Multi-node signal source cooperative processing method based on POST-INA integrated module

Abstract

The invention discloses a multi-node signal source cooperative processing method based on a POST-INNA integrated module, and relates to the technical field of multi-node signal source cooperative processing, and the method comprises the following steps: obtaining an initial network topology of a plurality of node signal sources; a graph structure self-healing mechanism is triggered through node offline monitoring, graph structure updating is carried out on the initial network topology, and the graph structure self-healing mechanism comprises edge weight distribution, topology reconstruction, isolated connection and load balancing; fusing the obtained signal source feature matrix with the adjacent matrix of the updated network topology to obtain a fused feature matrix; and predicting the signal state of each node through a preset first LSTM model according to the fusion feature matrix, starting an adjacent node cooperative compensation mechanism according to the predicted node signal state, and cooperatively processing the multi-node signal source. The method is used for solving the problems of node failure processing delay, low cooperation efficiency and poor adaptive capability of a multi-node signal monitoring system in the prior art.

Description

Technical Field

[0001] This invention relates to the field of multi-node signal source collaborative processing technology, and more specifically, to a multi-node signal source collaborative processing method based on the POST-INA integrated module. Background Technology

[0002] Current X-ray device systems generally employ an independent signal source control structure, meaning each high-voltage generating unit, modulation driver, detector, or auxiliary control module operates independently, relying on a central control system for unified scheduling. However, in multi-node deployments and long-term operation scenarios, complex signal coupling effects (such as electromagnetic interference and chain feedback), state drift (caused by heat, aging, and jitter), and edge anomalous diffusion due to node degradation exist between nodes. Furthermore, as X-ray devices develop towards large-scale and modular designs, the number of nodes increases significantly, and their distribution exhibits spatial irregularity and heterogeneity, further increasing the complexity of system integration and fault-tolerant control. Therefore, in the complex operating environment of multi-node X-ray devices, the importance of collaborative signal source processing in multi-node X-ray devices becomes increasingly prominent.

[0003] For example, the invention patent with announcement number CN111902781B discloses a device and method for controlling a system, which includes the following steps: S1, the system receives signal data from multiple signal sources through an input interface; S2, the system includes a pre-trained neural network model for diagnosing and judging the control state of the system, the neural network including a multi-layer structure, each layer consisting of multiple nodes; S3, each node in the input layer and the first hidden layer corresponds to a specific signal source; S4, each node in the input layer and the first hidden layer corresponds to a specific signal source, realizing a partial connection structure, avoiding redundant connections of unrelated nodes, and improving modeling efficiency; S5, the processor inputs the signal into the neural network to generate the current system control state, and the controller executes corresponding control actions (such as adjusting signals, issuing commands, etc.) based on the control state output by the neural network.

[0004] For example, the invention patent with announcement number CN117278188A discloses a signal source synchronization system and its synchronization method, which includes the following steps: S1, multiple signal sources are cascaded in sequence, and the event output terminal of the upper-level signal source is connected to the trigger receiving terminal of the lower-level signal source; S2, each level signal source receives a trigger signal through its trigger receiving terminal, and after receiving it, transmits the trigger signal to the next level through its event output terminal; S3, after receiving the trigger signal, each level signal source outputs a modulated signal through its modulated signal output terminal; S4, the delay adjustment unit of each level signal source obtains the corresponding preset delay time, and adjusts the time interval between receiving the trigger signal and outputting the modulated signal according to the time; S5, through delay adjustment, the modulated signals of all cascaded signal sources are output synchronously.

[0005] The above-disclosed technical solutions have at least the following technical problems:

[0006] Existing signal source management and synchronization systems generally suffer from the following core technical problems: First, they rely on static connection methods and centralized control, resulting in delayed node failure handling, low coordination efficiency, and poor adaptability. Second, their synchronization methods are fixed and lack scalability, graph modeling, and edge intelligence, making it difficult to cope with dynamic changes and coordinated control requirements in complex systems, and resulting in insufficient overall adaptability and stability.

[0007] To address the above problems, this invention proposes a solution. Summary of the Invention

[0008] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a multi-node signal source collaborative processing method based on the POST-INA integrated module. By using the POST-INA integrated module as a foundation and integrating edge intelligent computing and graph neural network modeling, a multi-node collaborative processing system including self-healing mechanism, LSTM prediction and collaborative compensation is constructed, which solves the problems of node failure handling delay, low collaborative efficiency and poor adaptive capability in the existing multi-node signal monitoring system.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] A collaborative processing method for multi-node signal sources based on a POST-INA integrated module includes the following steps: acquiring the initial network topology of several node signal sources; updating the initial network topology by triggering a graph structure self-healing mechanism through offline node monitoring, wherein the graph structure self-healing mechanism includes edge weight allocation, topology reconstruction, isolated connections, and load balancing; fusing the acquired signal source feature matrix with the adjacency matrix of the updated network topology to obtain a fused feature matrix; predicting the signal state of each node using a preset first LSTM model based on the fused feature matrix, and initiating a neighbor node collaborative compensation mechanism based on the predicted node signal state to collaboratively process the multi-node signal sources.

[0011] In a preferred embodiment, the node signal source includes a high-voltage power supply module, a power control unit, a radiation detector, and a POST-INA integrated module.

