Underground power distribution room communication system and method based on heterogeneous network and multi-mode fusion

Abstract

The invention relates to the technical field of communication, in particular to an underground power distribution room communication system and method based on heterogeneous network and multi-modal fusion, and the system comprises a three-dimensional heterogeneous network cooperation module, a cross-modal feature fusion module, a dynamic routing decision engine module, a self-adaptive interference suppression system and a cross-layer cooperation optimization module. The three-dimensional heterogeneous network cooperation module comprises a 4G / 5G wireless communication sub-layer, a power line carrier communication sub-layer and an edge computing management sub-layer; the cross-modal feature fusion module comprises a multi-modal encoder group, a shared feature projection layer, a twin network structure and a cross-modal attention fusion module; and the dynamic routing decision engine module comprises a state space monitoring unit, a routing optimization unit and a millisecond-level path switching strategy. Through the arrangement, the reliability, the real-time performance and the energy efficiency of communication of the underground power distribution room are systematically improved, and a high-robustness communication infrastructure is provided for an intelligent power grid.

Description

Technical Field

[0001] This application relates to the field of communication technology, and in particular to a communication system and method for underground power distribution rooms based on heterogeneous networks and multimodal fusion. Background Technology

[0002] In the construction of smart grids, underground substations serve as critical nodes in the power system, and their communication reliability directly impacts power supply security. However, traditional communication solutions relying solely on 4G / 5G wireless communication are susceptible to shielding effects in the underground environment. While power line carrier (PLC) utilizes existing lines, its transmission rate is low and noise interference is severe. The lack of a coordination mechanism between the two leads to a high transmission interruption rate. Simultaneously, substations need to process multimodal data such as infrared images, SCADA parameter text, and equipment audio. Traditional single-modal analysis models cannot capture cross-modal correlation features (such as the coupling relationship between partial discharge sound and temperature abrupt changes), resulting in a high rate of missed fault detections. Furthermore, they lack adaptability to dynamic environments: at the routing level, fixed-path strategies experience a surge in delays during electromagnetic interference fluctuations, failing to meet the real-time requirement of ≤10ms for power protection signals; at the anti-interference level, conventional filtering techniques have a rejection ratio of less than 15dB for frequency conversion interference, resulting in a high bit error rate; and at the energy efficiency level, continuous high-power communication significantly reduces the battery life of terminal devices. Summary of the Invention

[0003] To address the aforementioned technical problems, this application provides a communication system and method for underground power distribution rooms based on heterogeneous networks and multimodal fusion. Through three pillars—heterogeneous network redundancy, multimodal intelligent fusion, and AI-driven dynamic optimization—the system systematically improves the reliability, real-time performance, and energy efficiency of communication in underground power distribution rooms, providing a highly robust communication infrastructure for smart grids.

[0004] In a first aspect, embodiments of this application provide a communication system for underground power distribution rooms based on heterogeneous networks and multimodal fusion, comprising:

[0005] A three-dimensional heterogeneous network collaboration module includes a 4G / 5G wireless communication sublayer, a power line carrier communication sublayer, and an edge computing management sublayer. The power line carrier communication sublayer is connected to the 4G / 5G wireless communication sublayer and the edge computing management sublayer, respectively. The 4G / 5G wireless communication sublayer, the power line carrier communication sublayer, and the edge computing management sublayer achieve collaborative control through cross-layer signaling interaction.

[0006] The cross-modal feature fusion module includes a multimodal encoder group, a shared feature projection layer, a Siamese network structure, and a cross-modal attention fusion module. The shared feature projection layer is connected to the multimodal encoder group and the Siamese network structure, respectively. The Siamese network structure is also connected to the cross-modal attention fusion module. The multimodal encoder group is also connected to the edge computing management sublayer. The cross-modal feature fusion module extracts and fuses multimodal data features based on an improved Siamese neural network.

[0007] The dynamic routing decision engine module includes a state space monitoring unit, a routing optimization unit, and a millisecond-level path switching strategy. The state space monitoring unit is connected to the routing optimization unit and the cross-modal attention fusion module, respectively. The routing optimization unit is also connected to the millisecond-level path switching strategy. The routing optimization unit uses the STAM-MADDPG model and the deep Q network model to optimize the routing strategy in real time.

[0008] An adaptive interference suppression system includes a signal subspace separation module, a cross-modal cooperative anti-interference mechanism, and an interference prediction and power control module. The interference prediction and power control module is connected to the signal subspace separation module and the cross-modal cooperative anti-interference mechanism, respectively. The cross-modal cooperative anti-interference mechanism is also connected to the signal subspace separation module and the millisecond-level path switching strategy, respectively. The signal subspace separation module, the cross-modal cooperative anti-interference mechanism, and the interference prediction and power control module combine NOMA-STBC joint anti-interference technology, generative adversarial network interference prediction, and the cross-modal cooperative mechanism to suppress complex electromagnetic interference.

[0009] A cross-layer collaborative optimization module is connected to the edge computing management sub-layer and the state space monitoring unit, respectively. The cross-layer collaborative optimization module balances transmission rate, bit error rate and energy consumption based on non-dominated sorting genetic algorithm III.

[0010] According to some embodiments of the first aspect of this application, the 4G / 5G wireless communication sublayer adopts dynamic network slicing technology to allocate a dedicated 5MHz bandwidth for power protection signals so that the transmission bandwidth is ≥200Mbps; the power line carrier communication sublayer optimizes OFDM modulation based on the G3-PLC standard to achieve a rate of ≥1Mbps in the 6-500kHz frequency band; the edge computing management sublayer introduces lightweight model compression technology and implements hardware-accelerated inference of the model based on the OpenVINO framework. Combined with the task offloading strategy, the tasks of power equipment status monitoring and fault early warning with real-time requirements of ≤5ms are completed on the edge side. Moreover, the edge computing management sublayer automatically switches to the power line communication backup channel when the wireless link is interrupted through a cross-layer signaling interaction protocol.

[0011] The 4G / 5G wireless communication sublayer is used to transmit the collected power data and control command information to the edge computing management sublayer through a dedicated network slice. The power line carrier communication sublayer is used to utilize the existing power lines in the power distribution room to provide a stable data transmission channel for near-field devices, transmit the data to the edge computing management sublayer, and serve as a backup communication path when the 4G / 5G wireless communication link is interrupted, receiving and forwarding key commands from the edge computing management sublayer. The edge computing management sublayer is used to perform localized processing and intelligent decision-making on the data from the 4G / 5G wireless communication sublayer and the power line carrier communication sublayer respectively to generate control signals, and coordinate the resource allocation of the 4G / 5G wireless communication sublayer and the channel scheduling of the power line carrier communication sublayer through cross-layer signaling interaction based on the control signals.

[0012] According to some embodiments of the first aspect of this application, the multimodal encoder group includes a power equipment visual encoder, a power parameter text encoder, and an environmental audio encoder. The power equipment visual encoder, the power parameter text encoder, and the environmental audio encoder are used to process image data, text data, and audio data, respectively. The shared feature projection layer constructs a projection matrix by maximizing cross-modal mutual information. The twin network structure embeds a power equipment topology graph. The cross-modal attention fusion module dynamically allocates weights based on a three-dimensional attention mechanism.

[0013] According to some embodiments of the first aspect of this application, the state space monitoring unit collects signal strength, signal-to-noise ratio, and device power parameters in real time through multi-source heterogeneous data fusion sensing technology; the routing optimization unit generates routing strategies based on the STAM-MADDPG model; the routing optimization unit abstracts communication nodes as intelligent agents, realizes cross-modal link collaborative scheduling based on the STAM-MADDPG model, and adopts a deep Q-network model containing a three-layer fully connected neural network to minimize latency and packet loss rate, and accelerates routing strategy optimization with the help of an experience replay mechanism; the millisecond-level path switching strategy triggers switching within 15ms through dual thresholds.

[0014] According to some embodiments of the first aspect of this application, the signal subspace separation module separates 4G / 5G and power line carrier signals based on an improved singular value decomposition algorithm. The signal subspace separation module also constructs a dynamic space-time block coded (STBC) matrix to reassemble the signal, so as to maintain a bit error rate of less than or equal to 10 even under a -5dB signal-to-noise ratio environment. -3The interference prediction and power control module generates adversarial network interference predictions and dynamically adjusts the transmit power. The cross-modal cooperative anti-interference mechanism is used to trigger NOMA technology to reuse frequency band resources or switch transmission links when the performance of any link in 4G / 5G and power line carrier is lower than the threshold, extending the regulation strategy generated by the interference prediction and power control module to multi-modal cooperative scenarios.

[0015] Secondly, embodiments of this application provide a communication method for underground power distribution rooms based on heterogeneous networks and multimodal intelligent fusion, applied to an underground power distribution room communication system based on heterogeneous networks and multimodal fusion. The underground power distribution room communication system based on heterogeneous networks and multimodal fusion includes a three-dimensional heterogeneous network collaboration module, a cross-modal feature fusion module, a dynamic routing decision engine module, an adaptive interference suppression system, and a cross-layer collaborative optimization module. The three-dimensional heterogeneous network collaboration module includes a 4G / 5G wireless communication sublayer, a power line carrier communication sublayer, and an edge computing management sublayer. The power line carrier communication sublayer is connected to the 4G / 5G wireless communication sublayer, the power line carrier communication sublayer, and the edge computing management sublayer. The edge computing management sublayer is connected, and the 4G / 5G wireless communication sublayer, the power line carrier communication sublayer, and the edge computing management sublayer achieve coordinated control through cross-layer signaling interaction. The cross-modal feature fusion module includes a multimodal encoder group, a shared feature projection layer, a Siamese network structure, and a cross-modal attention fusion module. The shared feature projection layer is connected to the multimodal encoder group and the Siamese network structure, respectively. The Siamese network structure is also connected to the cross-modal attention fusion module. The multimodal encoder group is also connected to the edge computing management sublayer. The cross-modal feature fusion module extracts and fuses multimodal features based on an improved Siamese neural network. Modal data features; the dynamic routing decision engine module includes a state space monitoring unit, a routing optimization unit, and a millisecond-level path switching strategy. The state space monitoring unit is connected to the routing optimization unit and the cross-modal attention fusion module, respectively. The routing optimization unit is also connected to the millisecond-level path switching strategy. The routing optimization unit uses a STAM-MADDPG model and a deep Q-network model to optimize the routing strategy in real time. The adaptive interference suppression system includes a signal subspace separation module, a cross-modal collaborative anti-interference mechanism, and an interference prediction and power control module. The interference prediction and power control module is separated from the signal subspace. The module is connected to the cross-modal collaborative anti-interference mechanism, which is also connected to the signal subspace separation module and the millisecond-level path switching strategy. The signal subspace separation module, the cross-modal collaborative anti-interference mechanism, and the interference prediction and power control module combine NOMA-STBC joint anti-interference technology, generative adversarial network interference prediction, and cross-modal collaborative mechanism to suppress complex electromagnetic interference. The cross-layer collaborative optimization module is connected to the edge computing management sublayer and the state space monitoring unit. The cross-layer collaborative optimization module balances transmission rate, bit error rate, and energy consumption based on non-dominated sorting genetic algorithm III.

[0016] The method includes:

[0017] Cross-layer data preprocessing:

[0018] In the 4G / 5G wireless communication sublayer, spatial coordinate mapping of 4G / 5G baseband signals is performed through BeiDou / GNSS dual-system positioning.

[0019] In the power line carrier communication sublayer, the power line carrier signal is clock-synchronized to generate a unified data frame containing timestamps and device IDs;

[0020] In the edge computing management sublayer, two types of signal data are integrated and encapsulated into standardized data frames according to dynamic slicing requirements and channel detection requirements for subsequent processing;

[0021] Feature fusion coding:

[0022] In the multimodal encoder group, infrared images are processed by an improved ResNet50+CBAM architecture to compress brightness to the [0.2, 0.8] range; SCADA parameters are processed by dual-channel LSTM+time window attention; and abnormal noise features are extracted by Mel spectrum-wavelet transform.

[0023] In the shared feature projection layer, features are mapped through visual, text, and audio projection matrices, and weights are dynamically adjusted according to the environment.

[0024] In the twin network structure, multi-scale features are captured through dilated convolution, embedded in the device topology graph, and comparative learning is performed based on the improved InfoNCE loss function.

[0025] In the cross-modal attention fusion module, spatial attention is used to focus on key parts of the device, temporal attention is used to capture data 5 seconds before the fault occurs, modal attention is used to dynamically allocate weights, and fused features are output.

[0026] Dynamic route calculation:

[0027] The state space monitoring unit collects signal strength, signal-to-noise ratio, and device power in real time, and constructs a state vector by combining the space attenuation factor, time penalty, and interference threshold.

