Machine room monitoring method and system based on multi-source data fusion intelligent inspection robot

Abstract

The application discloses a computer room monitoring method and system based on a multi-source data fusion intelligent inspection robot, belongs to the technical field of automatic computer room monitoring, and comprises application of adversarial transfer learning on a four-dimensional fault semantic feature field, generation of a cross-modal causal graph atlas representing a fault evolution path through a graph neural network, and the like, wherein the application aligns the feature distribution of a standard computer room and a current computer room through a joint loss function, forcibly aligns the feature distribution of a source domain and a target domain through a gradient reversal layer, generates a cross-modal causal graph atlas in combination with an attention mechanism and a causal strength weight, the graph neural network further models the physical connection, functional dependence and time sequence correlation between nodes, encodes the causal relationship as an edge weight, filters noise correlation through a causal mask, and improves the stability of the causal graph atlas, and the cross-modal causal graph atlas can accurately capture a fault propagation path.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of automatic monitoring of machine rooms, and particularly relates to a machine room monitoring method and system based on a multi-source data fusion intelligent inspection robot. BACKGROUND

[0002] With the rapid development of information technology, as the core hub of data processing and communication, machine rooms have high equipment density and complex operating environments, and traditional manual inspection has problems such as low efficiency, high omission rate and delayed response. In recent years, intelligent inspection robots have been widely used in machine room monitoring, and through the use of visible light, infrared, ultrasonic and other sensors to collect environmental and equipment state data, automatic inspection is realized.

[0003] However, the machine room monitoring of the existing inspection robot still has certain defects. The existing multi-source data fusion usually relies on simple splicing or weighted average, lacks modeling of the spatio-temporal correlation and semantic consistency between heterogeneous data, and mainly uses pure neural network models in fault diagnosis. Although the pure neural network model has high precision, it lacks explainability and is difficult to trace the fault causal chain. The rule reasoning system has low computational efficiency and cannot meet the real-time requirements. Therefore, the machine room monitoring method and system based on the multi-source data fusion intelligent inspection robot are proposed. SUMMARY

[0004] The purpose of the present application is to provide a machine room monitoring method and system based on a multi-source data fusion intelligent inspection robot to solve the problems raised in the background.

[0005] To achieve the above purpose, the present application provides the following technical scheme: a machine room monitoring method based on a multi-source data fusion intelligent inspection robot, comprising the following steps:

[0006] S1, aligning the spatio-temporal reference of multi-source heterogeneous sensor data by synchronizing the quantum time stamp chip with the laser SLAM reflector plate positioning;

[0007] S2, based on the spatio-temporal aligned data, fusing visible light, depth, thermal and electromagnetic properties in a three-dimensional spatial grid to construct a four-dimensional fault semantic feature field;

[0008] S3, applying adversarial transfer learning on the four-dimensional fault semantic feature field to generate a cross-modal causal graph representing the fault evolution path through a graph neural network;

[0009] S4, based on the cross-modal causal graph, synchronously executing rule reasoning and graph neural network prediction through a neural-symbol hybrid chip to output a diagnosis confidence;

[0010] S5, when the diagnosis confidence result is lower than a preset threshold, performing causal subgraph distillation to inject a lightweight edge knowledge base through incremental learning;

[0011] S6, in combination with the updated lightweight edge knowledge base, simulate derived fault scenarios on the digital twin platform to generate an optimal decision tree, and dynamically adjust the robot inspection path;

[0012] S7, based on the national secret algorithm block chain, seal the whole chain evidence, generate the space-time credible record, and feed back the actual disposal effect of the robot to the cross-modal causal graph correction.

[0013] Preferably, S1, through quantum timestamp synchronization chip and laser SLAM reflector positioning, the time and space reference alignment of multi-source heterogeneous sensor data is carried out; quantum timestamp synchronization chip is installed on the key sensor node of the robot, the communication interface of the chip is configured, and the timestamp broadcast and synchronization of cross-modal sensors are realized;

[0014] Laser SLAM reflector is arranged in the environment, ensuring that the key path nodes of the robot activity area are covered, the relative position relationship between laser radar and reflector is calibrated, and the mapping table of reflector coordinate system and global coordinate system is established;

[0015] By automatically attaching quantum timestamp to multi-source heterogeneous sensor data during collection, the time consistency of cross-modal data is ensured, and the time offset between sensors is calibrated by using the atomic clock comparison mechanism of quantum timestamp synchronization chip;

[0016] It should be understood that through the distributed clock synchronization protocol, the sensor clock drift is monitored in real time during the movement of the robot, and dynamic correction is carried out through the feedback loop of the quantum chip, so as to maintain long-term synchronization stability, and the laser radar scans the reflection point cloud of the reflector in the environment;

[0017] The position of the reflector is identified by geometric features, the current pose of the laser radar relative to the reflector is calculated by combining the pre-defined reflector coordinate mapping table, the point cloud data is registered by laser SLAM algorithm, and the real-time pose of the robot in the global coordinate system is optimized;

[0018] Specifically, through the fixed position constraint of the reflector, the cumulative error in the SLAM process is corrected, the time and space reference alignment is carried out, the time dimension alignment is carried out: all sensor data is aligned on the time axis according to the quantum timestamp, so as to ensure that the multi-modal data at the same time can be jointly analyzed, and the space dimension alignment is carried out: the coordinate system of each sensor is converted to a unified global coordinate system, and the spatial transformation matrix of the non-laser sensor relative to the laser radar is determined by external parameter calibration.

[0019] Preferably, S2, based on the time and space aligned data, the visible light, depth, thermal and electromagnetic properties are fused in the three-dimensional space grid to construct a four-dimensional fault semantic feature field; based on the three-dimensional environment model constructed by laser SLAM, the physical space is divided into regular three-dimensional voxel grid, and each voxel corresponds to a space coordinate (x, y, z).