[0012] In a preferred embodiment, the acquisition of the initial network topology of several node signal sources is specifically as follows: the physical connection relationship between nodes is identified and recorded through the RS485 communication protocol to form a preliminary physical connection graph; based on the physical connection relationship, the normalized mutual information weight of any adjacent node pair is calculated; and a delay attenuation term is introduced, combined with historical fault records, to obtain the edge weight of any adjacent node pair; an initial adjacency matrix is ​​constructed based on the edge weight of any adjacent node pair to obtain the initial network topology.

[0013] In a preferred embodiment, the graph structure self-healing mechanism includes edge weight allocation, topology reconstruction, isolated connections, and load balancing, specifically as follows: Calculate the impact factor of offline nodes on neighboring nodes, where the impact factor is determined based on the historical state correlation between offline nodes and neighboring nodes; acquire isolated nodes caused by node offline status, calculate isolated weights based on the minimum impact factor and state similarity, and reconstruct isolated node edges; generate a repaired topology based on isolated node edges and isolated weights by minimizing the Laplace matrix perturbation; connect nodes according to the impact factor and neighbor overlap, calculate node variance, and adjust isolated node edge weights using a preset load balancing edge weight correction formula.

[0014] In a preferred embodiment, the process of fusing the acquired signal source feature matrix with the adjacency matrix of the updated network topology to obtain a fused feature matrix is ​​as follows: Signal source features are extracted based on current and historical signal data. These features include first-order physical features, second-order frequency domain statistical features, and third-order cross-module residual features. The signal source features are then input into the adjacency matrix of the updated network topology, fused using a graph convolutional network, and a cross-layer residual connection mechanism is introduced to obtain the fused feature matrix.

[0015] In a preferred embodiment, the step of predicting the signal state of each node using a preset first LSTM model and initiating a neighboring node collaborative compensation mechanism based on the predicted node signal state is as follows: the first LSTM model is used to analyze historical time step data to predict the future node signal state and confidence interval; when the residual between the predicted value and the actual value of the node signal state exceeds a first threshold, the collaborative compensation mechanism is triggered, and the signal source features are marked as abnormal.

[0016] In a preferred embodiment, the collaborative compensation mechanism includes a neighboring node signal buffering compensation mechanism, a collaborative interpolation compensation mechanism, and a redundant node activation and dynamic route reconstruction mechanism.

[0017] In a preferred embodiment, after the collaborative processing of multi-node signal sources, the method further includes: uploading the marked abnormal signal source features to a cloud server; the cloud server running a preset second LSTM model to process the abnormal signal source features and outputting the system node signal status; comparing the prediction results of the second LSTM model and the first LSTM model in the same time interval and calculating the prediction deviation between the models; when the prediction deviation exceeds a preset second threshold, extracting the weight difference between the two models to construct an incremental update package; and remotely distributing the weight update package to local nodes to dynamically optimize the parameters of the first LSTM model.

[0018] The technical effects and advantages of the multi-node signal source collaborative processing method based on the POST-INA integrated module of this invention are as follows:

[0019] This invention constructs a multi-node intelligent sensing network for signal sources with self-diagnosis, self-repair, and local self-adjustment capabilities by introducing a graph structure self-healing mechanism and a multi-level neighbor node collaborative compensation mechanism. At the structural level, the self-healing mechanism can identify network breakpoints in real time when nodes are offline or the network is damaged. It dynamically reconstructs topology edges based on impact factors and state similarity, and combines minimum Laplace perturbation and edge weight optimization strategies to maintain the continuity and stability of the topology, significantly enhancing system robustness and adapting to complex industrial environments. At the functional level, the collaborative compensation mechanism accurately identifies anomaly levels based on LSTM prediction residuals and quickly schedules healthy neighbor nodes for signal compensation and functional replacement, effectively suppressing the risk of local fault propagation. The two mechanisms work together to achieve a closed-loop control process from "prediction-compensation-repair-optimization," exhibiting high adaptability, low manual dependence, and strong environmental adaptability, significantly improving the system's stability, intelligence, and sustainable operation capabilities. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the structure of a multi-node signal source collaborative processing method based on a POST-I NA integrated module according to the present invention.

[0021] Figure 2 This is a diagram of a graph structure self-healing mechanism provided in an embodiment of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Example 1, Figure 1This invention presents a multi-node signal source collaborative processing method based on a POST-INA integrated module, comprising the following steps:

[0024] S1, Obtain the initial network topology of several node signal sources;

[0025] S2, triggering the graph structure self-healing mechanism through offline node monitoring, updating the graph structure of the initial network topology. The graph structure self-healing mechanism includes edge weight allocation, topology reconstruction, isolated connections, and load balancing.

[0026] S3, fuse the acquired signal source feature matrix with the adjacency matrix of the updated network topology to obtain the fused feature matrix;

[0027] S4. Based on the fusion feature matrix, the signal state of each node is predicted by the preset first LSTM model, and the neighbor node collaborative compensation mechanism is activated based on the predicted node signal state to collaboratively process the multi-node signal source.