[0028] In the routing optimization unit, the STAM-MADDPG model is used to coordinate the scheduling of 4G / 5G and power line carrier links, and the deep Q network model generates routing strategies with the goal of minimizing latency and packet loss rate.

[0029] In millisecond-level path switching strategies, when the latency is ≥30ms or the packet loss rate is ≥10... -4 When the pre-switching buffer is triggered, the path reselection is completed within 15ms. Real-time commands are transmitted via 4G / 5G, while historical data is transmitted via power line carrier.

[0030] Interference adaptive processing:

[0031] In the signal subspace separation module, 4G / 5G and power line carrier signals are separated according to the difference in singular values ​​based on the improved singular value decomposition algorithm, and a dynamic STBC matrix is ​​constructed to resist multipath fading.

[0032] In the interference prediction and power control module, the interference distribution is predicted by a generative adversarial network model, the transmit power is adjusted in 2dBm steps, and the LoRa spreading factor is dynamically switched according to the interference intensity.

[0033] In the cross-modal cooperative anti-interference mechanism, when the bit error rate of a certain link is >10 -4 If the signal-to-noise ratio is less than the threshold, NOMA technology is activated to reuse frequency band resources, or the connection is switched to a backup link.

[0034] Cross-layer collaborative optimization:

[0035] Design a weighted optimization function to dynamically adjust the weights; allocate the proportion of 4G / 5G and power line carrier transmission based on non-dominated sorting genetic algorithm III; when the 4G / 5G load is ≥70%, divert 30% of non-real-time services to power line carrier and improve spectrum efficiency through NOMA technology.

[0036] For device-level power consumption control, the transmission power is reduced by 50% when the battery level is below 20%, and the sleep cycle is extended to 500ms;

[0037] For network-level power consumption control, when the server CPU utilization is greater than or equal to 80%, the server CPU utilization is reduced by 0.3GHz for every 10% increase, and tasks are migrated to low-load servers.

[0038] The indicators are guaranteed through a constraint processing mechanism, and the solution set is updated using an elite retention strategy guided by reference points.

[0039] According to some embodiments of the second aspect of this application, the calculation formula for the attention weight in the modal attention dynamic allocation weight is as follows:

[0040]

[0041] Among them, e t =score(h t ,q) represents the attention score, h t Let α represent the hidden state of the LSTM at time t, q represent the query vector, and α represent the hidden state of the LSTM at time t. t The attention weight at time t is used to weight and sum the data at different times to obtain the feature representation; β t This represents the spatiotemporal coupling coefficient, reflecting the differences in the importance of data at different times, and is dynamically adjusted based on the dynamic changing trends of the power system; w mRepresents modal interaction weights, measuring the degree of correlation between multimodal data, and is dynamically learned through a cross-modal attention mechanism; ρ t,m This represents the equipment topology constraint terms, introduces a power equipment topology relationship graph, and quantifies the correlation between the current data and the fault propagation path;

[0042] This represents modal feature enhancement terms, which enhance the characteristics of different modal data.

[0043]

[0044] Wherein, s(x i ,x j ) represents sample x i and x j The similarity score; τ represents the temperature parameter, used to control the difficulty of contrastive learning, by minimizing the InfoNCE loss function, bringing positive sample pairs closer and pushing negative sample pairs further apart; φ i,i+ The weights represent the topological constraints of the devices. Based on the power equipment topology graph, the topological correlation between device nodes in a positive sample pair is calculated using a graph neural network. i,j ε represents the negative sample association penalty term, which incorporates prior knowledge in the power field to introduce a dynamic penalty mechanism for negative sample pairs; ε represents the smoothing factor to ensure the stability of the loss function calculation.

[0045] According to some embodiments of the second aspect of this application, after the state space monitoring unit collects signal strength, signal-to-noise ratio, and device power in real time, and constructs a state vector by combining spatial attenuation factor, time penalty, and interference threshold, the process includes:

[0046] The attention weight α is dynamically adjusted by a spatiotemporal attention mechanism, incorporating multimodal information fusion, dynamic environment perception, and network topology constraints. i The calculation formula is as follows:

[0047]

[0048] Where, f(s) i ) represents the attention scoring function, which obtains a score by calculating the correlation between each parameter and the current routing decision; α i Used for weighted aggregation of state vectors, prioritizing the capture of key parameters affecting routing decisions; γ t,i b represents the time-based dynamic adjustment coefficient, which dynamically adjusts the attention weights of parameters at different times based on historical latency fluctuations in the communication link and the urgency of real-time transmission tasks; β m,iThe modality fusion weights represent the multimodal features of 4G / 5G and power line carrier communication, dynamically allocating weights based on link type and interference status; g(G) represents the network topology constraint term, introducing the communication network topology graph G, and calculating the structural relationships between nodes through a graph convolutional network; δ n,i This represents the topology influence coefficient, which dynamically adjusts the influence of topology constraints based on the role of a node in the network topology.

[0049] According to some embodiments of the second aspect of this application, the method of coordinating the scheduling of 4G / 5G and power line carrier links through the STAM-MADDPG model, and generating routing strategies with the deep Q network model aiming to minimize latency and packet loss rate, includes:

[0050] The Deep Q-Network model learns the optimal routing strategy through an improved Deep Q-Network loss function, balancing immediate rewards and long-term benefits. The formula for calculating the improved Deep Q-Network loss function is as follows:

[0051] L = E s,a,r,s′ [w delay *(y delay -Q delay (s,a;θ)) 2 +w packet *(y packet -Q packet (s,a;θ)) 2 ]

[0052] Among them, y delay =r delay +γ delay max a' Q delay (S',a';θ - ),y packet =r packet +γ packet max a' Q packet (S',a′;θ - ); r represents the immediate reward, a positive reward for successful transmission and a negative reward for packet loss; γ represents the discount factor, ranging from 0 to 1, balancing immediate and long-term benefits; θ represents the current network parameters; θ - The target network parameters are represented and updated periodically to stabilize training. The (s,a,r,s′) are stored in the experience pool through the experience replay mechanism and randomly sampled for training.

[0053] According to some embodiments of the second aspect of this application, the interference adaptive processing further includes:

[0054] By leveraging generative adversarial networks to fuse historical interference data with real-time environmental parameters, the generator and discriminator learn adversarially to predict interference distribution. The generator attempts to produce realistic interference samples. The discriminator distinguishes between real interference and generated samples. The objective function of the generative adversarial network model is as follows:

[0055]

[0056] Where, max D This indicates that the discriminator D attempts to maximize its own discriminative power, min G Let V(D,G) represent the generator G's attempt to minimize the discriminative power of the discriminator, and let V(D,G) represent the optimization objective function of the generative adversarial network. Let log D(I) represent the expectation of the real disturbance data, where log D(I) follows the real data distribution P. data ; log D(i) represents the logarithmic probability of the discriminator D classifying the true sample I. This represents the noise vector log(1-D(G(P)). real The expectation of log(1-D(G(P)) real )) Follows the prior distribution P noise ;log(1-D(G(P) real )) indicates that the discriminator D evaluates the generated sample G(P) real The logarithm of the probability complement of the judgment, P real This represents real-time environmental parameters.

[0057] The beneficial effects of this application are reflected in the fact that the three-dimensional heterogeneous network collaborative module complements the 4G / 5G wireless communication sublayer (wide-area coverage) and the power line carrier communication sublayer (utilizing existing lines), avoiding single-point failures and monitoring the link status in real time (such as when PLC noise suddenly increases). It dynamically switches to the 4G / 5G link through the edge computing management sublayer, reducing the interruption rate; edge computing local decision-making: it can quickly respond to link failures locally in the power distribution room, avoiding cloud decision-making delays;

[0058] The cross-modal feature fusion module processes infrared images (visual), SCADA parameters (text / numerical), and device audio (time-series signals) separately through a multimodal encoder group to extract modality-specific features. It maps heterogeneous features to a unified space through a shared feature projection layer. The cross-modal attention module dynamically weights different modal features (e.g., when discharge audio occurs, the temperature feature weight of the corresponding region is enhanced) to capture coupling relationships. It directly outputs the fused joint feature vector for use by the fault diagnosis model to reduce the false negative rate.

[0059] The dynamic routing decision engine module collects network congestion, interference intensity, and link quality in real time through the state space monitoring unit (from the cross-modal feature fusion module); the STAM-MADDPG model combines the spatiotemporal attention mechanism (STAM) and multi-agent reinforcement learning (MADDPG) to predict interference change trends and generate the optimal route; the millisecond-level switching strategy prioritizes switching to low-latency paths when critical signals (such as protection instructions) are detected, ensuring transmission within ≤10ms.

[0060] The adaptive interference suppression system improves spectral efficiency through the joint NOMA-STBC technology (Non-Orthogonal Multiple Access (NOMA) and enhances anti-fading capability through the Space-Time Block Code (STBC); Generative Adversarial Network (GAN) interference prediction: GAN learns historical interference patterns and predicts future interference distribution; Cross-modal collaborative anti-interference: For example, after the audio sensor identifies interference at a specific frequency, it notifies the communication module to avoid that frequency band to improve the suppression ratio and reduce the bit error rate.

[0061] The non-dominated sorting genetic algorithm III in the cross-layer collaborative optimization module solves the Pareto optimal solution set in real time at the edge computing layer with multiple objectives such as transmission rate, bit error rate, and energy consumption; and dynamically adjusts the transmission power according to the interference prediction results (from the anti-interference module) and service priority; and selects the path with the lowest energy consumption to transmit non-urgent data, thereby extending the terminal's battery life.

[0062] In other words, the modules in this application form a closed loop through data flow (e.g., the anti-interference module outputs interference characteristics → the routing module adjusts the path → the cross-layer optimization module balances energy consumption) to achieve global optimization; the edge computing management sublayer acts as the core hub, coordinating network control, modal fusion, and routing decisions, reducing cloud dependence and improving response speed. This application systematically improves the reliability, real-time performance, and energy efficiency of underground power distribution room communication through three pillars: heterogeneous network redundancy, multimodal intelligent fusion, and AI-driven dynamic optimization, providing a highly robust communication infrastructure for the smart grid. Attached Figure Description

[0063] Figure 1 A connection diagram of an underground power distribution room communication system based on heterogeneous network and multimodal fusion provided for the first aspect of this application;

[0064] Figure 2 This is a flowchart illustrating the communication method for underground power distribution rooms based on the fusion of heterogeneous networks and multimodal intelligence provided in the second aspect of this application.

[0065] Figure 3 This is a schematic diagram of the feature fusion coding process provided in the second aspect embodiment of this application;

[0066] Figure 4This is a schematic diagram of the dynamic route calculation process provided in the second aspect embodiment of this application;

[0067] Figure 5 This is a schematic flowchart of the interference adaptive processing provided in the second aspect of this application. Detailed Implementation

[0068] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0069] The following description, in conjunction with the accompanying drawings, details a communication system and method for underground power distribution rooms based on heterogeneous networks and multimodal fusion, provided by the embodiments of this application, through specific implementations and application scenarios.

[0070] To address the aforementioned problems, the first aspect of this application proposes a communication system for underground power distribution rooms based on heterogeneous networks and multimodal fusion. The embodiments of this application will be further described below with reference to the accompanying drawings.

[0071] Reference Figure 1The first aspect of this application provides a communication system for underground power distribution rooms based on heterogeneous networks and multimodal fusion, including a three-dimensional heterogeneous network collaboration module, a cross-modal feature fusion module, a dynamic routing decision engine module, an adaptive interference suppression system, and a cross-layer collaborative optimization module. The three-dimensional heterogeneous network collaboration module includes a 4G / 5G wireless communication sublayer, a power line carrier communication sublayer, and an edge computing management sublayer. The power line carrier communication sublayer is connected to the 4G / 5G wireless communication sublayer and the edge computing management sublayer, respectively. The sub-layer and edge computing management sub-layer achieve collaborative control through cross-layer signaling interaction; the cross-modal feature fusion module includes a multimodal encoder group, a shared feature projection layer, a Siamese network structure, and a cross-modal attention fusion module. The shared feature projection layer is connected to the multimodal encoder group and the Siamese network structure, respectively. The Siamese network structure is also connected to the cross-modal attention fusion module. The multimodal encoder group is also connected to the edge computing management sub-layer. The cross-modal feature fusion module extracts and fuses multimodal data features based on an improved Siamese neural network; the dynamic routing decision engine module includes a state space. The system comprises a monitoring unit, a routing optimization unit, and a millisecond-level path switching strategy. The state space monitoring unit is connected to the routing optimization unit and the cross-modal attention fusion module, respectively. The routing optimization unit is also connected to the millisecond-level path switching strategy. The routing optimization unit uses the STAM-MADDPG model and a deep Q-network model to optimize the routing strategy in real time. An adaptive interference suppression system includes a signal subspace separation module, a cross-modal collaborative anti-interference mechanism, and an interference prediction and power control module. The interference prediction and power control module is connected to the signal subspace separation module and the cross-modal collaborative anti-interference mechanism, respectively. The cross-modal collaborative anti-interference mechanism is also connected to the signal subspace separation module and the millisecond-level path switching strategy, respectively. The signal subspace separation module, the cross-modal collaborative anti-interference mechanism, and the interference prediction and power control module combine NOMA-STBC joint anti-interference technology, generative adversarial network interference prediction, and the cross-modal collaborative mechanism to suppress complex electromagnetic interference. A cross-layer collaborative optimization module is connected to the edge computing management sublayer and the state space monitoring unit, respectively. The cross-layer collaborative optimization module balances transmission rate, bit error rate, and energy consumption based on a non-dominated sorting genetic algorithm III.