[0020] Specifically, the grid granularity is adjusted according to the sensor resolution, and the output spatio-temporal alignment data is mapped into the corresponding voxel according to the timestamp and spatial coordinates, and the spatio-temporal alignment data includes visible light, depth, thermal force, and electromagnetic properties, and feature extraction is performed on the spatio-temporal alignment data:

[0021] Visible light: color histogram, texture feature and object contour feature are extracted;

[0022] Depth: surface curvature, normal vector and obstacle distance distribution are calculated;

[0023] Thermal force: temperature gradient, hot spot area and abnormal temperature fluctuation pattern are extracted;

[0024] Electromagnetic: abnormal signal frequency, intensity peak and spatial distribution rule are identified;

[0025] The features of each modality are normalized to eliminate dimensional differences, and the dimensions of the heterogeneous data are unified, the multi-modal features are fused in early, middle and late stages, and the fused feature vectors are assigned to the corresponding voxels to form a three-dimensional spatial feature grid;

[0026] The voxel features at multiple time points are sequenced by timestamp to construct a spatio-temporal tensor to capture dynamic change trends, and the feature vector of each three-dimensional grid unit contains four-dimensional attributes, the fused feature vector is mapped to each grid unit to form a four-dimensional feature field, and the features are semantically encoded by a semantic model, and in the robot inspection process, new sensor data is collected in real time and the three-dimensional grid features are updated.

[0027] Preferably, the S3, the step of adversarial transfer learning comprises: setting a source domain and a target domain, the source domain is a fault data set of a standard machine room, and the target domain is real-time four-dimensional fault semantic feature field data x of the current machine room, and the feature distribution of the source domain and the target domain is aligned through a joint loss function of adversarial transfer learning, and the formula is:

[0028] ,

[0029] In the formula, represents the total value of the joint loss function, represents the comprehensive performance of the model in the source domain task, the target domain task and the domain adversarial task, represents a weight coefficient for balancing the contributions of the source domain task loss, the target domain task loss and the domain adversarial loss, represents the source domain task loss, and is represented as:

[0030] ,

[0031] In the formula, represents the number of source domain samples, denotes a loss function, denotes the true label of the i-th sample in the source domain, denotes the model prediction output;

[0032] denotes the target domain task loss, denoted as:

[0033] ,

[0034] In the formula, denotes the number of target domain samples, denotes a loss function, denotes the true label of the j-th sample in the target domain, denotes the model prediction output;

[0035] denotes the domain adversarial loss, F denotes the feature extractor, and the source domain and the target domain feature distribution are aligned by the gradient inversion layer, and is realized as:

[0036] ,

[0037] In the formula, D denotes the domain classifier, which determines whether the feature comes from the source domain or the target domain, F denotes the feature extractor, which maps the input data x to a feature vector, denotes the source domain input data, denotes the target domain input data.

[0038] Preferably, the S3, the step of constructing a cross-modal graph structure comprises: setting nodes and edges, the nodes include each sensor node as a node in the graph, the node features are the fusion of multi-modal data in the four-dimensional fault semantic feature field, the edges include physical connection relationship, functional dependence relationship and dynamic time sequence association, the feature vector of the node is generated by aggregating the neighbor node information, and the formula is realized as:

[0039] ,

[0040] In the formula, denotes the feature vector of node v in the l+1 layer, denotes an activation function, denotes the summation of all neighbor nodes u of node v, denotes the neighbor node set of node v, denotes the influence degree of node u on node v, which is the neighbor weight based on the attention mechanism, denotes , a denotes an attention parameter vector, denotes a splicing operation, W denotes a learnable weight matrix used for linear conversion of neighbor node features , the feature vector of node u in the l-th layer the feature vector of the l-th layer, represents the causal strength between nodes u and v, the causal influence of u on v;

[0041] by aligning the source domain and target domain feature distributions, and the distribution of approaches, is expressed as , represents the probability distribution of node v under the condition of node u, and the attention weight is combined with the causal strength , so that the node feature update considers both neighbor importance and causal relationship fusion;

[0042] Encode the causal relationship as the edge weight in the graph to form a cross-modal causal graph G, which is expressed as .

[0043] Preferably, the S3, cross-modal causal graph optimization step includes: combining the real-time four-dimensional fault semantic feature field, incrementally updating the edge weight in the causal graph, filtering noise association through a causal mask, and realizing the formula:

[0044] ,

[0045] In the formula, represents the value of the edge weight between nodes u and v after t+1 iterations, represents the value of the edge weight between nodes u and v in the t-th iteration, represents the learning rate, represents the causal consistency loss the gradient of the edge weight , is expressed as , represents the edge set of the graph, represents the causal influence of node u on v, represents a noise suppression coefficient for controlling the weight of the noise term, represents a Gaussian noise term for preventing fitting, is expressed as , represents a Gaussian distribution with a mean of 0 and a variance of ;

[0046] Compare the matching degree of the artificially labeled causal chain with the generation result of the cross-modal causal graph to test the stability of the cross-modal causal graph in different computer room scenarios.

[0047] Preferably, the S4, the neural-symbol hybrid chip step comprises: integrating a symbolic reasoning unit and a neural network unit, calculating through parallel of the neural-symbol hybrid chip, cross-modal alignment of real-time monitoring data and historical fault records to extract unified features;

[0048] The aligned feature vector is injected into the node V of the cross-modal causal graph atlas as the initial state of reasoning and prediction, and the node state of the causal graph atlas G is dynamically updated according to the real-time data stream.

[0049] Preferably, the S4, the quantitative confidence step comprises: based on the edge weight of the cross-modal causal graph atlas G, applying a predefined causal rule, if multiple causal rules produce contradictory conclusions for the same node, arbitrating through confidence weighting, assigning a fault probability distribution to each node, jointly quantifying the causal confidence, and realizing the formula:

[0050] ,

[0051] In the formula, represents the diagnostic confidence, represents the temperature coefficient, which controls the smoothness of the confidence distribution, and is obtained by jointly weighting the causal weight and the feature similarity, capturing the dynamic updated causal consistency, mapping the joint quantization result to the interval [0, 1], and directly outputting the explainable confidence.

[0052] Preferably, the S5, low-confidence diagnosis result screening and causal subgraph extraction: according to historical data statistics, set the diagnostic confidence threshold Tn, screen out the diagnostic results of all nodes v with confidence The diagnostic confidence of the node v is high, and the causal subgraph of the node v is extracted from the cross-modal causal graph atlas.

