edge computing nodes.

10. A neural network-based intelligent robot data cleaning system, used to implement the intelligent robot data cleaning method as described in any one of claims 1-9, characterized in that, The system includes: The robot data acquisition unit is used to collect robot data generated during robot operation at the edge computing node, and manage the robot data based on an adaptive circular buffer queue management mechanism. The noise reduction and desensitization feature extraction unit is used to input robot data into the noise reduction and desensitization feature extraction model based on neural network to generate an intermediate feature tensor that completes noise reduction and desensitization at the feature level. The structured reconstruction unit is used to input intermediate feature tensors into a structured reconstruction model built on a neural network to generate structured standard robot data with semantic labels and topological structures. The local model fine-tuning unit is used to calculate the confidence score of structured standard robot data. Based on the local training sample set with confidence scores below the score threshold, it performs local model fine-tuning, collects local model parameter updates and auxiliary update data, and encrypts and uploads them to the central server. The adaptive weighted aggregation unit is used on the central server to aggregate the local model parameter update amounts uploaded by several edge computing nodes based on the adaptive weighted aggregation mechanism, and then distribute the generated updated global model parameters to all edge computing nodes.

Images