[0072] Specifically, the three-dimensional heterogeneous network collaboration module adopts a hierarchical collaboration and dynamic resource allocation architecture to construct a three-dimensional fusion system consisting of a 4G / 5G wireless communication sublayer, a power line carrier communication sublayer, and an edge computing management sublayer; the cross-modal feature fusion module designs a multimodal power data feature extractor based on an improved Siamese neural network (Siamese-NN) model, which includes a multimodal encoder group, a shared feature projection layer, a Siamese network structure, and a cross-modal attention fusion module; through the cross-modal feature fusion module, feature extraction and fusion of image, text, and audio modalities are achieved, making it suitable for various cross-modal tasks; the dynamic routing decision engine module proposes a spatiotemporal attention-based approach. The mechanism employs a multi-agent deep deterministic policy gradient model (STAM-MADDPG) to construct a state-space monitoring unit, a routing optimization unit, and a millisecond-level path switching strategy. The state-space monitoring unit includes a spatial attenuation factor of 0.75 / km, a time delay penalty of 0.1 points per 1ms delay, and an interference strength threshold of -85dBm, monitoring parameters such as 4G / 5G signal strength, power line carrier noise level, and remaining device power in real time. The routing optimization unit generates the optimal routing strategy through a deep Q-network, achieving millisecond-level path switching with a handover latency ≤15ms. In an underground power distribution room environment, it achieves an end-to-end transmission latency ≤30ms and a packet loss rate ≤10%. -4 The millisecond-level path switching strategy achieves efficient transmission through dual-threshold triggering and multi-mode link complementarity; the adaptive interference suppression system adopts power domain NOMA-STBC joint anti-interference, including a signal subspace separation module and an interference prediction and power control module; the signal subspace separation module is based on the Singular Value Decomposition (SVD) algorithm to separate 4G / 5G signals with singular values ​​≥ 3 times the power line carrier, and constructs a space-time block coding (STBC) matrix to enhance the signal's resistance to multipath fading; the interference prediction and power control module is based on generative adversarial networks (GANs) to predict electromagnetic interference distribution, dynamically adjust the transmit power with a 2dBm transmit step size and the LoRa spreading factor to improve link reliability; the cross-layer collaborative optimization module constructs a transmission rate system that includes targets with a transmission rate ≥ 100Mbps and targets with a transmission rate ≤ 100Mbps. -6 The multi-objective optimization function aims to reduce bit error rate and system energy consumption by more than 30%. Based on the improved non-dominated sorting genetic algorithm NSGA-Ⅲ, it dynamically balances the transmission ratio of 4G / 5G and power line carrier. When the remaining power threshold of the device is ≤20%, a low-power mode is triggered. When the CPU utilization of the edge server is ≥80%, dynamic frequency adjustment is started to achieve energy efficiency optimization. The frequency adjustment range is 1.2-2.4GHz.

[0073] It should be noted that the connection between the 4G / 5G wireless communication sublayer and the edge computing management sublayer serves the following purpose: to transmit wide-area high-speed data (such as device status monitoring streams), ensuring bandwidth ≥ 5MHz through a dedicated 5MHz slice.

[0074] 200Mbps, latency ≤10ms; dynamically allocates resources based on the Slicing-Enabled architecture, and AI prediction algorithms prioritize power protection signals. Power line carrier communication sublayer and edge computing management sublayer connection: provides a stable local transmission channel (rate ≥1Mbps), automatically switching to a backup path when 4G / 5G is interrupted;

[0075] G3-PLC optimizes OFDM modulation and adaptively adjusts subcarrier allocation in real time based on spectrum sensing. The edge computing management sublayer connects with the 4G / 5G wireless communication sublayer and the power line carrier communication sublayer. Its function is to provide feedback control signals and coordinate resource allocation and channel scheduling. Technical implementation involves triggering PLC backup channel switching via cross-layer signaling protocols, dynamically offloading real-time tasks ≤5ms (such as accelerating inference using YOLOv5 compressed models).

[0076] It should be noted that the multimodal encoder group is connected to the shared feature projection layer: its function is to map the raw data to a unified feature space; the technical implementation is as follows: visual encoder: ResNet50+CBAM compresses the brightness to the range of [0.2, 0.8] (temperature error ≤ ±0.5℃); text encoder: dual-channel LSTM+time window attention (data weight of 30 minutes before the fault + 20%); audio encoder: Mel spectrum-wavelet transform extracts abnormal noise (recognition rate ≥90% at -5dB signal-to-noise ratio). The role of connecting the shared feature projection layer with the Siamese network structure: to preserve and dynamically weight power semantic features; technical implementation: projection matrix optimization (visual 2048×128 / text 1024×128 / audio 1024×128), dynamic weight adjustment (normal state: text weight 60%; fault state: visual weight 70%); the role of connecting the Siamese network structure with the cross-modal attention fusion module: to enhance feature discriminative power and spatiotemporal correlation; technical implementation: dilated convolution to capture multi-scale features, inject equipment topology relationships (such as circuit breaker-transformer connection), and improve the InfoNCE loss function to introduce prior power knowledge.

[0077] The cross-modal attention fusion output function is to generate a high-dimensional fusion feature vector, a three-dimensional attention mechanism (spatial focus on the terminal / temporal focus on the 5 seconds before tripping / audio weight of 50% during modal discharge fault), and GNN (Generative Adversarial Network) to correct the fault propagation probability.

[0078] It should be noted that the state space monitoring unit is connected to the route optimization unit: its function is to provide real-time decision-making basis; the technical implementation involves quantizing the state vector s = [signal strength, signal-to-noise ratio, device power], introducing a spatial attenuation factor (0.75 / km) and a time penalty (0.1 points deducted for every 1ms delay), and dynamically weighting key parameters using a spatiotemporal attention mechanism. The route optimization unit is connected to the millisecond-level path switching strategy: its function is to generate the optimal routing strategy and execute the switching; the technical implementation involves using the STAM-MADDPG model to coordinate the actions of multiple agents, the DQN network (deep Q-network) to output the Q-value (a three-layer fully connected neural network), and an experience replay mechanism to avoid overfitting. The state space monitoring unit is also connected to the millisecond-level path switching strategy: its function is to trigger a fast response with dual thresholds; the technical implementation requires a delay ≥30ms or a packet loss rate ≥10%. -4 The handover is initiated directly, and the pre-handover buffer strategy ensures that the path migration is completed within 15ms.

[0079] It should be noted that the signal subspace separation module is connected to the interference prediction and power control module. Its function is to provide the purified signal input. Technical implementation includes: SVD algorithm separation of 4G / 5G signals (singularity ≥ 3 times the threshold), and dynamic STBC matrix for multipath fading resistance (bit error rate ≤ 10 at -5dB SNR). -3 The interference prediction and power control module is connected to the cross-modal collaborative anti-interference mechanism: Its function is to transmit single-link control strategies; its implementation involves using GAN to predict interference distribution, adjusting transmit power in 2dBm steps, and dynamically switching LoRa spreading factors. The signal subspace separation module is also connected to the cross-modal collaborative anti-interference mechanism: Its function is cross-modal resource reuse; its implementation involves triggering NOMA technology to allocate subcarrier power at a 1:3 ratio when single-link performance is below a threshold, reusing spectrum resources of another mode (co-channel interference suppression ratio ≥25dB).

[0080] This system optimizes communication transmission in underground power distribution rooms from multiple aspects, specifically addressing existing technical deficiencies to achieve efficient, stable, and intelligent data transmission. Specific implementation examples are as follows: I. Solving the problem of unstable communication links: The adaptive interference suppression system adopts a power domain NOMA-STBC joint anti-interference scheme. The signal subspace separation module, based on the Singular Value Decomposition (SVD) algorithm, analyzes and processes 4G / 5G signals and power line carrier signals, accurately separating 4G / 5G signals with singular values ​​≥ 3 times that of the power line carrier, and then constructs a space-time block coding (STBC) matrix. When the system detects multipath fading interference caused by the complex electromagnetic environment of the underground power distribution room, the STBC matrix enhances the signal's anti-interference capability through spatial and temporal diversity techniques. The interference prediction and power control module utilizes generative adversarial networks (GANs) to generate adversarial networks and predict electromagnetic interference distribution in real time. If strong interference is predicted in a certain area, the system immediately and dynamically adjusts the transmit power with a 2dBm step size and the LoRa spreading factor to effectively avoid interference, ensure stable link transmission, and significantly improve data transmission reliability. II. Solving the Difficulty of Multimodal Data Processing: The cross-modal feature fusion module is designed based on an improved Siamese-NN model to extract multimodal power data features. When the system receives multimodal data such as images, text, and audio from an underground power distribution room, the multimodal encoder group performs preliminary encoding on different modalities of data. Then, the data is mapped to a unified feature space through a shared feature projection layer. The Siamese network structure further mines data features, and the cross-modal attention fusion module assigns attention weights according to the importance of different modalities of data, achieving deep fusion of multimodal data. Through this process, the system can efficiently process complex and diverse power data, providing accurate data support for applications such as equipment status monitoring and fault early warning, and meeting the needs of various cross-modal tasks. III. Solving the Problem of Inefficient Routing Decisions: The dynamic routing decision engine module works based on the STAM-MADDPG multi-agent deep deterministic policy gradient model with a spatiotemporal attention mechanism. The state space monitoring unit monitors parameters such as 4G / 5G signal strength, power line carrier noise level, and remaining device power in real time, using a space attenuation factor of 0.75 / km, a time delay penalty of 0.1 points per 1ms delay, and an interference strength threshold of -85dBm. If the 4G / 5G signal strength drops below -85dBm, or the power line carrier noise level is too high and affects transmission quality, the routing optimization unit immediately activates the deep Q network to comprehensively analyze the current network status and generate the optimal routing strategy. The millisecond-level path switching strategy, through dual-threshold triggering and multi-mode link complementarity, rapidly switches the transmission path when the transmission delay or packet loss rate exceeds the threshold, achieving a switching delay ≤15ms. In underground power distribution room environments, it ensures end-to-end transmission delay ≤30ms and packet loss rate ≤10%. -4This greatly improves routing decision efficiency and data transmission timeliness. IV. Solution to the problem of unreasonable communication resource allocation: The three-dimensional heterogeneous network collaborative module innovatively adopts a layered collaborative and dynamic resource allocation architecture. When the system is running, the 4G / 5G wireless communication sublayer, the power line carrier communication sublayer, and the edge computing management sublayer constitute a three-dimensional fusion system. During data transmission, resources are dynamically allocated according to different scenario requirements. For example, when transmitting equipment status monitoring data with high real-time requirements, the high-speed transmission advantage of the 4G / 5G wireless communication sublayer is prioritized; for conventional data with low bandwidth requirements but requiring stable transmission, it is allocated to the power line carrier communication sublayer. The edge computing management sublayer preprocesses and analyzes the data, reducing data backhaul pressure, improving overall communication efficiency, and achieving efficient utilization and reasonable allocation of communication resources. V. Solution to the problem of imbalance between transmission performance and energy consumption: The cross-layer collaborative optimization module constructs a multi-objective optimization function, setting a transmission rate of ≥100Mbps and a target of ≤10 -6 The optimization goals are to reduce bit error rate and system energy consumption by more than 30%. Based on the improved non-dominated sorting genetic algorithm NSGA-Ⅲ, the system dynamically balances the transmission ratio of 4G / 5G and power line carrier. When the remaining power threshold of the device is ≤20%, a low-power mode is automatically triggered to reduce system energy consumption; when the edge server CPU utilization is ≥80%, dynamic frequency adjustment is initiated, adjusting the frequency within the range of 1.2-2.4GHz. While ensuring transmission performance, energy efficiency is optimized, effectively resolving the contradiction between transmission performance and energy consumption, and improving the overall efficiency of the system.