[0053] Preferably, the S5, the causal subgraph distillation step comprises: a teacher model and a student model, the teacher model: based on the neural-symbol hybrid chip, predicting the causal subgraph to output high-precision causal chain, the student model: constructing a lightweight model, learning the associated causal logic and classification boundary of the teacher model in the causal subgraph;

[0054] The prediction probability distribution of the node in the causal subgraph by the teacher model is used as a supervision signal, and the student model is trained through a KL divergence loss function, explicit causal rules are extracted from the symbolic reasoning results of the teacher model, and logical forms are injected into the knowledge base of the student model, and a lightweight causal subgraph knowledge representation;

[0055] It should be understood that the edge knowledge base is stored in the form of causal graph fragments and rule index table, the newly distilled causal subgraph knowledge is detected for conflict and merged with the existing knowledge in the edge knowledge base, the existing student model is fine-tuned on the edge device using the new knowledge, and the newly added entries of the edge knowledge base are time-stamped and versioned.

[0056] Preferably, S6, based on the extracted causal rules in the lightweight edge knowledge base, the compressed model parameters and the low confidence node supplementary data, maps the causal rules in the knowledge base to the virtual device model of the digital twin platform according to the device state, constructs a dynamic causal relationship network; the compressed model parameters are deployed to the reasoning engine of the digital twin platform for real-time prediction of potential failure scenarios;

[0057] Specifically, based on the causal rules in the knowledge base, the possible derived failure scenarios are simulated, the multi-dimensional failure hypotheses are generated combining real-time sensor data and historical failure patterns, the risk score of each failure scenario is calculated through the compressed model, the scenario is prioritized using the causal chain, and the failure type that needs to be processed first is determined, taking the failure scenario with the highest risk score as the root node, and gradually splitting the child nodes downward, each node containing decision conditions, action suggestions and expected returns;

[0058] The splitting logic of the decision tree is constrained by the causal rules, the prediction results of the lightweight model are introduced as the confidence evaluation of the leaf nodes, the low-efficiency branches are eliminated, the key actions are extracted from the decision tree as the task list of the robot inspection, and the initial path is generated combining the real-time state of the digital twin platform;

[0059] The task order is dynamically adjusted according to the urgency of the actions in the decision tree, the feasibility of the path is verified through virtual simulation of the digital twin, collision is avoided and the shortest path is selected, if a new failure signal is detected during the execution of the robot, local update of the decision tree is triggered and the path is re-planned, the actual execution result of the robot is compared with the predicted value of the decision tree, and the effectiveness of the decision tree is evaluated;

[0060] New causal relationships during execution are recorded on the digital twin platform, the newly discovered causal relationships and the optimized decision tree logic are injected into the edge knowledge base, and through the incremental learning mechanism of S5, the updated knowledge base is fed back to the neuro-symbolic hybrid chip of S4 to optimize the global causal graph.

[0061] Preferably, S7, based on the national secret algorithm block chain, the whole chain evidence is sealed, the space-time credible evidence is generated, and the actual disposal effect of the robot is fed back to the cross-modal causal graph correction; the fault diagnosis result, the quantitative confidence and the causal chain output by S4 are obtained, the optimal decision tree, the robot inspection path planning and the execution record generated by S6 are obtained, the sensor feedback of the actual execution action of the robot is obtained, and the above evidences are uniformly encoded into structured data blocks, including the following fields:

[0062] Timestamp: Record the time of event occurrence;

[0063] Geolocation: Associate the physical location of the device;

[0064] Evidence type: Differentiate diagnostic, decision or treatment evidence;

[0065] Causal chain identifier: Reference the node / edge identifier in the cross-modal causal graph G;

[0066] Hash value: Calculate the hash value of the data block by the national encryption algorithm to ensure data integrity;

[0067] Specifically, the structured data block is encrypted by the national encryption algorithm, the encrypted data block is signed by the national encryption SM2 algorithm, the authenticity of the data source is verified, the encrypted data block and the signature result are packaged as a blockchain transaction, and the national encryption algorithm is written into the alliance chain;

[0068] Get strong trusted timestamp through trusted timestamp agency, combine GPS module and Internet of Things sensor to verify the authenticity of the geographic location, and write the geographic coordinate hash value into the blockchain. From the sensor data after the robot performs the action, extract the actual treatment effect. If the treatment effect is consistent with the S4 diagnosis, update the weight of the related edge in the cross-modal causal graph G. If the treatment effect is inconsistent with the diagnosis, adjust the attribute label and feature vector of the node v. The corrected cross-modal causal graph is fed back to the edge knowledge base and digital twin platform.

[0069] Preferably, the machine room monitoring system based on multi-source data fusion intelligent patrol robot comprises:

[0070] Multi-source spatio-temporal reference alignment module: through quantum timestamp synchronization chip and laser SLAM reflector positioning, spatio-temporal reference alignment of multi-source heterogeneous sensor data is performed;

[0071] Fault semantic feature field construction module: based on spatio-temporal alignment data, visible light, depth, thermal and electromagnetic properties are fused in a three-dimensional space grid to construct a four-dimensional fault semantic feature field;

[0072] Cross-modal causal graph migration module: apply adversarial transfer learning on the four-dimensional fault semantic feature field to generate a cross-modal causal graph G representing the fault evolution path through a graph neural network;

[0073] Neural-symbol joint reasoning module: based on the cross-modal causal graph G, rule reasoning and graph neural network prediction are synchronously executed through a neural-symbol hybrid chip to output a diagnosis confidence;

[0074] Incremental causal distillation learning module: when the diagnosis confidence result is lower than the preset threshold, perform causal subgraph distillation to inject a lightweight edge knowledge base through incremental learning;

[0075] Virtual-real linkage digital twin module: combined with the updated lightweight edge knowledge base, the optimal decision tree is generated by simulating the derived fault scene in the digital twin platform, and the robot inspection path is dynamically adjusted;

[0076] Fault tracing on-chain evidence module: based on the national secret algorithm block chain, the whole chain evidence is sealed, the space-time credible evidence is generated, and the actual disposal effect of the robot is fed back to the cross-modal causal graph correction.