[0028] This embodiment constructs a multi-node intelligent sensing network for signal sources with self-diagnosis, self-repair, and local self-adjustment capabilities by introducing a graph structure self-healing mechanism and a multi-level neighbor node collaborative compensation mechanism. At the structural level, the self-healing mechanism can identify network breakpoints in real time when nodes are offline or the network is damaged. It dynamically reconstructs topology edges based on impact factors and state similarity, and combines minimum Laplace perturbation and edge weight optimization strategies to maintain the continuity and stability of the topology, significantly enhancing system robustness and adapting to complex industrial environments. At the functional level, the collaborative compensation mechanism accurately identifies anomaly levels based on LSTM prediction residuals and quickly schedules healthy neighbor nodes for signal compensation and functional replacement, effectively suppressing the risk of local fault propagation. The two mechanisms work together to achieve a closed-loop control process from "prediction-compensation-repair-optimization," exhibiting high adaptability, low manual dependence, and strong environmental adaptability, significantly improving the system's stability, intelligence, and sustainable operation capabilities.

[0029] S1, obtain the initial network topology of several node signal sources.

[0030] In this embodiment, the node signal source includes a high-voltage power supply module, a power control unit, a radiation detector, and a POST-INA integrated module.

[0031] In this embodiment, a secure network is constructed based on the signal source node through dynamic environment adaptation and chain-based identity authentication, as detailed below:

[0032] The system collects environmental noise signals and calculates the power spectral density, simultaneously acquires temperature sensor data, and dynamically calibrates the sampling voltage of the acquired signals using a temperature drift compensation formula.

[0033] Each signal source node performs two-way authentication using the ECDSA algorithm. After successful authentication, an encrypted communication channel is established to build a secure network.

[0034] The formula for calculating the power spectral density of noise is as follows:

[0035]

[0036] In the formula, P(f) is the power spectral density, T is the signal sampling time, f is the frequency component, S(t) is the time-domain noise signal, and e -j2πft It is a complex exponential function, corresponding to the frequency component of the Fourier transform.

[0037] The specific formula for temperature drift compensation is as follows:

[0038]

[0039] In the formula, V raw The original sampling voltage, V cal The compensated voltage, α is the temperature drift coefficient, T ref This is a reference temperature.

[0040] In this embodiment, each signal source node authenticates the two-way signature using the ECDSA algorithm. After successful authentication, an encrypted communication channel is established to construct a secure network, as detailed below:

[0041] The master node generates an ECDSA key pair (using secp256r1 elliptic curve), including a private key Sk1 and a public key Pk1, and embeds the public key into the device digital certificate;

[0042] The slave node sends an authentication request to the master node, which includes: the slave node's own digital certificate Cert2 and a random number R1 generated by a true random number generator;

[0043] After receiving R1, the master node signs R1 using its private key Sk1, generating Sig1 = Sign_Sk1(R1). Subsequently, the master node returns to the slave node: the signature Sig1 and a new random number R2 randomly generated by the master node.

[0044] The slave node verifies the validity of Sig1 to confirm the identity of the master node. If the verification is successful, the slave node uses its own private key Sk2 to sign R2, generating Sig2 = Sign_Sk2(R2) and sending it to the master node.

[0045] The master node verifies Sig1, completing the authentication of the slave node. At this point, two-way authentication is complete, and the identities of both communicating parties have been confirmed.

[0046] The master node and slave node each use the other's public key and their own private key to negotiate and generate a shared key K based on the ECDH protocol.shared :

[0047] K shared =ECDH(SK2,PK2)=ECDH(SK2,PK1)

[0048] Based on K shared A session key K for symmetric encryption is derived from KDF. AES ;

[0049] Both communicating parties establish a secure communication channel using AES-128-CTR mode. The initial vector IV is provided by a hardware true random number generator (TRNG) to ensure that the communication has anti-replay and forward security capabilities.

[0050] In this embodiment, the acquisition of the initial network topology of several node signal sources is specifically as follows:

[0051] The physical connection relationships between nodes are identified and recorded using the RS485 communication protocol to form a preliminary physical connection diagram;

[0052] Based on physical connection relationships, calculate the normalized mutual information weight of any pair of adjacent nodes.

[0053] Furthermore, a delay attenuation term is introduced, and combined with historical fault records, the edge weights of any pair of adjacent nodes are obtained;

[0054] The initial adjacency matrix is ​​constructed based on the edge weights of any pair of adjacent nodes, thus obtaining the initial network topology.

[0055] It should be noted that the initial network topology is obtained based on a secure network setup.

[0056] In this embodiment, based on physical connection relationships, the normalized mutual information weight of any pair of adjacent nodes is calculated, and a delay attenuation term is introduced. Combined with historical fault records, the edge weight of any pair of adjacent nodes is obtained, as follows:

[0057] For each pair of connection points (i,j), calculate the following three sub-weights:

[0058] The specific formula for mutual information weighting is as follows:

[0059]

[0060] In the formula, w MI (i,j) are the mutual information weights, I(x) i ;x j ) represents the mutual information of information sequences.

[0061] The formula for the delay attenuation term is as follows:

[0062]

[0063] In the formula, w delay (i,j) represents the delay decay term, T ij T0 is the average signal propagation delay between nodes, and T0 is a normalization constant.

[0064] Historical fault correlation w fail (i,j): Normalized fault occurrence frequency.