[0081] The beneficial effects of this application are reflected in the fact that the three-dimensional heterogeneous network collaborative module complements the 4G / 5G wireless communication sublayer (wide-area coverage) and the power line carrier communication sublayer (utilizing existing lines), avoiding single-point failures and monitoring the link status in real time (such as when PLC noise suddenly increases). It dynamically switches to the 4G / 5G link through the edge computing management sublayer, reducing the interruption rate; edge computing local decision-making: it can quickly respond to link failures locally in the power distribution room, avoiding cloud decision-making delays;

[0082] The cross-modal feature fusion module processes infrared images (visual), SCADA parameters (text / numerical), and device audio (time-series signals) separately through a multimodal encoder group to extract modality-specific features. It maps heterogeneous features to a unified space through a shared feature projection layer. The cross-modal attention module dynamically weights different modal features (e.g., when discharge audio occurs, the temperature feature weight of the corresponding region is enhanced) to capture coupling relationships. It directly outputs the fused joint feature vector for use by the fault diagnosis model to reduce the false negative rate.

[0083] The dynamic routing decision engine module collects network congestion, interference intensity, and link quality in real time through the state space monitoring unit (from the cross-modal feature fusion module); the STAM-MADDPG model combines the spatiotemporal attention mechanism (STAM) and multi-agent reinforcement learning (MADDPG) to predict interference change trends and generate the optimal route; the millisecond-level switching strategy prioritizes switching to low-latency paths when critical signals (such as protection instructions) are detected, ensuring transmission within ≤10ms.

[0084] The adaptive interference suppression system improves spectral efficiency through the joint NOMA-STBC technology (Non-Orthogonal Multiple Access (NOMA) and enhances anti-fading capability through the Space-Time Block Code (STBC); Generative Adversarial Network (GAN) interference prediction: GAN learns historical interference patterns and predicts future interference distribution; Cross-modal collaborative anti-interference: For example, after the audio sensor identifies interference at a specific frequency, it notifies the communication module to avoid that frequency band to improve the suppression ratio and reduce the bit error rate.

[0085] The non-dominated sorting genetic algorithm III in the cross-layer collaborative optimization module solves the Pareto optimal solution set in real time at the edge computing layer with multiple objectives such as transmission rate, bit error rate, and energy consumption; and dynamically adjusts the transmission power according to the interference prediction results (from the anti-interference module) and service priority; and selects the path with the lowest energy consumption to transmit non-urgent data, thereby extending the terminal's battery life.

[0086] In other words, the modules in this application form a closed loop through data flow (e.g., the anti-interference module outputs interference characteristics → the routing module adjusts the path → the cross-layer optimization module balances energy consumption) to achieve global optimization; the edge computing management sublayer acts as the core hub, coordinating network control, modal fusion, and routing decisions, reducing cloud dependence and improving response speed. This application systematically improves the reliability, real-time performance, and energy efficiency of underground power distribution room communication through three pillars: heterogeneous network redundancy, multimodal intelligent fusion, and AI-driven dynamic optimization, providing a highly robust communication infrastructure for the smart grid.

[0087] Understandably, the 4G / 5G wireless communication sublayer employs dynamic network slicing technology, allocating a dedicated 5MHz bandwidth for power protection signals to achieve a transmission bandwidth ≥200Mbps; the power line carrier communication sublayer optimizes OFDM modulation based on the G3-PLC standard, achieving a rate ≥1Mbps in the 6–500kHz frequency band; the edge computing management sublayer introduces lightweight model compression technology and implements hardware-accelerated model inference based on the OpenVINO framework. Combined with a task offloading strategy, tasks requiring real-time performance of ≤5ms for power equipment status monitoring and fault early warning are completed at the edge. Furthermore, the edge computing management sublayer automatically switches to the power line communication backup channel when the wireless link is interrupted through a cross-layer signaling interaction protocol; 4G / 5G wireless... The communication sublayer transmits collected power data and control commands to the edge computing management sublayer via dedicated network slices. The power line carrier communication sublayer utilizes existing power lines in the distribution room to provide a stable data transmission channel for nearby devices and transmits the data to the edge computing management sublayer. It also serves as a backup communication path when the 4G / 5G wireless communication link is interrupted, receiving and forwarding critical commands from the edge computing management sublayer. The edge computing management sublayer performs localized processing and intelligent decision-making on data from both the 4G / 5G wireless communication sublayer and the power line carrier communication sublayer to generate control signals. Based on these control signals, it coordinates resource allocation in the 4G / 5G wireless communication sublayer and channel scheduling in the power line carrier communication sublayer through cross-layer signaling interaction.

[0088] Specifically, the three-dimensional heterogeneous network collaboration module includes a 4G / 5G wireless communication sublayer, a power line carrier communication sublayer, and an edge computing management sublayer. Each sublayer achieves collaborative control through cross-layer signaling interaction. The 4G / 5G wireless communication sublayer implements dynamic network slicing based on a Slicing-Enabled architecture. Through an AI-driven resource prediction algorithm, it pre-allocates a dedicated 5MHz slice for power protection signals. Combined with the dynamic priority scheduling mechanism of URLLC and eMBB, it can still guarantee a transmission bandwidth ≥200Mbps and an end-to-end latency ≤10ms even in complex electromagnetic environments. The power line carrier communication sublayer breaks through the traditional PLC speed bottleneck, based on… The OFDM modulation scheme is optimized based on the G3-PLC standard. Adaptive spectrum sensing technology is used in the 6-500kHz frequency band to detect the power line channel status in real time and dynamically adjust subcarrier allocation. Anti-interference subchannels are constructed through interference cancellation algorithms to achieve a stable data transmission rate of ≥1Mbps. Lightweight model compression technology is introduced into the edge computing management sublayer, compressing the YOLOv5 power equipment anomaly detection model by 80%. Hardware-accelerated inference of the model is implemented based on the OpenVINO framework. Combined with a task offloading strategy, the tasks of power equipment status monitoring and fault early warning with real-time requirements of ≤5ms are completed on the edge side, reducing cloud-based tasks by more than 50%. Data interaction, along with cross-layer signaling interaction protocols, enables automatic switching to the PLC backup channel when the wireless link is interrupted, ensuring communication continuity. This ultimately forms an integrated collaborative network covering wide-area transmission, local interconnection, and intelligent decision-making, effectively solving the problems of poor communication reliability and slow response in underground power distribution rooms. The 4G / 5G wireless communication sublayer is interconnected with the edge computing management sublayer; the power line carrier communication sublayer is also interconnected with the edge computing management sublayer. The 4G / 5G wireless communication sublayer is primarily responsible for wide-area high-speed data transmission, transmitting collected power data and control command information to the edge computing management sublayer via dedicated network slicing; the power line carrier communication sublayer utilizes… The power distribution room has existing power lines, providing a stable data transmission channel for nearby devices. Its data can also be transmitted to the edge computing management sublayer. At the same time, it serves as a backup communication path when the 4G / 5G wireless communication link is interrupted, receiving and forwarding key instructions from the edge computing management sublayer. The edge computing management sublayer, as the core hub, receives data from the 4G / 5G wireless communication sublayer and the power line carrier communication sublayer. After localization processing and intelligent decision-making, it feeds back control signals to each sublayer. Through cross-layer signaling interaction, it coordinates the resource allocation of the 4G / 5G wireless communication sublayer and the channel scheduling of the power line carrier communication sublayer, ensuring the efficient and stable operation of the entire three-dimensional heterogeneous network.

[0089] In some embodiments, the technical essence of the three-dimensional heterogeneous network collaborative module is to construct an integrated collaborative network that deeply integrates 4G / 5G wireless communication, power line carrier communication and edge computing, and realize the collaborative control of each sub-layer through cross-layer signaling interaction, thereby solving the problems of poor communication reliability and slow response in underground power distribution rooms. During the technical implementation process, the 4G / 5G wireless communication sublayer uses a Slicing-Enabled architecture for dynamic network slicing. Utilizing AI-driven resource prediction algorithms, it pre-allocates a dedicated 5MHz slice for power protection signals. Combined with URLLC and eMBB dynamic priority scheduling mechanisms, it ensures transmission bandwidth ≥200Mbps and end-to-end latency ≤10ms in complex electromagnetic environments. The power line carrier communication sublayer optimizes the OFDM modulation scheme according to the G3-PLC standard. Adaptive spectrum sensing technology is used to detect channel status in the 6-500kHz frequency band, dynamically adjusting subcarrier allocation. Interference cancellation algorithms are used to construct anti-interference subchannels, achieving stable transmission of ≥1Mbps. The edge computing management sublayer uses lightweight model compression technology to compress the YOLOv5 model by 80%, accelerates inference based on the OpenVINO framework, and combines task offloading strategies to process tasks with real-time requirements ≤5ms at the edge, reducing cloud data interaction by more than 50%. Simultaneously, through cross-layer signaling interaction protocols, it automatically switches to the PLC backup channel when the wireless link is interrupted. The principle behind this system is that the 4G / 5G wireless communication sublayer utilizes network slicing technology to isolate different services and dynamically allocates resources according to power service needs; the power line carrier communication sublayer overcomes the problems of high interference and instability in power line channels by optimizing modulation schemes and spectrum sensing technology; and the edge computing management sublayer reduces data transmission pressure and improves processing efficiency through model compression and task offloading. These three layers achieve resource sharing and collaborative scheduling through cross-layer signaling interaction. 4G / 5G handles wide-area high-speed transmission, power line carrier communication provides local interconnection and backup paths, and edge computing enables intelligent decision-making, forming a complementary and collaborative relationship. This module ultimately achieves integrated collaboration of wide-area transmission, local interconnection, and intelligent decision-making, effectively improving the reliability and response speed of communication in underground power distribution rooms, ensuring the stability and real-time performance of power data transmission, and providing a solid communication foundation for intelligent management of power distribution rooms.

[0090] It is understandable that the multimodal encoder group includes a power equipment visual encoder, a power parameter text encoder, and an environmental audio encoder. The power equipment visual encoder, the power parameter text encoder, and the environmental audio encoder are used to process image data, text data, and audio data, respectively. The shared feature projection layer constructs a projection matrix by maximizing cross-modal mutual information. The power equipment topology graph is embedded in the twin network structure. The cross-modal attention fusion module dynamically allocates weights based on a three-dimensional attention mechanism.

[0091] Specifically, the multimodal encoder group includes a power equipment visual encoder, a power parameter text encoder, and an environmental audio encoder. The power equipment visual encoder is based on an improved ResNet50+CBAM architecture, which enables equipment status recognition in underground low-light environments. It introduces an adaptive illumination compensation module and, through the fusion of histogram equalization and the Retinex algorithm, compresses the dynamic range of image brightness to the [0.2, 0.8] interval, thereby improving the extraction accuracy of equipment temperature distribution features in infrared thermal imaging, with a temperature recognition error ≤ ±0.5℃; The power parameter text encoder designs a time-series-semantic dual-path LSTM network to process multi-dimensional power parameters collected by the SCADA system, including current, voltage, and frequency. Through a time window attention mechanism, it assigns weights to data at different time scales, with a 20% increase in weight for data from the 30 minutes prior to a fault, capturing the abnormal evolution trend of the power system; The environmental audio encoder develops a dual feature extractor based on Mel-spectrum-wavelet transform to extract features from abnormal noises in distribution room equipment, including discharge sounds and abnormal motor vibrations. Through joint time-frequency domain enhancement, it can still identify ≥90% of abnormal equipment sounds even in an environment with a signal-to-noise ratio of -5dB. The shared feature projection layer proposes a projection matrix optimization algorithm that preserves power semantics. It constructs projection matrices by maximizing cross-modal mutual information: a visual projection matrix (2048×128) retains equipment spatial structure information; a text projection matrix (1024×128) strengthens the physical correlation between electrical parameters; and an audio projection matrix (1024×128) highlights equipment vibration frequency characteristics. The shared feature projection layer designs an adaptive adjustment mechanism for modal importance, dynamically adjusting projection weights based on the real-time environment of the power distribution room: under normal conditions, text feature weights account for 60%, relying on SCADA data; when a fault is suspected, the visual feature weight increases to 70%, relying on... Relying on infrared images; the twin network structure adopts a three-branch weight-sharing architecture, each branch containing: a spatiotemporal feature enhancement layer, which captures multi-scale features through dilated convolution with dilation rates of 2, 4, and 8; a power domain knowledge injection layer, embedding a power equipment topology graph and circuit breaker-transformer connection relationships; a contrastive learning layer, based on an improved InfoNCE loss function, introducing prior knowledge from the power domain, with positive sample pairs representing different modal data of the same equipment and negative sample pairs representing data from different equipment and at different times; a cross-modal attention fusion module proposes a three-dimensional attention fusion mechanism, simultaneously considering spatial, temporal, and modal dimensions; spatial attention, focusing on equipment relationships. Key components include transformer terminals; time-based attention, highlighting critical time points in fault evolution, particularly the 5 seconds before tripping; modal attention, adaptively allocating weights for each mode, increasing audio weight to 50% during discharge faults; a fault propagation path perception module is designed, modeling electrical connections between devices using a graph neural network (GNN), and using fault propagation probability as a correction factor for attention weights, improving fault location accuracy to over 98%; the output of the multimodal encoder group is connected to the input of the shared feature projection layer; the output of the shared feature projection layer is connected to the input of the twin network structure; and the output of the twin network structure is connected to the input of the cross-modal attention fusion.