[0077] Compared with the prior art, the beneficial effects of the present application are:

[0078] 1、The present application aligns the feature distribution of the standard machine room and the current machine room by joint loss function, forces the source domain and target domain feature distribution to align by gradient reversal layer, combines attention mechanism and causal strength weight to generate cross-modal causal graph, graph neural network further models the physical connection, functional dependence and time sequence correlation between nodes, encodes the causal relationship as edge weight, filters noise correlation through causal mask, improves the stability of causal graph, and the cross-modal causal graph can accurately capture the fault propagation path;

[0079] 2、The present application realizes dynamic updating of the cross-modal causal graph by parallel execution of rule reasoning and graph neural network prediction through the neural-symbol hybrid chip, injects the real-time monitoring data and historical fault records into the causal graph nodes after alignment, applies the pre-defined causal rules through the symbolic reasoning unit, and simultaneously predicts the node state change through the neural network, when multiple causal rules produce contradictory conclusions for the same node, the confidence weighted arbitration is used to output the explainable confidence, the neural-symbol hybrid calculation improves the real-time performance, and the fault diagnosis result is both efficient and transparent;

[0080] 3、The present application compresses the high-precision teacher model of the neural-symbol hybrid chip into a lightweight student model through the causal subgraph distillation technology, extracts the causal subgraph with low confidence nodes as the core when the diagnosis confidence is lower than the threshold, trains the student model through the KL divergence loss function, retains the key causal logic, dynamically updates the knowledge base through the incremental learning mechanism, the student model can quickly learn and inject it into the knowledge base, avoids missing detection caused by outdated models, the lightweight causal subgraph is stored as a causal graph fragment and version number mark, supports multi-version knowledge conflict detection and merging, significantly reduces the storage and computing overhead of edge devices, and improves the real-time response ability of the model;

[0081] 4、The application constructs a four-dimensional fault semantic feature field containing spatial coordinates and semantic attributes through three-dimensional voxel grid division and multi-modal feature fusion strategy, maps visible light, depth, thermal and electromagnetic data into voxels according to timestamps, eliminates dimensional differences through normalization and early-middle-late fusion strategy, generates high-dimensional feature vectors, constructs a space-time tensor through time serialization, can dynamically monitor the fault evolution process, and the semantic coding of the four-dimensional feature field further converts physical features into interpretable fault labels, significantly improves the accuracy and interpretability of fault location. BRIEF DESCRIPTION OF DRAWINGS

[0082] Figure 1 The operation flow of the machine room monitoring method of the multi-source data fusion intelligent inspection robot Figure 1 ;

[0083] Figure 2 The operation flow of the machine room monitoring method of the multi-source data fusion intelligent inspection robot Figure 2 ;

[0084] Figure 3 The operation flow of the machine room monitoring method of the multi-source data fusion intelligent inspection robot Figure 3 ;

[0085] Figure 4 The structural schematic diagram of the machine room monitoring system of the multi-source data fusion intelligent inspection robot. DETAILED DESCRIPTION

[0086] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the application.

[0087] Embodiment:

[0088] Please refer to Figures 1-4 The application provides a technical solution comprising the following steps:

[0089] S1, time-space reference alignment of multi-source heterogeneous sensor data is performed by quantum timestamp synchronization chip and laser SLAM reflector positioning.

[0090] S2, based on the time-space aligned data, visible light, depth, thermal and electromagnetic properties are fused in a three-dimensional space grid to construct a four-dimensional fault semantic feature field.

[0091] S3, apply adversarial transfer learning on the four-dimensional fault semantic feature field, and generate a cross-modal causal graph representing the fault evolution path through a graph neural network;

[0092] S4, based on the cross-modal causal graph, perform rule reasoning and graph neural network prediction simultaneously through a neural-symbol hybrid chip, and output a diagnosis confidence;

[0093] S5, when the diagnosis confidence result is lower than the preset threshold, perform causal subgraph distillation, and inject lightweight edge knowledge base through incremental learning;

[0094] S6, in combination with the updated lightweight edge knowledge base, generate an optimal decision tree by simulating a derived fault scenario on a digital twin platform, and dynamically adjust the robot inspection path;

[0095] S7, based on the national cryptographic algorithm blockchain, seal the entire chain of evidence, generate a spatiotemporal trusted record, and feed back the actual disposal effect of the robot to the cross-modal causal graph for correction.

[0096] Preferably, in S1, the quantum timestamp synchronization chip and the laser SLAM reflector are positioned to align the spatiotemporal reference of multi-source heterogeneous sensor data; the quantum timestamp synchronization chip is installed on the key sensor node of the robot, the communication interface of the chip is configured, and the timestamp broadcast and synchronization of cross-modal sensors are realized;

[0097] In this embodiment, the laser SLAM reflector is arranged in the environment to ensure that the key path nodes in the robot activity area are covered, the relative position relationship between the laser radar and the reflector is calibrated, and the mapping table of the reflector coordinate system and the global coordinate system is established;

[0098] The quantum timestamp is automatically attached to the multi-source heterogeneous sensor data during collection, ensuring the temporal consistency of cross-modal data, and the atomic clock comparison mechanism of the quantum timestamp synchronization chip is used to calibrate the time offset between sensors;

[0099] It should be understood that through the distributed clock synchronization protocol, the sensor clock drift is monitored in real time during the movement of the robot, and dynamic correction is performed through the feedback loop of the quantum chip to maintain long-term synchronization stability, and the laser radar scans the reflection point cloud of the reflector in the environment;

[0100] In this embodiment, the position of the reflector is identified through geometric features, the pre-defined reflector coordinate mapping table is combined, the current pose of the laser radar relative to the reflector is calculated, the point cloud data is registered through the laser SLAM algorithm, and the real-time pose of the robot in the global coordinate system is optimized;

[0101] Specifically, by fixing the position of the reflector, the cumulative error in the SLAM process is corrected, the space-time reference is aligned, the time dimension is aligned: all sensor data are aligned on the time axis according to the quantum timestamp, ensuring that multi-modal data at the same time can be jointly analyzed, and the spatial dimension is aligned: the coordinate system of each sensor is converted to a unified global coordinate system, and the spatial transformation matrix of the non-laser sensor relative to the laser radar is determined by the external parameter calibration.