[0065] The final weighted sum formula for edge weights is as follows:

[0066] w ij =λ1w MI +λ2w delay +λ3w fail

[0067] In the formula, w ij Let λ1, λ2, and λ3 be the final edge weight between node i and node j, and let λ1, λ2, and λ3 be the weight coefficients, satisfying λ1 + λ2 + λ3 = 1.

[0068] In this embodiment, an initial adjacency matrix is ​​constructed based on the edge weights of any pair of adjacent nodes to obtain the initial network topology.

[0069] The specific formula for defining a matrix is ​​as follows:

[0070] A = [a ij ] n×n (n=|V|)

[0071] In the formula, A is the initial adjacency matrix; a ij Let a be the connection weight between nodes i and j. ij ∈[0,1]; n is the total number of nodes in the network; i and j are node numbers; V is the set of communication nodes (devices).

[0072] The specific rules for element assignment are as follows:

[0073]

[0074] Note: a ii =0 (i∈{1,2,...,n})

[0075] In the formula, w ij Let e ​​be the edge weights of nodes i and j; ij Let be the edge that physically connects nodes i and j; E is the set of all physically connected edges; a ii This is the node's own connection weight, fixed at 0, indicating that the node does not form a valid communication link with itself.

[0076] S2, triggering the graph structure self-healing mechanism through offline node monitoring, updates the graph structure of the initial network topology. The graph structure self-healing mechanism includes edge weight allocation, topology reconstruction, isolated connections, and load balancing.

[0077] In this embodiment, the graph structure self-healing mechanism triggered by offline node monitoring updates the initial network topology. This self-healing mechanism includes edge weight allocation, topology reconstruction, isolated connections, and load balancing, as detailed below:

[0078] Calculate the impact factor of the offline node on the neighboring node, the impact factor being determined based on the historical state correlation between the offline node and the neighboring node;

[0079] Identify isolated nodes caused by node offline, calculate the isolation weight based on the minimum impact factor and state similarity, and reconstruct the edges of isolated nodes;

[0080] Based on isolated node edges and isolated weights, a repaired topology is generated by minimizing the Laplace matrix perturbation.

[0081] Connect nodes based on the impact factor and neighbor overlap, calculate node variance, and adjust the edge weights of isolated nodes using a preset load balancing edge weight correction formula.

[0082] In this embodiment, the impact factor of the offline node on neighboring nodes is calculated. The impact factor is determined based on the correlation between the historical states of the offline node and its neighboring nodes, as follows:

[0083] When node k goes offline, its impact factor on each neighbor node i is calculated. This factor reflects the magnitude of the impact of node k going offline on the stability of its neighbors, and is specifically determined by the ratio of the state covariance of the two nodes to the state variance of node k.

[0084] When node k is offline, the specific formula for defining the impact factor on its neighboring node i is as follows:

[0085]

[0086] In the formula, δ ik Let w be the impact factor of offline node k in neighboring node i. ik Let s be the edge weight between nodes i and k. i s k For node state vectors, Let be the variance of node i, representing the strength of the influence of node k offline on the stability of node i, and let Cov(s) be the variance of node i. i ,s k Var(s) represents the covariance between the state vectors of node i and node k. k ) represents the variance of the state vector of node k.

[0087] In this embodiment, isolated nodes caused by node offline are obtained, and isolation weights are calculated based on the minimum impact factor and state similarity. The edges of the isolated nodes are then reconstructed. The isolation weights are obtained based on the minimum impact factor and the obtained state similarity, as follows:

[0088] For each neighboring node affected by the offline node, a new edge is created. The weight of the new edge is the weight of the original disconnected edge multiplied by the smaller of the impact factors of the two nodes on the offline node k (i.e., min(δ)). ik ,δ jk The system dynamically allocates nodes by considering the state similarity of the nodes at both ends of the new edge (adjusted by an exponential decay function).

[0089] For each original edge (i,k) of offline node k, its weight is adjusted by δ. ik The new edge (i,j) is dynamically assigned, and the specific formula is as follows:

[0090]

[0091] In the formula, Let δ be the edge weight of the reconstructed edge (i,j). ik Let δ be the impact factor of offline node k on neighboring node i. jk Let w be the impact factor of offline node k at neighboring node j. ik N represents the edge weight of the original edge (i,k). off For offline node sets, The weights are for state similarity (T is the temperature parameter).

[0092] In this embodiment, based on isolated node edges and isolated weights, the repaired topology is generated by minimizing the Laplace matrix perturbation, as detailed below:

[0093] The topology is repaired by adjusting the Laplacian matrix of the graph. Specifically, the parts corresponding to offline nodes are removed, and then the topology is reconstructed by minimizing the spectral difference (i.e., minimizing the Frobenius norm difference between the new Laplacian matrix and the Laplacian matrix after removing offline nodes), while maintaining the connectivity of the graph.

[0094] The specific formula for defining the structural gap matrix ΔL is as follows:

[0095]

[0096] In the formula, ΔL is the structural gap matrix, L is the graphical Laplace matrix, and e k Let N be the standard basis vector of node k. off This is an offline node set.