[0092] It should be noted that the technical essence of the cross-modal feature fusion module is based on the improved Siamese-NN type of twin neural network to build a multimodal power data feature extractor, which realizes efficient fusion and feature extraction of multimodal power data such as images, text, and audio, and accurately identifies the operating status and faults of power equipment.

[0093] In some embodiments, the three encoders of the multimodal encoder group each perform their respective functions: the power equipment visual encoder adopts an improved ResNet50+CBAM architecture, integrating histogram equalization and the Retinex algorithm for illumination compensation, compressing the image brightness range, and improving the accuracy of equipment temperature feature extraction; the power parameter text encoder processes power parameters through a time-series-semantic dual-path LSTM network combined with a time window attention mechanism; and the environmental audio encoder utilizes a Mel-spectrum-wavelet transform dual feature extractor to enhance time-frequency domain features, enabling the identification of abnormal equipment noise under low signal-to-noise ratio conditions. The shared feature projection layer retains key information of each modality by optimizing the projection matrix and dynamically adjusts the projection weights according to the environment. The three-branch architecture of the Siamese network, through spatiotemporal feature enhancement, domain knowledge injection, and comparative learning, mines multi-scale features and the relationship between the device. The cross-modal attention fusion module allocates attention from three dimensions: space, time, and modality, and corrects the weights based on a graph neural network to locate faults. The principle behind this system is that a multimodal encoder group designs a dedicated encoding method for different modal data characteristics to extract original features; a shared feature projection layer maps multimodal features to a unified space, strengthening the correlation between modalities; a Siamese network structure utilizes contrastive learning and domain knowledge to enhance feature discriminative power; and a cross-modal attention fusion module dynamically focuses on key information based on fault characteristics, mining equipment correlations through graph neural networks. These components work together to achieve end-to-end processing of multimodal data from feature extraction and mapping to fusion. The final result is significant: this module can accurately extract multimodal power data features in complex environments, effectively identify equipment anomalies, and improve fault location accuracy to over 98%. This provides reliable data support for status monitoring and fault early warning of power equipment in underground substations, effectively ensuring the safe and stable operation of the power system.

[0094] Understandably, the state space monitoring unit collects signal strength, signal-to-noise ratio, and device power parameters in real time through multi-source heterogeneous data fusion sensing technology; the routing optimization unit generates routing strategies based on the STAM-MADDPG model; the routing optimization unit abstracts communication nodes as intelligent agents, realizes cross-modal link collaborative scheduling based on the STAM-MADDPG model, and adopts a deep Q-network model containing a three-layer fully connected neural network to minimize latency and packet loss rate, and accelerates routing strategy optimization with the help of an experience replay mechanism; the millisecond-level path switching strategy triggers switching within 15ms through dual thresholds.

[0095] Specifically, the dynamic routing decision engine module constructs a multi-agent deep deterministic policy gradient model STAM-MADDPG based on a spatiotemporal attention mechanism, achieving intelligent scheduling through a three-level routing decision framework. The dynamic routing decision engine module includes a state space monitoring unit, a routing optimization unit, and a millisecond-level path switching strategy. The state space monitoring unit utilizes multi-source heterogeneous data fusion sensing technology to collect parameters such as signal strength, signal-to-noise ratio, and device power in real time. It constructs a dynamic state space by combining a spatial attenuation weight of 0.75 / km, a time penalty of 0.1 points deducted every 1ms, and an interference threshold of -85dBm. It dynamically adjusts parameter weights through a spatiotemporal attention mechanism, prioritizing the capture of key information. The routing optimization unit abstracts communication nodes as agents, achieving cross-modal link collaborative scheduling based on the STAM-MADDPG model. It also employs a deep Q-network model containing a three-layer fully connected neural network to minimize latency and packet loss rate, accelerating strategy optimization through an experience replay mechanism. The millisecond-level path switching strategy uses a 30ms delay and 10... -4 A dual-threshold trigger for packet loss rate recalculation, combined with a pre-switching buffer strategy, enables seamless switching within 15ms. A hybrid link transmission scheme is designed based on business requirements, transmitting real-time commands via 4G / 5G and historical data via power line carrier. A millisecond-level path switching strategy, through dual-threshold triggering and multi-modal link complementarity, achieves efficient transmission, ensuring end-to-end transmission latency ≤30ms and packet loss rate ≤10%. -4 Based on this, the system energy consumption is reduced by more than 25%, significantly improving the reliability and flexibility of the communication network; the state space monitoring unit is interconnected with the millisecond-level path switching strategy; the path optimization unit is interconnected with the millisecond-level path switching strategy.

[0096] It should be noted that the technical essence of the dynamic routing decision engine module is based on the STAM-MADDPG multi-agent deep deterministic policy gradient model with a spatiotemporal attention mechanism. This model constructs a three-level routing decision framework to achieve intelligent scheduling of 4G / 5G and power line carrier communication links, solving the problems of inefficient routing decision-making and unstable transmission in complex underground power distribution room environments. During implementation, the state space monitoring unit uses multi-source heterogeneous data fusion sensing technology to collect parameters such as signal strength, signal-to-noise ratio, and device power in real time. It constructs a dynamic state space by combining preset spatial attenuation weights, time penalties, and interference thresholds, and dynamically adjusts parameter weights through a spatiotemporal attention mechanism to quickly capture key information. The routing optimization unit abstracts communication nodes as agents and achieves cross-modal link collaborative scheduling based on the STAM-MADDPG model. It utilizes a deep Q-network model containing a three-layer fully connected neural network to minimize latency and packet loss rate. An experience replay mechanism accelerates policy optimization and generates the optimal routing policy. The millisecond-level path switching policy is set with a 30ms delay and 10... -4The module employs a dual-threshold packet loss rate mechanism. When monitored data triggers a threshold, route recalculation is initiated. Combined with a pre-switching buffer strategy, seamless switching is achieved within 15ms. A hybrid link transmission scheme is designed based on business requirements to achieve efficient transmission of different data types. The principle behind this is that the state-space monitoring unit efficiently processes complex and ever-changing network state information through data fusion and attention mechanisms; the routing optimization unit, based on multi-agent reinforcement learning, enables collaborative decision-making among communication nodes; and the deep Q-network model continuously optimizes routing strategies by learning from historical experience. A millisecond-level path switching strategy, based on threshold triggering and buffering mechanisms, achieves rapid path switching. The hybrid link transmission scheme combines the advantages of two communication methods to ensure data transmission. These three elements work together to achieve intelligent routing decisions in dynamic environments. Ultimately, this module ensures end-to-end transmission latency ≤30ms and a packet loss rate ≤10%. -4 Based on this, the system energy consumption is reduced by more than 25%, significantly improving the reliability and flexibility of the communication network and effectively meeting the needs of underground power distribution rooms for efficient, stable, and energy-saving data transmission.

[0097] Understandably, the signal subspace separation module separates 4G / 5G signals from power line carrier signals based on an improved singular value decomposition algorithm. The module also constructs a dynamic space-time block coded (STBC) matrix to reassemble the signal, ensuring a bit error rate of less than or equal to 10 even at a signal-to-noise ratio of -5dB. -3 The interference prediction and power control module generates adversarial network interference predictions and dynamically adjusts the transmit power. The cross-modal cooperative anti-interference mechanism is used to trigger NOMA technology to reuse frequency band resources or switch transmission links when the performance of either 4G / 5G or power line carrier is below a threshold, extending the regulation strategy generated by the interference prediction and power control module to multi-modal cooperative scenarios.

[0098] Specifically, the signal subspace separation module utilizes an improved Singular Value Decomposition (SVD) algorithm. Based on the differences in high-frequency broadband and low-frequency narrowband characteristics between 4G / 5G and power line carrier signals, it achieves accurate separation by using a 3x singular value threshold, avoiding frequency band overlap interference. Furthermore, considering the multipath fading problem caused by metal equipment in the power distribution room, it constructs a dynamic space-time block coding (STBC) matrix to reassemble the signal, maintaining a bit error rate ≤10 even under a -5dB signal-to-noise ratio environment. -3The interference prediction and power control module utilizes generative adversarial networks to integrate historical interference data with real-time environmental parameters. It achieves interference distribution prediction with an error of ≤6% through adversarial learning between the discriminator and the generator. It also employs a dual-dimensional dynamic adjustment strategy: on the one hand, it dynamically adjusts the transmit power in 2dBm steps to reduce co-channel interference; on the other hand, it dynamically switches the LoRa spreading factor based on the interference intensity to balance transmission efficiency and anti-interference capability. The cross-modal collaborative anti-interference mechanism establishes an interference compensation linkage strategy between 4G / 5G and power line carrier. When one mode is interfered with, it automatically switches and reuses the resources of another mode to achieve dual-modal link complementarity. The output of the signal subspace separation module is connected to the input of the interference prediction and power control module; the interference prediction and power control module is interconnected with the cross-modal collaborative anti-interference mechanism; the signal subspace separation module is interconnected with the cross-modal collaborative anti-interference mechanism; the signal subspace separation module separates 4G / 5G and power line carrier signals based on the improved SVD algorithm, and after enhancing anti-fading capability through a dynamic STBC matrix, outputs a stable signal to the interference prediction and power control module; the interference prediction and power control module uses a generative adversarial network model to predict the electromagnetic interference distribution, and dynamically adjusts the transmit power and switches the LoRa spreading factor in 2dBm steps based on the prediction results to optimize single-link transmission performance; the cross-modal collaborative anti-interference mechanism monitors the status of the two-mode links in real time, and when the performance of either the 4G / 5G or power line carrier link is lower than the threshold, it triggers NOMA technology to reuse frequency band resources or switch transmission links, extending the control strategy generated by the interference prediction and power control module to multi-modal collaborative scenarios, and the cross-modal collaborative anti-interference mechanism is interconnected with the signal subspace separation module.

[0099] It should be noted that the core technology of the adaptive interference suppression system is to construct a comprehensive interference suppression system through the coordinated operation of three modules: signal subspace separation, interference prediction and power control, and cross-modal collaborative anti-interference. This system addresses the issues of communication signals being susceptible to interference and unstable transmission in the complex electromagnetic environment of underground power distribution rooms. In the implementation process, the signal subspace separation module utilizes an improved SVD algorithm to accurately separate signals based on the frequency characteristics differences between 4G / 5G and power line carrier signals, using a 3x singularity threshold to avoid frequency band overlap interference. It also constructs a dynamic STBC matrix to reassemble signals to address multipath fading. The interference prediction and power control module uses a generative adversarial network to fuse historical and real-time data, predicting interference distribution through adversarial learning. It employs a dual-dimensional dynamic control strategy, adjusting the transmit power in 2dBm steps and switching the LoRa spreading factor based on interference intensity. The cross-modal collaborative anti-interference mechanism monitors the status of the 4G / 5G and power line carrier links in real time. When the performance of either link degrades, it triggers NOMA technology to reuse frequency band resources or switch transmission links, achieving dual-modal link complementarity. The underlying principle is as follows: the signal subspace separation module enhances signal anti-interference and anti-fading capabilities based on signal characteristic differences and matrix recombination technology; the interference prediction and power control module predicts interference using the adversarial learning characteristics of generative adversarial networks and reduces co-channel interference and balances transmission efficiency through dynamic adjustment strategies; and the cross-modal collaborative anti-interference mechanism, based on link status monitoring and NOMA technology, achieves multi-modal resource complementarity and interference compensation. These three modules work together to form a complete anti-interference link, from signal separation enhancement and interference prediction and control to multi-modal collaboration. Ultimately, the system can maintain a bit error rate ≤10% in an environment with a signal-to-noise ratio of -5dB. -3 With an interference distribution prediction error of ≤6%, it effectively improves the stability and reliability of communication links in complex electromagnetic environments, ensures the accuracy and continuity of data transmission in underground power distribution rooms, and significantly reduces the probability of communication failures caused by interference.