[0102] Preferably, S2, based on the space-time alignment data, fuses the visible light, depth, thermal and electromagnetic properties in the three-dimensional space grid to construct a four-dimensional fault semantic feature field; based on the three-dimensional environment model constructed by laser SLAM, the physical space is divided into a regular three-dimensional voxel grid, and each voxel corresponds to a spatial coordinate (x, y, z);

[0103] Specifically, the grid granularity is adjusted according to the sensor resolution, and the output space-time alignment data is mapped to the corresponding voxel according to the timestamp and spatial coordinate. The space-time alignment data includes visible light, depth, thermal, and electromagnetic properties. Feature extraction is performed on the space-time alignment data:

[0104] Visible light: extracting color histogram, texture feature and object contour feature;

[0105] Depth: calculating surface curvature, normal vector and obstacle distance distribution;

[0106] Thermal: extracting temperature gradient, hot spot area and abnormal temperature fluctuation pattern;

[0107] Electromagnetic: identifying abnormal signal frequency, intensity peak and spatial distribution rule;

[0108] Normalizing each modal feature to eliminate dimensional differences, unifying the dimensions of heterogeneous data, early, middle and late fusion of multi-modal features, and assigning the fused feature vector to the corresponding voxel to form a three-dimensional spatial feature grid;

[0109] In this embodiment, the voxel feature sequences of multiple time points are constructed by timestamp to build a space-time tensor to capture dynamic change trends. The feature vector of each three-dimensional grid unit contains four-dimensional properties. The fused feature vector is mapped to each grid unit to form a four-dimensional feature field. The features are semantically encoded by a semantic model. In the robot inspection process, new sensor data is collected in real time and the three-dimensional grid features are updated.

[0110] Preferably, the S3, the steps of the adversarial transfer learning include: setting a source domain and a target domain, the source domain is a standard machine room fault data set, and the target domain is the real-time four-dimensional fault semantic feature field data x of the current machine room. The feature distribution of the source domain and the target domain is aligned by the joint loss function of the adversarial transfer learning, and the formula is realized as:

[0111] ,

[0112] In the formula, represents the total value of the joint loss function, represents the comprehensive performance of the model in the source domain task, the target domain task and the domain adversarial task, represents the weight coefficient, used to balance the contribution of the source domain task loss, the target domain task loss and the domain adversarial loss, represents the source domain task loss, which is represented as:

[0113] ,

[0114] In the formula, represents the number of source domain samples, represents the loss function, represents the true label of the i-th sample in the source domain, represents the model prediction output;

[0115] represents the target domain task loss, which is represented as:

[0116] ,

[0117] In the formula, represents the number of target domain samples, represents the loss function, represents the true label of the j-th sample in the target domain, represents the model prediction output;

[0118] represents the domain adversarial loss, F represents the feature extractor, which is forced to align the source domain and target domain feature distribution through the gradient inversion layer, and is realized as:

[0119] ,

[0120] In the formula, D represents the domain classifier, which judges whether the feature comes from the source domain or the target domain, F represents the feature extractor, which maps the input data x to the feature vector, represents the source domain input data, represents the target domain input data.

[0121] Preferably, the S3, the step of constructing a cross-modal graph structure comprises: setting nodes and edges, the nodes include each sensor node as a node in the graph, the node features are the fusion of multi-modal data in the four-dimensional fault semantic feature field, the edges include physical connection relationship, functional dependence relationship and dynamic time sequence association, the feature vector of the node is generated by aggregating the neighbor node information, and the formula is realized as:

[0122] ,

[0123] In the formula, represents the feature vector of node v at the l+1 layer, represents the activation function, represents the summation of all neighbor nodes u of node v, represents the neighbor node set of node v, represents the influence degree of node u on node v, based on the neighbor weight of attention mechanism, represents , a represents the attention parameter vector, represents the concatenation operation, W represents the learnable weight matrix for linear conversion of neighbor node features , represents the feature vector of node u at the l layer layer, represents the causal strength between nodes u and v, and the causal influence of u on v;

[0124] In this embodiment, by aligning the feature distribution of the source domain and the target domain, the distributions of and are close, represents , represents the probability distribution of node v under the condition of node u, and the attention weight is combined with the causal strength , so that the node feature update considers both the importance of neighbors and the causal relationship;

[0125] The causal relationship is encoded as the edge weight in the graph to form a cross-modal causal graph G, which is represented as .

[0126] Preferably, the S3, the cross-modal causal graph optimization step includes: combining the real-time four-dimensional fault semantic feature field to incrementally update the edge weight in the causal graph, filtering the noise association through the causal mask, and realizing the formula:

[0127] ,

[0128] In this embodiment, represents the value of the edge weight between nodes u and v after t+1 iterations, represents the value of the edge weight between nodes u and v at the tth iteration, represents the learning rate, represents the causal consistency loss to the gradient of the edge weight , represents , a set of edges of a graph, a causal influence of a node u on v, a noise suppression coefficient for controlling the weight of the noise term, a Gaussian noise term for preventing overfitting, a Gaussian distribution with mean 0 and variance , a Gaussian distribution with mean 0 and variance

[0129] The matching degree of the artificially labeled causal chain and the cross-modal causal graph generation result is compared to test the stability of the cross-modal causal graph in different computer room scenarios.

[0130] Preferably, the S4, the neural-symbol hybrid chip step includes: integrating a symbolic reasoning unit and a neural network unit, performing calculation through parallel of the neural-symbol hybrid chip, and extracting unified features through cross-modal alignment of real-time monitoring data and historical fault records;

[0131] The aligned feature vector is injected into the node V of the cross-modal causal graph as an initial state for reasoning and prediction, and the node state of the causal graph G is dynamically updated according to the real-time data stream.