[0097] The specific formula for reconstructing the topology by minimizing the spectral difference is as follows:

[0098]

[0099] Constraint: rank(L') = n - |N off |

[0100] In the formula, L' is the reconstructed Laplacian matrix, ΔL is the structural gap matrix, F is the Frobenius norm, λ is the weight, and tr(L'11) T Let ) be the inner product of matrix L' and the matrix of all 1s, rank(L') be the rank of the reconstructed Laplace matrix, n be the number of nodes, and N be the number of nodes. off This is an offline node set.

[0101] In this embodiment, the connection of nodes based on the impact factor and the obtained neighbor overlap is specifically as follows:

[0102] For nodes that become isolated due to going offline, select node pairs to fill the gaps. The selection criteria are to prioritize connecting node pairs with large aggregation impact factors (i.e., those greatly affected by offline nodes) and high neighbor overlap, in order to quickly restore connectivity.

[0103] For the isolated node set I caused by k going offline, the replacement node pair (a,b) is selected as follows:

[0104]

[0105] In the formula, (a * ,b * For optimal edge-filling node pairs, a,b∈I means that nodes a and b must be nodes in the isolated node set I. Let be the impact factors of nodes a and b respectively, and Comm(a,b) be the neighbor overlap between nodes a and b.

[0106] In this embodiment, the node variance is calculated, and the edge weights of isolated nodes are adjusted in conjunction with a preset load balancing edge weight correction formula, as follows:

[0107] After assigning new edge weights, a load balancing adjustment is performed based on the load of each node (represented by variance). Specifically, a load factor (1 minus the relative deviation of the node variance) is multiplied by the new edge weight, so that nodes with high loads are assigned fewer weights, thereby balancing the load of the entire network.

[0108] The specific formula for adjusting the load balancing edge weights is as follows:

[0109]

[0110] In the formula, For the final edge weights, Let σ be the original edge weight between node i and node j. iLet be the variance of node i. The mean variance of all nodes at present, max m σ m This represents the maximum variance of all nodes in the network.

[0111] Figure 2 This is a diagram of a graph structure self-healing mechanism provided in an embodiment of the present invention.

[0112] S3. The obtained signal source feature matrix is ​​fused with the adjacency matrix of the updated network topology to obtain the fused feature matrix.

[0113] In this embodiment, the process of fusing the acquired signal source feature matrix with the adjacency matrix of the updated network topology to obtain the fused feature matrix is ​​as follows:

[0114] Signal source features are extracted based on current and historical signal data. These signal source features include first-order physical features, second-order frequency domain statistical features, and third-order cross-module residual features.

[0115] The signal source features are input into the adjacency matrix of the updated network topology, fused through a graph convolutional network, and a cross-layer residual connection mechanism is introduced to obtain the fused feature matrix.

[0116] In this embodiment, the first-order physical characteristics include effective voltage, effective current, average power, and main signal frequency; the second-order frequency domain statistical characteristics include power fluctuation rate, power spectral density, and STFT (Short Time Fourier Transform) energy; and the third-order cross-module residual characteristics include the actual output signal of the current node, the predicted signal between adjacent nodes, and the residual signals between multiple signal modules that affect each other due to physical connections, signal interference, or structural resonance.

[0117] The specific formula for defining the characteristic matrix of a signal source is as follows:

[0118] X∈R N×F

[0119] In the formula, X is the feature matrix of the signal source, R is the set of real numbers, N is the number of nodes, and F is the number of feature dimensions of each node.

[0120] Once the network topology completes its self-healing update and obtains a new adjacency matrix, the specific formula for defining the adjacency matrix of the updated network topology is as follows:

[0121] A (new) ∈R n×n

[0122] In the formula, A (new) R is the adjacency matrix of the updated network topology; R is the set of real numbers, and n is the number of nodes.

[0123] In this embodiment, a graph convolutional neural network is introduced to fuse and model the updated adjacency matrix with the signal source features.

[0124] The specific formula for fusion modeling is as follows:

[0125]

[0126] In the formula, H (l+1) To obtain the output feature matrix of the (l+1)th layer after calculation; H (l) Let l be the input feature matrix of the l-th layer; l is the layer number in the graph convolutional neural network. The normalized adjacency matrix is ​​calculated as follows: D is the node degree matrix; W (l) Let be the learnable weight matrix of the graph convolution at layer l; σ(·) is the activation function.

[0127] In this embodiment, to enhance the expressive power of the deep model and alleviate the oversmoothing of features, a cross-layer residual connection mechanism is introduced to fuse the outputs of multiple layers. The specific formula is as follows:

[0128]

[0129] In the formula, H fusion H represents the fused residual feature matrix; L represents the total number of layers in the graph convolutional network; H represents the number of layers in the graph convolutional network. (L) α represents the final output feature of the Lth layer of the graph convolutional network. l H represents the residual weighting coefficients corresponding to the feature matrix of the l-th layer; (l) Let be the input feature matrix of the l-th layer.

[0130] In this embodiment, the fused feature matrix is ​​normalized and activated to stabilize the training process and enhance its expressive power. The specific formula is as follows:

[0131] H out =Normalize(σ(H) fusion ))

[0132] In the formula, H out σ is the final output fused feature matrix; Normalize(·) is the normalization function (LayerNorm or BatchNorm normalization can be used); σ(·) is the nonlinear activation function.