[0100] In one possible implementation, the cross-layer collaborative optimization module achieves multi-objective balance by designing a weighted optimization function that includes transmission rate, reliability, and system energy consumption. The transmission rate target is set to be above 100 Mbps, corresponding to a weight of 0.4; reliability is defined as a bit error rate not exceeding 10^6. -6For standard settings, a weight of 0.3 is assigned; system energy consumption requirements are reduced by more than 30%, with a weight of 0.3; a dynamic weight adjustment mechanism is introduced, increasing the reliability weight to 0.5 when transmitting power protection signals; reducing the transmission rate weight to 0.2 when equipment power is insufficient, prioritizing energy efficiency; in resource allocation, a non-dominated sorting genetic algorithm III is used to dynamically allocate the proportion of 4G / 5G and power line carrier transmission, real-time sensing that the 4G / 5G signal strength is not lower than -85dBm and the power line carrier signal-to-noise ratio is not lower than 20dB; when the 4G / 5G load rate reaches more than 70%, 30% of non-real-time services are diverted to the power line carrier link, and frequency band resources are reused through NOMA technology, improving spectrum efficiency by more than 25%; in terms of energy consumption control, when the equipment power drops below 20%, the transmission power is automatically reduced to 50% of the default value, and the sleep cycle is extended. Within 500ms, unnecessary sensors are shut down; at the network level, the CPU utilization of edge servers is adjusted. When the utilization exceeds 80%, dynamic frequency modulation of 1.2-2.4GHz is initiated, and a stepwise strategy of reducing the frequency by 0.3GHz for every 10% increase in utilization is adopted. Computationally intensive tasks are migrated to servers with CPU utilization below 30% to reduce peak energy consumption at single points; at the algorithm optimization level, a constraint processing mechanism is used to ensure that all solutions meet the hard indicators of transmission rate and bit error rate, and exponential penalties are imposed on solutions that do not meet the conditions; an elite retention strategy guided by reference points is adopted, with reference points evenly distributed on the Pareto front, and the solution set is updated every 10 iterations to retain the optimal transmission energy efficiency ratio solution; this module ultimately reduces system energy consumption by 35%-40% while ensuring transmission performance, significantly improving the greenness and reliability of the communication system.

[0101] It should be noted that the technical essence of the cross-layer collaborative optimization module is to construct a multi-objective weighted optimization system. Through dynamic weight adjustment and intelligent resource allocation strategies, it achieves a balance between transmission rate, reliability, and energy consumption in the underground power distribution room communication system, thereby improving the overall performance and greenness of the communication system. In the implementation process, the module first designs a weighted optimization function that includes transmission rate, reliability, and energy consumption, setting a transmission rate ≥ 100Mbps (weight 0.4) and a bit error rate ≤ 10. -6The system aims to reduce energy consumption by ≥30% (weight 0.3) and introduces a dynamic weight adjustment mechanism to flexibly adjust weights based on service type and device status. In terms of resource allocation, the NSGA-Ⅲ algorithm is used to perceive the status of 4G / 5G and power line carrier links, dynamically allocating transmission ratios. When the 4G / 5G load exceeds 70%, non-real-time services are diverted to the power line carrier link, and NOMA technology is used to improve spectrum efficiency. Regarding energy consumption control, at the device level, transmission power is reduced and sleep cycles are extended when power is insufficient; at the network level, frequency is dynamically adjusted and tasks are migrated based on edge server CPU utilization. In algorithm optimization, constraint processing and reference point guidance strategies ensure that the solution meets performance indicators and retains the optimal solution. The principle is that the weighted optimization function and dynamic weight mechanism quantify and integrate different requirements, enabling the system to dynamically adjust the optimization direction according to the actual situation; the NSGA-Ⅲ algorithm achieves efficient resource allocation based on link status data; device-level and network-level energy consumption control strategies reduce energy consumption from the hardware and resource scheduling levels; and the algorithm optimization strategy ensures the effectiveness of the optimization process and the feasibility of the solution. All parts work together to achieve a dynamic balance of multiple objectives. Ultimately, this module reduces system energy consumption by 35%-40% while ensuring transmission performance, significantly improving the reliability and greenness of the communication system. It effectively solves the problem of balancing performance and energy consumption in underground power distribution room communication, and provides strong support for the sustainable operation of intelligent power distribution systems.

[0102] It should be noted that, amidst the wave of smart grids and digital transformation, underground substations, as key hubs of the power system, play a decisive role in the stability and reliability of power supply through their safe and efficient operation. With the continuous development and application of emerging technologies such as the Internet of Things, big data, and artificial intelligence, the demand for intelligent management of underground substations is increasing, placing higher demands on communication systems both inside and outside the substations. Traditional single communication technologies face numerous challenges in underground substation scenarios. While 4G / 5G wireless communication technology boasts advantages such as high speed and wide coverage, in the enclosed and electromagnetically complex environment of underground substations, signals are easily blocked, reflected, and interfered with, leading to severe signal attenuation and unstable transmission. Power line carrier communication technology, which uses power lines to transmit data, is convenient to deploy, but suffers from low transmission rates, high noise interference, and rapid signal attenuation, making it difficult to meet the needs of large-scale real-time data transmission. Furthermore, the substations contain diverse multimodal data types, such as equipment operation status monitoring data and environmental parameter data, and traditional communication technologies cannot effectively achieve efficient fusion and transmission of multimodal data. Existing communication systems also suffer from inflexible routing decisions, insufficient interference suppression capabilities, and high energy consumption during data transmission. Traditional routing algorithms struggle to dynamically adjust transmission paths based on the complex and ever-changing environment of power distribution rooms, resulting in high transmission delays and packet loss rates. Furthermore, they lack effective interference suppression methods to address the strong electromagnetic interference generated by electrical equipment in power distribution rooms, severely impacting communication quality. At the same time, a single communication method cannot balance transmission performance and energy consumption, leading to energy waste and hindering the long-term stable operation of the system.

[0103] Based on this, this application provides a communication system for underground power distribution rooms based on heterogeneous networks and multimodal fusion. The adaptive interference suppression system adopts power domain NOMA-STBC joint anti-interference, separates signals through the SVD algorithm, constructs an STBC matrix to enhance anti-fading capability, and uses a generative adversarial network to predict interference and dynamically adjust power and spreading factor to improve link reliability. The dynamic routing decision engine module is based on the STAM-MADDPG model, monitors key parameters in real time, generates the optimal strategy through a deep Q network, and achieves millisecond-level path switching to ensure low latency and low packet loss transmission. The cross-modal feature fusion module extracts and fuses multimodal data features based on an improved Siamese-NN model. The cross-layer collaborative optimization module uses the NSGA-Ⅲ algorithm to balance the transmission ratio, taking into account high speed, low bit error rate, and energy saving. The three-dimensional heterogeneous network collaborative module integrates a multi-layer architecture to achieve efficient resource allocation. This application significantly improves the security, stability, and intelligence of data transmission in underground power distribution rooms through multi-module collaboration, assists in intelligent management, achieves efficient, stable, and low-power transmission of data in underground power distribution rooms, and has cross-layer data fusion transmission capabilities, meeting the urgent needs of intelligent management of underground power distribution rooms.

[0104] It should be noted that, in terms of data transmission performance, the dynamic routing decision engine module and the adaptive interference suppression system work together. The dynamic routing decision engine, based on the STAM-MADDPG model, monitors key parameters such as signal strength and noise level in real time, and generates the optimal routing strategy by combining a deep Q network, achieving millisecond-level path switching. This results in an end-to-end transmission latency of ≤30ms and a packet loss rate of ≤10%. -4 The adaptive interference suppression system employs a power-domain NOMA-STBC joint anti-interference technology, utilizing the SVD algorithm and a generative adversarial network model to effectively enhance signal anti-fading capabilities, accurately predict interference, and dynamically adjust power, significantly improving link reliability and stability and ensuring efficient data transmission in complex electromagnetic environments. Regarding resource utilization efficiency, the three-dimensional heterogeneous network collaborative module and the cross-layer collaborative optimization module play crucial roles. The three-dimensional fusion system constructed by the three-dimensional heterogeneous network collaborative module enables dynamic allocation of 4G / 5G wireless communication, power line carrier communication, and edge computing resources; the cross-layer collaborative optimization module, based on the NSGA-Ⅲ algorithm, ensures a transmission rate ≥100Mbps and a bit error rate ≤10%. -6 Simultaneously, it reduces system energy consumption by over 30% and can dynamically optimize energy efficiency based on equipment power consumption and server load, significantly improving system resource utilization efficiency and reducing operating costs. In terms of application adaptability, the cross-modal feature fusion module, based on an improved Siamese-NN model, can extract and fuse features from multimodal power data such as images, text, and audio. This meets the diverse data processing needs of underground power distribution rooms, providing strong support for intelligent applications such as equipment status monitoring and fault early warning, greatly expanding the system's application scenarios and value. This application effectively solves the communication problem in underground power distribution rooms through this setup, significantly improving the overall performance and intelligence level of the communication system, which is of great significance for promoting intelligent management of underground power distribution rooms.

[0105] Secondly, referring to Figure 2 , Figure 2This application illustrates a second aspect of an embodiment of an underground power distribution room communication method based on heterogeneous network and multimodal intelligent fusion. This method is applied to the underground power distribution room communication system based on heterogeneous network and multimodal fusion provided in the first aspect embodiment. The underground power distribution room communication system based on heterogeneous network and multimodal fusion includes a three-dimensional heterogeneous network collaboration module, a cross-modal feature fusion module, a dynamic routing decision engine module, an adaptive interference suppression system, and a cross-layer collaborative optimization module. The three-dimensional heterogeneous network collaboration module includes a 4G / 5G wireless communication sublayer, a power line carrier communication sublayer, and an edge computing management sublayer. The power line carrier communication sublayer is connected to the 4G / 5G wireless communication sublayer and the edge computing management sublayer, respectively. The 4G / 5G wireless communication sublayer, the power line carrier communication sublayer, and the edge computing management sublayer achieve coordinated control through cross-layer signaling interaction. The cross-modal feature fusion module includes a multimodal encoder group, a shared feature projection layer, a twin network structure, and a cross-modal attention fusion module. The shared feature projection layer is connected to the multimodal encoder group and the twin network structure, respectively. The twin network structure is also connected to the cross-modal attention fusion module. The multimodal encoder group is also connected to the edge computing management sublayer. The cross-modal feature fusion module is based on an improved twin network structure. A neural network extracts and fuses features from multimodal data. The dynamic routing decision engine module includes a state space monitoring unit, a routing optimization unit, and a millisecond-level path switching strategy. The state space monitoring unit is connected to the routing optimization unit and the cross-modal attention fusion module, respectively. The routing optimization unit is also connected to the millisecond-level path switching strategy. The routing optimization unit uses a STAM-MADDPG model and a deep Q-network model to optimize the routing strategy in real time. The adaptive interference suppression system includes a signal subspace separation module, a cross-modal collaborative anti-interference mechanism, and an interference prediction and power control module. The interference prediction and power control module is connected to... The signal subspace separation module and the cross-modal collaborative anti-interference mechanism are connected. The cross-modal collaborative anti-interference mechanism is also connected to the signal subspace separation module and the millisecond-level path switching strategy. The signal subspace separation module, the cross-modal collaborative anti-interference mechanism, and the interference prediction and power control module combine NOMA-STBC joint anti-interference technology, generative adversarial network interference prediction, and the cross-modal collaborative mechanism to suppress complex electromagnetic interference. The cross-layer collaborative optimization module is connected to the edge computing management sublayer and the state space monitoring unit. The cross-layer collaborative optimization module balances transmission rate, bit error rate, and energy consumption based on the non-dominated sorting genetic algorithm III. This underground power distribution room communication method based on heterogeneous network and multimodal intelligent fusion includes the following steps:

[0106] Step S100, cross-layer data preprocessing:

[0107] Step S110: In the 4G / 5G wireless communication sublayer, spatial coordinate mapping of 4G / 5G baseband signals is performed through BeiDou / GNSS dual-system positioning.

[0108] Step S120: In the power line carrier communication sublayer, the power line carrier signal is clock-synchronized to generate a unified data frame containing timestamp and device ID;

[0109] Step S130: In the edge computing management sublayer, the two types of signal data are integrated and encapsulated into standardized data frames according to dynamic slicing requirements and channel detection requirements for subsequent processing.