[0132] Preferably, the S4, the quantification of confidence step includes: based on the edge weight of the cross-modal causal graph G, applying a predefined causal rule, if multiple causal rules produce contradictory conclusions for the same node, arbitrating through confidence weighting, assigning a fault probability distribution to each node, and jointly quantifying the causal confidence, achieving the formula:

[0133] ,

[0134] In the formula, diagnostic confidence, temperature coefficient, controls the smoothness of the confidence distribution, and is captured through joint causal weight and feature similarity, captures the dynamic updated causal consistency, maps the joint quantization result to the interval [0, 1], and directly outputs the explainable confidence.

[0135] Preferably, the S5, low-confidence diagnosis result screening and causal subgraph extraction: according to historical data statistics, set the diagnostic confidence threshold Tn, screen out the diagnostic results of all nodes v with confidence , and extract the causal subgraph of the node v with high and low diagnostic confidence from the cross-modal causal graph.

[0136] ​Preferably, S5, the causal subgraph distillation step includes: a teacher model that predicts the causal subgraph based on a neural-symbol hybrid chip to output a high-precision causal chain, and a student model that constructs a lightweight model to learn the correlation causal logic and classification boundary of the causal subgraph of the teacher model;

[0137] In this embodiment, the prediction probability distribution of the nodes in the causal subgraph by the teacher model is used as a supervision signal, the student model is trained by a KL divergence loss function, explicit causal rules are extracted from the symbolic reasoning result of the teacher model, and the knowledge base of the student model is injected in a logical form, and a lightweight causal subgraph knowledge representation is obtained.

[0138] It should be understood that the edge knowledge base is stored in the form of a causal graph atlas fragment and a rule index table, the newly distilled causal subgraph knowledge is detected for conflict and merged with the existing knowledge in the edge knowledge base, the existing student model is fine-tuned using the new knowledge on the edge device, and the newly added entries in the edge knowledge base are time-stamped and versioned.

[0139] Preferably, S6, based on the causal rules extracted from the lightweight edge knowledge base, compressed model parameters, and low-confidence node supplementary data, the causal rules in the knowledge base and the device state are mapped to the virtual device model of the digital twin platform to construct a dynamic causal relationship network; and the compressed model parameters are deployed to the inference engine of the digital twin platform for real-time prediction of potential failure scenarios.

[0140] Specifically, based on the causal rules in the knowledge base, possible derivative failure scenarios are simulated, multi-dimensional failure hypotheses are generated in combination with real-time sensor data and historical failure patterns, risk scores of each failure scenario are calculated through the compressed model, the scenarios are prioritized using the causal chain, the failure types that need to be processed first are determined, the failure scenario with the highest risk score is taken as the root node, and the sub-nodes are gradually split downward, each node contains decision conditions, action suggestions, and expected benefits.

[0141] In this embodiment, the splitting logic of the decision tree is constrained by the causal rules, the prediction results of the lightweight model are introduced as the confidence evaluation of the leaf nodes, the inefficient branches are eliminated, the key actions are extracted from the decision tree as the task list of the robot inspection, and the initial path is generated in combination with the real-time state of the digital twin platform.

[0142] The task order is dynamically adjusted according to the urgency of the actions in the decision tree, the path feasibility is verified through virtual simulation of the digital twin, collision is avoided and the shortest path is selected, if a new failure signal is detected during the execution of the robot, local update of the decision tree is triggered and the path is re-planned, the actual execution result of the robot is compared with the predicted value of the decision tree, and the effectiveness of the decision tree is evaluated.

[0143] The new causal relationship discovered during the execution process is recorded through the digital twin platform, the new discovered causal relationship is injected into the edge knowledge base with the optimized decision tree logic, the updated knowledge base is fed back to the neuro-symbolic hybrid chip of S4 through the incremental learning mechanism of S5, and the global causal graph is optimized.

[0144] Preferably, S7 stores the whole-chain evidence based on the national secret algorithm blockchain, generates the space-time credible record, and feeds back the actual treatment effect of the robot to the cross-modal causal graph correction; obtains the fault diagnosis result, the quantitative confidence and the causal chain output by S4, obtains the optimal decision tree, the robot inspection path planning and the execution record generated by S6, and the sensor feedback of the actual execution action of the robot, and encodes the above evidence into a structured data block, which includes the following fields:

[0145] Timestamp: records the time of the event;

[0146] Geographical location: associated with the physical location of the device;

[0147] Evidence type: distinguish between diagnosis, decision or treatment evidence;

[0148] Causal chain identifier: reference to the node / edge identifier in the cross-modal causal graph G;

[0149] Hash value: calculate the hash value of the data block through the national secret algorithm to ensure data integrity;

[0150] Specifically, the structured data block is encrypted by the national secret algorithm, the encrypted data block is signed by the national secret SM2 algorithm, the data source is verified, the encrypted data block and the signature result are packaged as a blockchain transaction, and the national secret algorithm is written into the alliance chain supporting the national secret algorithm;

[0151] In this embodiment, a strongly trusted timestamp is obtained through a trusted timestamp agency, the GPS module and the Internet of Things sensor are combined to verify the authenticity of the geographical location, and the geographical coordinate hash value is written into the blockchain. The actual treatment effect is extracted from the sensor data after the robot executes the action, if the treatment effect is consistent with the diagnosis of S4, the weight of the related edge in the cross-modal causal graph G is updated, if the treatment effect is inconsistent with the diagnosis, the attribute label and the feature vector of the node v are adjusted, and the corrected cross-modal causal graph is fed back to the edge knowledge base and the digital twin platform.