[0133] S4. Based on the fusion feature matrix, the signal state of each node is predicted by the preset first LSTM model, and the neighbor node collaborative compensation mechanism is activated based on the predicted node signal state to collaboratively process the multi-node signal source.

[0134] In this embodiment, the step of predicting the signal state of each node using a preset first LSTM model and initiating a neighbor node collaborative compensation mechanism based on the predicted node signal state is as follows:

[0135] By analyzing historical time step data using the first LSTM model, the future node signal state and confidence interval are predicted.

[0136] When the residual between the predicted and actual values ​​of a node signal state exceeds the first threshold, a collaborative compensation mechanism is triggered, and the signal source features are marked as abnormal.

[0137] It should be noted that the node signal state includes signal strength state, frequency domain characteristic state, and time series state.

[0138] In this embodiment, the prediction of future node states and confidence intervals by analyzing historical time step data through the first LSTM model is as follows: input feature data of 60 historical time steps, and output the predicted state values ​​and confidence intervals for the next 5 steps.

[0139] For each time step τ∈[t-59,t], the state calculation inside the first LSTM model is as follows:

[0140] Forgotten Gate: f τ =σ(W f ·[h τ-1 ,x τ ]+b f )

[0141] Input gate: i τ =σ(W i ·[h τ-1 ,x τ ]+b i )

[0142]

[0143] Cell status update:

[0144] Output gates and hidden states: o τ =σ(W o ·[h τ-1 ,x τ ]+b o )

[0145] h τ =o τ ⊙tanh(C τ )

[0146] In the formula, τ is the current time step; t is the current time; f τThe output of the forget gate controls how much of the memory state from the previous time step is retained; its value is in the range [0,1]. τ-1 ,x τ [] represents the previous hidden state and the current input to be concatenated; σ is the Sigmoid activation function, with an output range of (0,1); W f h is the weight matrix of the forget gate; τ-1 The hidden state of the previous time step; x τ b is the input feature vector at the current time step; f Forget gate bias term; i τ For input gate output; W i b is the weight matrix of the input gate; i This is the bias term for the input gate; For candidate memory states; tanh is the hyperbolic tangent function, with an output range of [-1, 1]; W C b is the candidate state weight matrix; C For candidate state bias terms; C τ The cell state at the current time step; C τ-1 The cell state at the previous time step; o τ For output gate output; W o b is the weight matrix of the output gate; o For the bias term of the output gate; h τ This represents the hidden state at the current time step, and is also the LSTM output at that moment.

[0147] Finally, the model will hide the state h at the last moment. t The input is fed into a fully connected layer to generate a multi-step prediction sequence for the future:

[0148] The confidence interval is:

[0149] In the formula, ε represents the output value predicted by the model at time step t+k, indicating the state at the future time step k predicted by the first LSTM model based on the feature data of the previous 60 time steps, where k∈{1,2,3,4,5}; k h represents the error range predicted at time step t+k. t The hidden state at the last time step; W out b is the weight matrix of the fully connected output layer; out This is the bias term for the fully connected output layer.

[0150] The formula for predicting residuals is as follows:

[0151]

[0152] In the formula, residual(t+k) is the predicted residual value. The model predicts the state value at time step t+k, and y(t+k) is the actual observed value at time step t+k.

[0153] In this embodiment, the abnormal node signal state is recorded when the prediction residual exceeds the first threshold, and the signal source features are marked as abnormal.

[0154] In this embodiment, the predicted residual (t+k) is compared with a set anomaly threshold θ, employing a multi-level triggering mechanism. θ1 is the first threshold, and θ2 and θ3 are other thresholds used to classify the levels. Specifically, it is divided into the following three levels:

[0155] Level I anomaly (mild fluctuation): If the residual (t+k) satisfies θ1<residual(t+k)≤θ2, it is judged as a Level I anomaly;

[0156] Level II anomaly (moderate deviation): If the residual (t+k) satisfies θ2<residual(t+k)≤θ3, it is judged as a Level II anomaly;

[0157] Level III anomaly (severely out of control): If the residual (t+k) satisfies residual(t+k)>θ3, it is judged as a Level III anomaly.

[0158] In this embodiment, the collaborative compensation mechanism includes a neighboring node signal buffering compensation mechanism, a collaborative interpolation compensation mechanism, and a redundant node activation and dynamic route reconstruction mechanism.

[0159] In this embodiment, based on the determined trigger level, corresponding neighbor node collaborative compensation measures are initiated, as follows:

[0160] Level I anomaly (mild fluctuation):

[0161] When the node prediction residual meets the mild fluctuation threshold (i.e., Level I anomaly), the system activates the neighboring node signal buffer compensation mechanism, specifically as follows:

[0162] Retrieve historical stable signal data of the direct neighboring nodes of the current abnormal node;

[0163] The data is then weighted and averaged within a time window, and used as a short-term alternative input for outlier nodes.

[0164] The status of abnormal nodes is adjusted in real time and recorded as a temporary status mapping;

[0165] At the same time, keep the original prediction module running online and periodically assess the state recovery status.