[0110] It should be noted that the cross-layer data preprocessing uses BeiDou and GNSS dual-system positioning to control the planar positioning error within 0.8 meters, performs high-precision spatial coordinate mapping between the 4G / 5G baseband signal with a sampling rate of 1.92MHz and the power line carrier signal with a center frequency of 200kHz, and uses high-precision clock synchronization technology to achieve time synchronization within 1 microsecond. The integrated signal data generates a unified data frame containing timestamps, device IDs and signal characteristics. Combined with the dynamic slicing requirements of the 4G / 5G wireless communication sublayer and the channel detection requirements of the power line carrier communication sublayer, it provides basic data for the localized processing and intelligent decision-making of the edge computing management sublayer, supporting the collaborative operation of three-dimensional heterogeneous networks.

[0111] Step S200, Feature fusion encoding:

[0112] In step S210, in the multimodal encoder group, the infrared image is processed by improving the ResNet50+CBAM architecture to compress the brightness to the [0.2, 0.8] range; SCADA parameters are processed by dual-channel LSTM+time window attention; and abnormal noise features are extracted by Mel spectrum-wavelet transform.

[0113] Step S220: In the shared feature projection layer, features are mapped through visual, text, and audio projection matrices, and weights are dynamically adjusted according to the environment.

[0114] Step S230: In the twin network structure, multi-scale features are captured by dilated convolution, embedded into the device topology map, and comparative learning is performed based on the improved InfoNCE loss function.

[0115] In step S240, in the cross-modal attention fusion module, spatial attention is used to focus on key parts of the device, temporal attention is used to capture data 5 seconds before the fault, modal attention is dynamically weighted, and fused features are output.

[0116] It should be noted that the feature fusion encoding is based on an improved Siamese neural network. First, the data is processed separately by a multimodal encoder group: the visual encoder uses an improved ResNet50+CBAM architecture and adaptive illumination compensation to extract equipment status features; the text encoder uses a temporal-semantic dual-path LSTM combined with time window attention to process power parameters; and the audio encoder uses Mel spectrum-wavelet transform to extract abnormal sound features. Then, the power semantic projection matrix and modal weights of the shared feature projection layer are adaptively adjusted to map the features of each modality to a unified space. Subsequently, the feature representation is optimized through spatiotemporal feature enhancement, power knowledge injection, and contrastive learning of the Siamese network structure. Finally, the multimodal features are dynamically fused by the three-dimensional attention mechanism and fault propagation perception of the cross-modal attention fusion module to achieve high-precision extraction and fusion of power equipment status features.

[0117] Step S300, Dynamic route calculation:

[0118] Step S310: The signal strength, signal-to-noise ratio, and device power are collected in real time by the state space monitoring unit, and a state vector is constructed by combining the space attenuation factor, time penalty, and interference threshold.

[0119] Step S320: In the routing optimization unit, the 4G / 5G and power line carrier links are coordinated and scheduled through the STAM-MADDPG model, and the deep Q network model generates routing strategies with the goal of minimizing latency and packet loss rate.

[0120] Step S330: In the millisecond-level path switching strategy, when the latency is ≥30ms or the packet loss rate is ≥10... -4 When the pre-switching buffer is triggered, the path reselection is completed within 15ms. Real-time commands are transmitted via 4G / 5G, while historical data is transmitted via power line carrier.

[0121] It should be noted that dynamic routing calculation relies on the STAM-MADDPG model based on a spatiotemporal attention mechanism. The state space monitoring unit collects parameters such as signal strength, signal-to-noise ratio, and device power in real time. This data is combined with a spatial attenuation weight of 0.75 / km, a time penalty of 0.1 points deducted every 1ms, and an interference threshold of -85dBm to construct a dynamic state space. Key information is then filtered through the spatiotemporal attention mechanism. The routing optimization unit abstracts nodes as intelligent agents and optimizes cross-modal link scheduling strategies based on STAM-MADDPG and a deep Q-network model containing a three-layer fully connected neural network, aiming to minimize latency and packet loss. When the transmission latency reaches 30ms or the packet loss rate reaches 10... -4 When a millisecond-level path switching strategy is triggered, the pre-switching buffer is used to complete path recalculation and switching within 15ms. Transmission tasks for 4G / 5G and power line carrier links are allocated according to service requirements, ultimately achieving end-to-end transmission latency ≤30ms and packet loss rate ≤10%. -4 This reduces system energy consumption by over 25%.

[0122] Step S400, Interference Adaptive Processing:

[0123] Step S410: In the signal subspace separation module, 4G / 5G and power line carrier signals are separated according to the difference of singular values ​​based on the improved singular value decomposition algorithm, and a dynamic STBC matrix is ​​constructed to resist multipath fading.

[0124] Step S420: In the interference prediction and power control module, the interference distribution is predicted by a generative adversarial network model, the transmit power is adjusted in 2dBm steps, and the LoRa spreading factor is dynamically switched according to the interference intensity.

[0125] Step S430: In the cross-modal cooperative anti-interference mechanism, when the bit error rate of a certain link is >10... -4 If the signal-to-noise ratio is less than the threshold, NOMA technology is activated to reuse frequency band resources, or the connection is switched to a backup link.

[0126] It should be noted that the interference adaptive processing adopts a three-level collaborative mechanism: the signal subspace separation module utilizes an improved SVD algorithm, based on the singularity difference between 4G / 5G and power line carrier signals, to achieve accurate separation with a threshold of 3 times, and constructs a dynamic STBC matrix to enhance anti-multipath fading capability; the interference prediction and power control module predicts interference distribution through a generative adversarial network model, with the error controlled within 6%, dynamically adjusts the transmit power in 2dBm steps, and switches the LoRa spreading factor according to the interference intensity; the cross-modal collaborative anti-interference mechanism monitors the link status in real time, and when the performance of a single link is lower than the threshold, it activates the NOMA interference alignment algorithm to allocate subcarrier power at a 1:3 ratio, reuses resources of another mode, and improves the co-channel interference suppression ratio to over 25dB, ensuring that the communication link reliability reaches 99.5% and the bit error rate is no higher than 10%. -4 .

[0127] Step S500, Cross-layer collaborative optimization:

[0128] Step S510: Design a weighted optimization function and dynamically adjust the weights; allocate the proportion of 4G / 5G and power line carrier transmission based on the non-dominated sorting genetic algorithm III. When the 4G / 5G load is ≥70%, divert 30% of non-real-time services to the power line carrier and improve spectrum efficiency through NOMA technology.

[0129] Step S520: For device-level power consumption control, reduce the transmission power by 50% when the battery level is below 20%, and extend the sleep cycle to 500ms.

[0130] Step S530: For network-level power consumption control, when the server CPU utilization is greater than or equal to 80%, the server CPU utilization is reduced by 0.3GHz for every 10% increase, and tasks are migrated to low-load servers.

[0131] Step S540: The indicators are guaranteed through the constraint processing mechanism, and the solution set is updated using an elite retention strategy guided by the reference point.

[0132] It should be noted that underground substations, as key nodes in urban power systems, undertake important tasks such as power distribution, equipment monitoring, and fault handling. Their safe and stable operation highly depends on efficient and reliable communication systems. However, the unique environment of underground substations (enclosed, multi-metallic structures, strong electromagnetic interference) poses severe challenges to communication technologies. Existing solutions generally suffer from the following problems: Traditional reliance on a single communication method (such as using only 4G / 5G wireless communication or only power line carrier communication PLC) is insufficient to cope with the complex and ever-changing underground environment. 4G / 5G signals are easily affected by shielding and multipath fading in underground spaces, resulting in uneven coverage and frequent interruptions; PLCs are susceptible to power grid load fluctuations and impulse noise interference, limiting bandwidth and stability. A single link failure can lead to communication interruptions, affecting monitoring and emergency response. The fusion of multi-source heterogeneous data is difficult: modern intelligent substations need to process sensing data from multiple modalities, such as infrared images, SCADA operating parameters, and equipment abnormal audio. These data differ significantly in feature space and temporal scale. Existing methods lack effective cross-modal feature extraction and fusion mechanisms, making it difficult to comprehensively and accurately characterize equipment status and identify potential faults, resulting in insufficient information value mining. Weak adaptability to dynamic environments: Underground environments suffer from complex and variable electromagnetic interference; equipment movement, switching operations, and the operation of neighboring equipment can all cause drastic fluctuations in channel status. Traditional routing strategies and interference suppression methods are mostly static or rule-driven, unable to perceive environmental changes in real time and make optimal decisions, leading to high transmission latency and high packet loss rates, making it difficult to meet the millisecond-level response requirements of real-time monitoring and control. Insufficient energy efficiency optimization: Underground power distribution rooms have numerous and widely distributed devices, and communication terminals (such as sensors and cameras) typically rely on battery power. Existing communication systems lack refined cross-layer (physical layer, network layer, application layer) collaborative optimization mechanisms, making it difficult to effectively reduce overall system energy consumption while ensuring communication quality (high speed, low error rate), posing challenges to equipment endurance and system sustainability. Limited anti-interference capabilities: Complex electromagnetic interference (such as same-channel interference, impulse noise, and multipath effects) severely affects communication quality. Traditional anti-interference techniques (such as simple filtering or power control) have limited effectiveness in environments with strong dynamic interference, and lack a comprehensive anti-interference system that combines signal processing (such as spatial separation), intelligent prediction (interference distribution), and cross-modal resource coordination. Therefore, there is an urgent need for an underground power distribution room communication method that can deeply integrate heterogeneous network resources, intelligently process multimodal data, dynamically adapt to complex environments, efficiently suppress strong interference, and achieve global energy efficiency optimization, in order to improve the reliability, real-time performance, intelligence, and sustainability of the communication system.

[0133] Based on this, the second aspect of this application provides a communication method for underground power distribution rooms based on heterogeneous networks and multimodal intelligent fusion, which has the following beneficial effects: (1) By constructing a three-dimensional heterogeneous network collaborative module (integrating 4G / 5G wireless communication sublayer, power line carrier communication sublayer, and edge computing management sublayer), dual-channel redundant communication is realized; taking advantage of their respective advantages (4G / 5G has large bandwidth and good mobility; PLC utilizes existing power lines and has strong penetration), and achieving collaborative control through cross-layer signaling interaction. When the performance of a link degrades, the system can quickly switch or divert traffic, significantly improving the overall reliability and robustness of the communication system and overcoming the inherent defects of a single communication method. (2) Solving the problem of difficult fusion of multi-source heterogeneous data: Through the cross-modal feature fusion module (including multimodal encoder group, shared feature projection layer, twin network structure, and cross-modal attention fusion module), especially based on improved twin neural network and three-dimensional attention mechanism (space, time, modality), efficient and accurate feature extraction and deep fusion of multimodal data such as infrared images, SCADA parameters, and audio noise are realized. This provides more comprehensive and accurate feature inputs for subsequent advanced applications such as status monitoring and fault diagnosis, greatly enhancing the value of information utilization and intelligent analysis capabilities. (3) Solving the problem of weak adaptability to dynamic environments: The dynamic routing decision engine module (state space monitoring unit, routing optimization unit, millisecond-level path switching strategy) is the key. It uses the STAM-MADDPG model (multi-agent deep deterministic policy gradient combined with spatiotemporal attention mechanism) and deep Q network model to monitor the network status (signal strength, signal-to-noise ratio, power consumption, etc.) in real time, construct a dynamic state space, and intelligently optimize the routing strategy. The millisecond-level path switching strategy can complete a fast switch within 15ms when the link performance deteriorates (delay ≥30ms or packet loss rate exceeds the standard), ensuring that real-time instructions are given priority through low-latency links (4G / 5G) and historical data is passed through high-bandwidth links (PLC). This significantly reduces end-to-end transmission delay (≤30ms) and packet loss rate, meeting the stringent requirements of real-time communication in underground power distribution rooms. (4) Addressing the issue of limited anti-interference capability: The adaptive interference suppression system (signal subspace separation module, interference prediction and power control module, and cross-modal collaborative anti-interference mechanism) constructs a three-level collaborative anti-interference system: Signal separation: The improved SVD algorithm accurately separates 4G / 5G and PLC signals, combined with a dynamic STBC matrix to resist multipath fading. Intelligent prediction and control: The GAN model predicts the interference distribution, guiding the fine adjustment of power in 2dBm steps and dynamic switching of LoRa spreading factors. Cross-modal collaboration: When the performance of a single link deteriorates, NOMA technology is activated to reuse the spectrum resources of another mode (allocating power at a 1:3 ratio), or the link is quickly switched, increasing the co-channel interference suppression ratio to over 25dB. This mechanism significantly improves the reliability of communication links (≥99.5%) and reduces the bit error rate under complex electromagnetic interference environments.(5) Solving the problem of insufficient energy efficiency optimization: Function: The cross-layer collaborative optimization module is the core of system-level optimization. It is based on the improved non-dominated sorting genetic algorithm III (NSGA-III) and constructs a multi-objective optimization framework: Service diversion and spectrum efficiency enhancement: Dynamically allocate the transmission ratio of 4G / 5G and PLC. When 4G / 5G is under high load (≥70%), non-real-time services are diverted (30%) to PLC, and NOMA is used to improve spectrum efficiency. (6) Refined energy consumption control: Equipment level: When the power is below 20%, the transmit power is significantly reduced (50%) and the sleep cycle is extended (500ms); Network level: When the server CPU utilization is high (≥80%), the frequency is dynamically reduced (0.3GHz is increased for every 10% utilization) and the task is migrated to the low-load server; Global optimization guarantee: The target (rate ≥100Mbps, low bit error rate, energy consumption reduction ≥30%) is set through the weighted optimization function. Combined with the constraint processing mechanism and the reference point-guided elite strategy, the optimization results are ensured to meet the requirements. This mechanism achieves a globally optimal balance between transmission performance and energy consumption, significantly reducing overall system energy consumption (35%-40%). The technical solution presented in this application systematically and specifically addresses key bottlenecks in existing underground power distribution room communication technologies in terms of reliability, multi-source data fusion, dynamic environmental adaptability, strong interference resistance, and energy efficiency optimization through five core technological innovations: deep integration of heterogeneous networks, multi-modal intelligent feature fusion, dynamic intelligent routing decision-making, adaptive collaborative anti-interference, and cross-layer global optimization. It not only improves the quality and reliability of communication links and ensures the real-time performance of critical services, but also significantly reduces system energy consumption through intelligent data processing and resource scheduling. This provides a solid communication foundation for building an intelligent, highly reliable, green, and sustainable underground power distribution room operation system. Its technical content is a direct, effective, and innovative response and solution to the problems and challenges described in the background technology.