[0152] Preferably, the machine room monitoring system of the intelligent patrol robot based on multi-source data fusion includes:

[0153] Multi-source space-time reference alignment module: through the quantum timestamp synchronization chip and the laser SLAM reflector positioning, the space-time reference alignment of multi-source heterogeneous sensor data is performed;

[0154] A fault semantic feature field construction module: based on spatiotemporal alignment data, visible light, depth, thermal and electromagnetic properties are fused in a three-dimensional spatial grid to construct a four-dimensional fault semantic feature field;

[0155] A cross-modal causal graph migration module: applying adversarial transfer learning on the four-dimensional fault semantic feature field, a cross-modal causal graph representing the fault evolution path is generated through a graph neural network;

[0156] A neural-symbol joint reasoning module: based on the cross-modal causal graph, rule reasoning and graph neural network prediction are simultaneously performed through a neural-symbol hybrid chip to output a diagnostic confidence;

[0157] An incremental causal distillation learning module: when the diagnostic confidence result is lower than a preset threshold, causal subgraph distillation is performed to inject a lightweight edge knowledge base through incremental learning;

[0158] A virtual-real interaction digital twin module: combined with the updated lightweight edge knowledge base, an optimal decision tree is generated by simulating the derived fault scenario on a digital twin platform to dynamically adjust the robot inspection path;

[0159] A fault tracing chain notarization module: based on a national encryption algorithm blockchain, the entire chain of evidence is sealed and stored, a spatiotemporal trusted notarization is generated, and the actual disposal effect of the robot is fed back to the cross-modal causal graph for correction.

[0160] Working principle: through quantum timestamp synchronization chips and laser SLAM retroreflective panel positioning technology, precise alignment of multi-source heterogeneous sensor data in time and space dimensions is achieved. Quantum timestamp synchronization chips eliminate microsecond-level time offset between sensors through atomic clock comparison mechanism and distributed clock synchronization protocol. Laser SLAM retroreflective panels correct SLAM cumulative error through fixed position constraints. Combined with a pre-defined coordinate mapping table, laser radar point cloud and global coordinate system are aligned, and non-laser sensor coordinate systems are unified through external parameter calibration. Based on the three-dimensional environment model constructed by laser SLAM, the physical space is divided into a regular three-dimensional voxel grid, each voxel corresponds to a spatial coordinate. Through spatiotemporal alignment data, multi-modal features are extracted, normalized, and early, middle, and late stage fusion is performed to generate high-dimensional feature vectors and map them into voxels. Through timestamp serialization of voxel features, a spatiotemporal tensor is constructed to capture dynamic change trends, and finally a four-dimensional fault semantic feature field containing spatial coordinates and semantic attributes is formed;

[0161] According to the alignment standard of the anti-migration learning framework, the feature distribution of the current computer room is aligned with that of the standard computer room, the source domain and the target domain feature distribution are aligned through the gradient reverse layer, and the cross-modal causal graph is constructed by combining the attention mechanism and the causal strength weight. The node is the sensor node, the edge is the physical connection, the functional dependence and the time sequence association, the feature vector is generated by aggregating the neighbor node information, and the causal relationship is encoded as the edge weight. The noise association is filtered by combining the causal mask, the stability of the graph is dynamically optimized, the fault evolution path is accurately captured, the symbolic reasoning unit and the neural network unit are integrated, the rule reasoning and the graph neural network prediction are executed in parallel through the neural-symbol hybrid chip, the node state of the cross-modal causal graph is dynamically updated, the pre-defined causal rules are applied based on the edge weight of the causal graph, the confidence weighted arbitration is used to resolve the contradictory conclusions, the fault probability distribution is assigned to the node, and the joint causal weight and the feature similarity are mapped to the [0, 1] interval to output the explainable diagnosis confidence. When the diagnosis confidence is lower than the threshold, the causal subgraph is extracted with the low-confidence node as the core, and the high-precision causal chain is compressed into a lightweight model through the teacher-student model distillation strategy. The student model learns the association logic and the classification boundary of the teacher model, and is trained through the KL divergence loss function. Explicit causal rules are extracted from the symbolic reasoning results and injected into the edge knowledge base. New knowledge is stored in the form of causal graph fragments and rule index table, supporting conflict detection and merging. The model is optimized through fine-tuning to solve the contradiction between limited computing resources of edge devices and insufficient generalization ability of the model, and to realize dynamic updating of knowledge. Combined with the updated edge knowledge base, a dynamic causal relationship network of the virtual device model is constructed in the digital twin platform, and the compressed model parameters are deployed for real-time prediction of potential fault scenarios. Based on the causal rules, the derived fault scenarios are simulated, multi-dimensional fault hypotheses are generated combined with real-time sensor data, the fault types that need to be processed in priority are determined through risk scoring and causal chain priority sorting, a decision tree is generated and the robot inspection path is dynamically adjusted, inefficient branches are eliminated, and the task order and path feasibility are optimized, realizing efficient fault response and adaptive optimization of inspection strategy. The data such as fault diagnosis results, decision tree and robot execution record are encrypted and signed through the national encryption algorithm, and structured data blocks are written into the alliance chain supporting national encryption. The spatiotemporal authenticity is verified through trusted timestamp and GPS, the data integrity is ensured through hash value, and the geographic coordinate hash value is recorded. The actual disposal effect of the robot is fed back to the causal graph. If it is consistent with the diagnosis, the edge weight is updated. If it is inconsistent, the node attribute is adjusted, forming a closed loop correction.

[0162] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

[0163] The above describes the present application and its embodiments, which are not limited, and the drawings only show one of the embodiments of the present application, and the actual structure is not limited thereto. In general, if a person skilled in the art is inspired thereby, without departing from the purpose of the present application, without creative design, similar structure and embodiments of the technical solution are not creative, and should belong to the protection scope of the present application.