[0166] Level II anomaly (moderate shift):

[0167] When the node prediction residual reaches the moderate offset threshold (i.e., Level II anomaly), the system initiates a collaborative interpolation compensation mechanism, specifically as follows:

[0168] Construct a feature graph structure of an abnormal node and its first-order neighborhood nodes;

[0169] A weighted distance-based collaborative interpolation algorithm is used to compensate for and predict the state of the current abnormal node. The compensation formula is as follows:

[0170]

[0171] In the formula, Let N(i) be the compensation state value of the abnormal node i at time t, and let N(i) be its set of neighboring nodes. ij ∈[0,1] represents the signal similarity weights between outlier node i and node j, s j (t) represents the state value of the adjacent point j.

[0172] Level III anomaly (severely out of control):

[0173] When the node prediction residual exceeds the severe runaway threshold (i.e., Level III anomaly), the system initiates a redundant node activation and dynamic route reconstruction mechanism, specifically as follows:

[0174] Determine whether redundant nodes (backup sensors / channels) are deployed in the subnetwork to which the current abnormal node belongs;

[0175] If redundant nodes exist, their sensing channels are activated to take over the signal.

[0176] Meanwhile, the graph neural network structure adjustment module recalculates the edge weights and communication paths of the network graph structure and performs dynamic route reconstruction to bypass severely failed nodes.

[0177] Mark the node as "out of control" and push it to the remote monitoring terminal for manual intervention and assessment;

[0178] If there are no redundant nodes, a "node offline" warning is sent to the edge processing module, and environment-level compensation logic is activated.

[0179] In this embodiment, after the collaborative processing of the multi-node signal source, the method further includes:

[0180] The characteristics of the marked abnormal signal sources are uploaded to the cloud server;

[0181] The cloud server runs a pre-set second LSTM model to process the features of abnormal signal sources and outputs the signal status of system nodes.

[0182] Compare the prediction results of the second LSTM model and the first LSTM model in the same time interval, and calculate the prediction bias between the models;

[0183] When the prediction deviation exceeds the preset second threshold, the weight difference between the two models is extracted to construct an incremental update package;

[0184] The weight update package is remotely sent to the local node to dynamically optimize the parameters of the first LSTM model.

[0185] In this embodiment, the features of node signal sources marked as abnormal during the collaborative compensation process are uploaded to the cloud server. After receiving the uploaded data, the cloud server uses the second LSTM model to independently predict the node signal state. The specific prediction formula is as follows:

[0186] The internal state calculation of the second LSTM model is as follows:

[0187] Forgotten Gate: f t =σ(A f ·x t +B f ·h t-1 +d f )

[0188] Input gate: i t =σ(A i ·x t +B i ·h t-1 +d i )

[0189]

[0190] Cell status update:

[0191] Output gates and hidden states: o t =σ(A o ·x t +B o ·h t-1 +d o )

[0192] h t =o t ⊙tanh(c t )

[0193] In the formula, t is the current time step; f t σ is the forget gate control value, ranging from [0,1]; σ is the Sigmoid activation function, with an output range of (0,1); A f A i A c A oB is the input weight matrix (used for the forget gate, input gate, candidate unit, and output gate, respectively); f B i B c B o Here are the hidden state weight matrices (used for the forget gate, input gate, candidate units, and output gate, respectively); d f d i d c d o These are bias terms (used in the forget gate, input gate, candidate unit, and output gate, respectively); h t-1 The hidden state of the previous time step; x t i is the input feature vector at the current time step; t For input gate output; Candidate memory states; tanh is the hyperbolic tangent function, with an output range of [-1, 1]; c t The cell state at the current time step; c t-1 The cell state at the previous time step; o t For output gate output; h t This represents the hidden state at the current time step, and is also the LSTM output at that moment.

[0194] The final output is transformed into a predicted value through a linear layer:

[0195] In the formula, h t The hidden state at the last time step; W y b is the weight matrix of the fully connected output layer; y This is the bias term for the fully connected output layer.

[0196] In this embodiment, the first LSTM model is retrieved. Historical prediction results, compared with the second LSTM model The prediction results are compared, and the prediction deviation is calculated using the following formula:

[0197]

[0198] In the formula, ΔY represents the prediction bias; The results are the predictions of the first LSTM model and the second LSTM model, respectively.

[0199] The normalized mean squared error (nMSE) index is calculated using the following formula:

[0200]

[0201] In the formula, nMSE is the normalized mean square error; N is the number of comparison nodes; Let Variance be the historical error variance of the i-th node; The results are the predictions of the first LSTM model and the second LSTM model, respectively.

[0202] In this embodiment, a second threshold δ2 for system deviation is set. If the following conditions are met:

[0203] nMSE>δ2

[0204] If the prediction performance of the first LSTM model deviates from the normal range, the remote dynamic optimization mechanism will be triggered.

[0205] In this embodiment, after comparing the parameter sets of the second LSTM model and the first LSTM model in the cloud, the weight difference vector is calculated, and the specific formula is as follows:

[0206] Δθ=θ cloud -θ edge

[0207] In the formula, Δθ is the difference vector between the parameters of the cloud and edge models; θ edge θ cloud These are the sets of all trainable parameters in the first and second LSTM models, respectively.

[0208] In this embodiment, an incremental update package is constructed based on the sparsity threshold ε, and the specific formula is as follows:

[0209] P update ={Δθ j ||Δθ j >ε}

[0210] In the formula, P update For incremental update of the package; Δθ j ε represents the difference of the j-th parameter; ε is the sparsity threshold, set to a positive number to filter parameter differences with significant changes. This update package contains all significant parameter offsets for subsequent remote distribution, reducing communication bandwidth burden.

[0211] In this embodiment, after the update package is sent to the POST-INA module via an encrypted channel, the local LSTM model parameters are updated. The specific expression is as follows:

[0212] θ edge ←θ edge +η·P update

[0213] In the formula, P update For incremental update packages, η is the dynamic learning rate, η = f(CI, ρ), CI is the confidence interval width of the current edge node model output, and ρ is the current computational load of the edge node; θ edge This is the set of trainable parameters for the first LSTM model.

[0214] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0215] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0216] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0217] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0218] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0219] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-node signal source collaborative processing method based on a POST-INA integrated module, characterized in that, Includes the following steps: Obtain the initial network topology with several node signal sources; The graph structure self-healing mechanism is triggered by offline node monitoring to update the initial network topology. The graph structure self-healing mechanism includes edge weight allocation, topology reconstruction, isolated connections, and load balancing. The obtained signal source feature matrix is ​​fused with the adjacency matrix of the updated network topology to obtain the fused feature matrix; Based on the fusion feature matrix, the signal state of each node is predicted by the preset first LSTM model, and the neighbor node collaborative compensation mechanism is activated based on the predicted node signal state to collaboratively process the multi-node signal source.

2. The multi-node signal source collaborative processing method based on a POST-INA integrated module according to claim 1, characterized in that, The node signal source includes a high-voltage power supply module, a power control unit, a radiation detector, and a POST-INA integrated module.

3. The multi-node signal source collaborative processing method based on a POST-INA integrated module according to claim 2, characterized in that, The initial network topology for obtaining the signal sources from several nodes is as follows: The physical connection relationships between nodes are identified and recorded using the RS485 communication protocol to form a preliminary physical connection diagram; Based on physical connection relationships, calculate the normalized mutual information weight of any pair of adjacent nodes. Furthermore, a delay attenuation term is introduced, and combined with historical fault records, the edge weights of any pair of adjacent nodes are obtained; The initial adjacency matrix is ​​constructed based on the edge weights of any pair of adjacent nodes, thus obtaining the initial network topology.

4. The multi-node signal source collaborative processing method based on a POST-INA integrated module according to claim 3, characterized in that, The graph structure self-healing mechanism, triggered by offline node monitoring, updates the initial network topology. This self-healing mechanism includes edge weight allocation, topology reconstruction, isolated connections, and load balancing, as detailed below: Calculate the impact factor of the offline node on the neighboring node, the impact factor being determined based on the historical state correlation between the offline node and the neighboring node; Identify isolated nodes caused by node offline, calculate the isolation weight based on the minimum impact factor and state similarity, and reconstruct the edges of isolated nodes; Based on isolated node edges and isolated weights, a repaired topology is generated by minimizing the Laplace matrix perturbation. Connect nodes based on the impact factor and neighbor overlap, calculate node variance, and adjust the edge weights of isolated nodes using a preset load balancing edge weight correction formula.

5. The multi-node signal source collaborative processing method based on the POST-INA integrated module according to claim 4, characterized in that, The obtained signal source feature matrix is ​​fused with the adjacency matrix of the updated network topology to obtain the fused feature matrix, as follows: Signal source features are extracted based on current and historical signal data. These signal source features include first-order physical features, second-order frequency domain statistical features, and third-order cross-module residual features. The signal source features are input into the adjacency matrix of the updated network topology, fused through a graph convolutional network, and a cross-layer residual connection mechanism is introduced to obtain the fused feature matrix.

6. The multi-node signal source collaborative processing method based on the POST-INA integrated module according to claim 5, characterized in that, The process involves predicting the signal state of each node using a preset first LSTM model and then initiating a neighbor node collaborative compensation mechanism based on the predicted node signal state, as detailed below: By analyzing historical time step data using the first LSTM model, the future node signal state and confidence interval are predicted. When the residual between the predicted and actual values ​​of a node signal state exceeds the first threshold, a collaborative compensation mechanism is triggered, and the signal source features are marked as abnormal.

7. The multi-node signal source collaborative processing method based on the POST-INA integrated module according to claim 6, characterized in that, The collaborative compensation mechanism includes a neighboring node signal buffering compensation mechanism, a collaborative interpolation compensation mechanism, and a redundant node activation and dynamic route reconstruction mechanism.

8. The multi-node signal source collaborative processing method based on the POST-INA integrated module according to claim 7, characterized in that, After the collaborative processing of multi-node signal sources, the method further includes: The characteristics of the marked abnormal signal sources are uploaded to the cloud server; The cloud server runs a pre-set second LSTM model to process the features of abnormal signal sources and outputs the signal status of system nodes. Compare the prediction results of the second LSTM model and the first LSTM model in the same time interval, and calculate the prediction bias between the models; When the prediction deviation exceeds the preset second threshold, the weight difference between the two models is extracted to construct an incremental update package; The weight update package is remotely sent to the local node to dynamically optimize the parameters of the first LSTM model.

Images