Claims

1. A method for monitoring a computer room based on a multi-source data fusion intelligent inspection robot, characterized in that, The method comprises the following steps: S1, synchronizing the quantum timestamp chip with the laser SLAM reflector positioning, and performing space-time reference alignment of multi-source heterogeneous sensor data; S2, based on the space-time alignment data, fusing visible light, depth, thermal and electromagnetic properties in a three-dimensional space grid to construct a four-dimensional fault semantic feature field; S3, applying adversarial transfer learning on the four-dimensional fault semantic feature field, and generating a cross-modal causal graph representing the fault evolution path through a graph neural network; S4, based on the cross-modal causal graph, performing rule reasoning and graph neural network prediction simultaneously through a neural-symbol hybrid chip, and outputting a diagnosis confidence; S5, when the low-confidence diagnosis result is lower than the preset threshold, performing causal subgraph distillation, and injecting lightweight edge knowledge base through incremental learning; S6, combining the updated lightweight edge knowledge base, generating an optimal decision tree in a digital twin platform to simulate derived fault scenarios, and dynamically adjusting the robot inspection path; S7, based on the national encryption algorithm blockchain, sealing the whole chain of evidence, generating a space-time credible record, and feeding back the actual disposal effect of the robot to the cross-modal causal graph for correction; The S3, the step of constructing the cross-modal graph structure comprises: setting nodes and edges, the nodes include each sensor node as a node in the graph, the node features are the fusion of multi-modal data in the four-dimensional fault semantic feature field, the edges include physical connection relationship, functional dependence relationship and dynamic time sequence association, the feature vector of the node is generated by aggregating the neighbor node information, and the formula is: , In the formula, denotes the feature vector of node v at the l+1 layer, denotes the activation function, denotes the summation of all neighbor nodes u of node v, denotes the neighbor node set of node v, denotes the influence degree of node u on node v, W denotes a learnable weight matrix, denotes the feature vector of node u at the l layer the feature vector of the layer, denotes the causal strength between nodes u and v, encodes the causal relationship as the edge weight in the graph, forms a cross-modal causal graph G, and F denotes a feature extractor. 2.The method according to claim 1, characterized in that: The S3, the adversarial transfer learning step comprises: setting a source domain and a target domain, the source domain is a fault data set of a standard machine room, and the target domain is real-time four-dimensional fault semantic feature field data x of the current machine room; through the joint loss function of adversarial transfer learning, the feature distribution of the source domain and the target domain is aligned, and the formula is: ， In the formula, denotes the total value of the joint loss function, denotes the weight coefficient, denotes the source domain task loss, denotes the target domain task loss, denotes the domain adversarial loss, and F denotes a feature extractor. 3.The method of claim 2, wherein the method further comprises: The S3, the cross-modal causal graph optimization step comprises: combining the real-time four-dimensional fault semantic feature field, incrementally updating the edge weight in the causal graph, filtering noise association through a causal mask, and the formula is: ， In the formula, denotes the value of the edge weight between nodes u and v after the t+1 iteration, denotes the value of the edge weight between nodes u and v in the t iteration, denotes the learning rate, denotes the causal consistency loss the gradient of the edge weight , denotes the noise suppression coefficient, denotes the Gaussian noise term, and the matching degree of the artificial labeled causal chain and the cross-modal causal graph generation result is compared to test the stability of the cross-modal causal graph in different computer room scenarios.

4. The method according to claim 1, wherein the method further comprises: The S4, the neural-symbol hybrid chip step comprises: integrating a symbolic reasoning unit and a neural network unit, calculating in parallel through the neural-symbol hybrid chip, cross-modally aligning and extracting unified features from real-time monitoring data and historical fault records, injecting the aligned feature vector into the node V of the cross-modal causal graph as the initial state of reasoning and prediction, and dynamically updating the node state of the causal graph G according to the real-time data stream.

5. The method according to claim 4, wherein the method further comprises: The S4, the quantification confidence step comprises: based on the edge weight of the cross-modal causal graph G, applying a predefined causal rule to assign a fault probability distribution to each node, and jointly quantifying the causal confidence, the formula is: ， In the formula, represents the diagnostic confidence, represents the temperature coefficient, and is obtained by The joint causal weight and the feature similarity capture the dynamically updated causal consistency, map the joint quantization result to the [0, 1] interval, and directly output the explainable confidence. 6.The method of monitoring the computer room based on the multi-source data fusion intelligent inspection robot according to claim 1, characterized in that: S5, low confidence diagnosis result screening and causal subgraph extraction: according to historical data statistics, set the diagnosis confidence threshold Tn, screen out the diagnosis results of all nodes v With the core of nodes v with high and low diagnosis confidence, extract the causal subgraph from the cross-modal causal graph.

7. The method according to claim 6, wherein the method further comprises: The S5, the causal subgraph distillation step comprises: a teacher model and a student model, the prediction probability distribution of the nodes in the causal subgraph is used as a supervision signal through the teacher model, the student model is trained through the KL divergence loss function, the explicit causal rule is extracted from the symbolic reasoning result of the teacher model, and the rule is injected into the knowledge base of the student model in the form of logic, the lightweight causal subgraph knowledge representation, and the edge knowledge base is stored in the form of causal graph fragments and rule index table, the new distilled causal subgraph knowledge and the existing knowledge in the edge knowledge base are detected and merged, the new knowledge is used to fine-tune the existing student model on the edge device, and the new entries in the edge knowledge base are time-stamped and versioned.

8. The intelligent inspection robot-based machine room monitoring system implemented by the method of claim 1, comprising: Multi-source spatio-temporal reference alignment module: through the quantum timestamp synchronization chip and the laser SLAM reflector positioning, the spatio-temporal reference alignment of multi-source heterogeneous sensor data is performed; Fault semantic feature field construction module: based on the spatio-temporal alignment data, the visible light, depth, thermal and electromagnetic properties are fused in the three-dimensional space grid to construct a four-dimensional fault semantic feature field; Cross-modal causal graph migration module: applying adversarial transfer learning on the four-dimensional fault semantic feature field, a cross-modal causal graph representing the fault evolution path is generated through a graph neural network; Neural-symbol joint reasoning module: based on the cross-modal causal graph, rule reasoning and graph neural network prediction are synchronously performed through a neural-symbol hybrid chip to output a diagnosis confidence; Incremental causal distillation learning module: when the diagnosis confidence result is lower than the preset threshold, causal subgraph distillation is performed, and a lightweight edge knowledge base is injected through incremental learning; Virtual-real linkage digital twin module: combined with the updated lightweight edge knowledge base, an optimal decision tree is generated by simulating the derived fault scenario on a digital twin platform to dynamically adjust the robot inspection path; Fault tracing on-chain notarization module: based on the national secret algorithm blockchain, the whole chain evidence is stored, a spatio-temporal trusted notarization is generated, and the actual disposal effect of the robot is fed back to the cross-modal causal graph for correction.

Images