[0134] It should be noted that cross-layer collaborative optimization constructs a three-level collaborative mechanism through an improved non-dominated sorting genetic algorithm III: first, a weighted optimization function is designed, with a transmission rate ≥100Mbps and a bit error rate ≤10. -6 The system aims to reduce energy consumption by ≥30% and dynamically adjusts weights. Then, a non-dominated sorting genetic algorithm (NOMA) is used to allocate the proportion of 4G / 5G and power line carrier transmission. When the 4G / 5G load exceeds 70%, 30% of non-real-time services are offloaded, and NOMA is used to improve spectrum efficiency. Energy consumption control is then implemented: when device battery level is below 20%, transmit power is reduced and sleep cycles are extended; when server CPU utilization exceeds 80%, frequency is dynamically adjusted and tasks are migrated. Finally, a constraint processing mechanism ensures the performance targets are met, and an elite retention strategy guided by reference points is used to update the solution set, ultimately achieving a 35%-40% reduction in system energy consumption.

[0135] Understandably, the formula for calculating the attention weights in the dynamic allocation of modal attention weights is as follows:

[0136]

[0137] Among them, e t =score(h t ,q) represents the attention score, h t Let α represent the hidden state of the LSTM at time t, q represent the query vector, and α represent the hidden state of the LSTM at time t. t The attention weight at time t is represented by β, which is used to weight and sum the data at different times to obtain the feature representation. t This represents the spatiotemporal coupling coefficient, reflecting the differences in the importance of data at different times, and is dynamically adjusted based on the dynamic changing trends of the power system; w m Represents modal interaction weights, measuring the degree of correlation between multimodal data, and is dynamically learned through a cross-modal attention mechanism; ρ t,m This represents the equipment topology constraint terms, introduces a power equipment topology relationship graph, and quantifies the correlation between the current data and the fault propagation path; This represents modal feature enhancement terms, which enhance the characteristics of different modal data.

[0138]

[0139] Wherein, s(x i ,x j ) represents sample x i and x j The similarity score; τ represents the temperature parameter, which is used to control the difficulty of contrastive learning. By minimizing the InfoNCE loss function, the distance between positive sample pairs is shortened, and the distance between negative sample pairs is widened. The topology constraint weights of the devices are represented by θ, which is calculated using a graph neural network based on the power equipment topology graph to determine the topological correlation between device nodes in a positive sample pair. i,j ε represents the negative sample association penalty term, which incorporates prior knowledge in the power field to introduce a dynamic penalty mechanism for negative sample pairs; ε represents the smoothing factor to ensure the stability of the loss function calculation.

[0140] Specifically, refer to Figure 3 The feature fusion encoding is based on the improved Siamese-NN model, which achieves deep feature extraction and fusion of multimodal power data through multi-module collaboration. Step S200 also includes, but is not limited to, the following steps:

[0141] Step S201, Multimodal data feature extraction:

[0142] The power equipment visual encoder adopts an improved ResNet50+CBAM architecture, combined with an adaptive illumination compensation module to process infrared thermal imaging images. By fusing histogram equalization with the Retinex algorithm, the dynamic range of image brightness is compressed to the [0.2, 0.8] interval, improving the accuracy of temperature distribution feature extraction. The power parameter text encoder uses a time-series-semantic dual-path LSTM network to process multi-dimensional power parameters (current I, voltage V, and frequency f) collected by the SCADA system. A time window attention mechanism is introduced to assign weights to data at different time scales, with a 20% increase in weight for data from the 30 minutes before a fault, highlighting the constant evolution trend. Combining the dynamic characteristics of the power system with multimodal data correlation, spatiotemporal coupling coefficients, modal interaction weights, and equipment topology constraints are introduced. The attention weights are calculated as follows:

[0143]

[0144] In formula (1), e t =score(h t ,q) represents the attention score, h t Here, α is the hidden state of the LSTM at time t, q is the query vector, and α is the hidden state. t Let β be the attention weight at time t. The feature representation is obtained by weighted summation of data from different times using this weight. t The spatiotemporal coupling coefficient reflects the difference in importance of data at different times and is dynamically adjusted based on the dynamic changing trend of the power system; w m Modal interaction weights, measuring the degree of correlation between multimodal data, are dynamically learned through a cross-modal attention mechanism; ρ t,m For equipment topology constraints, a power equipment topology graph is introduced to quantify the correlation between current data and fault propagation paths; The modal feature enhancement term enhances the characteristics of different modal data. This formula significantly improves the adaptability and representation capability of the attention mechanism to complex power system scenarios by integrating spatiotemporal dynamics, multimodal interaction, topological constraints, and feature enhancement. The environmental audio encoder uses a dual feature extractor based on Mel spectrum-wavelet transform to extract features of abnormal noises from equipment in the power distribution room. Through joint time-frequency domain enhancement, it achieves an abnormal sound recognition rate of ≥90% in an environment with a signal-to-noise ratio of -5dB.

[0145] Step S202, Shared feature projection:

[0146] By maximizing cross-modal mutual information (MI), three types of projection matrices are constructed: visual projection matrix P. vis ∈

[0147] R 2048*128 Preserve equipment spatial structure information; text projection matrix P txt ∈R 2048*128Strengthen the physical correlation of electrical parameters; audio projection matrix P aud ∈R 2048*128 Highlighting the vibration frequency characteristics of the equipment;

[0148] The projection weights are dynamically adjusted based on the real-time environment of the power distribution room. Under normal conditions, the weight of text features accounts for 60%; when a fault is suspected, the weight of visual features is increased to 70%.

[0149] Step S203, Siamese network feature enhancement and contrastive learning:

[0150] Dilated convolutions with dilation rates of 2, 4, and 8 are employed to capture multi-scale features and enhance feature representation capabilities. A power equipment topology graph is embedded, integrating the circuit breaker-transformer connection relationships into the network to improve the model's understanding of the power system structure. Comparative learning is performed based on an improved InfoNCE loss function, incorporating prior knowledge from the power domain. Positive sample pairs represent different modalities of the same device, while negative sample pairs represent data from different devices and at different times. To enhance the adaptability and complexity of the InfoNCE loss function in the power domain, dynamic adjustment mechanisms for device topology constraints, fault propagation probability, and modal weights are introduced. The loss function is as follows:

[0151] In formula (2), s(x i ,x j ) is sample x i and x j The similarity score, where τ is a temperature parameter used to control the difficulty of contrastive learning. By minimizing this loss function, the distance between positive sample pairs is shortened, while the distance between negative sample pairs is widened. As the topology constraint weights for equipment, based on the power equipment topology graph, a graph neural network (GNN) is used to calculate the topological correlation degree between equipment nodes in positive sample pairs; θ i,j For negative sample association penalty terms, a dynamic penalty mechanism is introduced for negative sample pairs based on prior knowledge in the power field; ε is a smoothing factor to ensure the stability of the loss function calculation.

[0152] Step S204, Cross-modal attention fusion:

[0153] A three-dimensional attention mechanism is constructed, including spatial attention, temporal attention, and modal attention; the electrical connection relationship between devices is modeled based on graph neural network (GNN), and the fault propagation probability is used as a correction factor for attention weights to improve the fault location accuracy to over 98%; finally, high-dimensional feature vectors are generated by fusion to complete the deep fusion of multimodal data.

[0154] For example, by using a time window attention mechanism, the weight of data in the 30 minutes prior to a fault is increased by 20%, effectively highlighting the abnormal evolution trend of the power system and improving the early warning capability of faults. The dynamic adjustment mechanism of the spatiotemporal coupling coefficient β_t enables the model to adapt to the dynamic changes in the power system, improving the response speed to time-varying faults. The dynamic learning mechanism of modal interaction weights w_m quantifies the correlation between data from different modalities, promoting cross-modal information complementarity. The equipment topology constraint term ρ_(t,m) introduces a power equipment topology graph, enabling the model to understand the power system structure and improving its ability to analyze fault propagation paths. The dilated convolution multi-scale feature capture mechanism (dilation rates 2, 4, 8) enhances the model's ability to perceive features of different granularities and improves the richness of feature representation. The improved InfoNCE loss function introduces equipment topology constraints. A penalty term θ_(i,j) is associated with negative samples to enhance the contrastive learning effect and improve the clustering compactness of similar samples and the separation of dissimilar samples. A three-dimensional attention mechanism combined with a graph neural network (GNN) uses the fault propagation probability as an attention weight correction factor, significantly improving fault localization accuracy to over 98%. A dynamic modality weight adjustment mechanism enables the model to automatically focus on key modal data in different scenarios, enhancing its diagnostic capabilities for complex faults. Table 1 shows the effect and data analysis of Siamese network feature enhancement and contrastive learning.

[0155] Table 1. Effects and Data Analysis of Twin Network Feature Enhancement and Contrastive Learning

[0156]

[0157]

[0158] It is understandable that after step S310, the following steps are included:

[0159] Step S311: The parameter weights are dynamically adjusted through a spatiotemporal attention mechanism, and multimodal information fusion, dynamic environment perception and network topology constraints are introduced.

[0160] In this step, the attention weight α of the parameter weights is dynamically adjusted through a spatiotemporal attention mechanism. i The calculation formula is as follows:

[0161]

[0162] Where, f(s) i ) represents the attention scoring function, which obtains a score by calculating the correlation between each parameter and the current routing decision; α i Used for weighted aggregation of state vectors, prioritizing the capture of key parameters affecting routing decisions; γ t,ib represents the time-based dynamic adjustment coefficient, which dynamically adjusts the attention weights of parameters at different times based on historical latency fluctuations in the communication link and the urgency of real-time transmission tasks; β m,i The modality fusion weights represent the multimodal features of 4G / 5G and power line carrier communication, dynamically allocating weights based on link type and interference status; g(G) represents the network topology constraint term, introducing the communication network topology graph G, and calculating the structural relationships between nodes through a graph convolutional network; δ n,i This represents the topology influence coefficient, which dynamically adjusts the influence of topology constraints based on the role of a node in the network topology.

[0163] It is understandable that in step S320, the 4G / 5G and power line carrier links are coordinated and scheduled using the STAM-MADDPG model, and the deep Q network model generates routing strategies with the goal of minimizing latency and packet loss rate, including:

[0164] Step S321: The deep Q-network model learns the optimal routing strategy through the improved deep Q-network loss function, balancing immediate rewards and long-term benefits. The calculation formula for the improved deep Q-network loss function is as follows:

[0165] L = E s,a,r,s′ [w delay *(y delay -Q delay (s,a;θ)) 2 +w packet

[0166] *(y packet -Q packet (s,a;θ)) 2 ]

[0167] Among them, y delay =r delay +γ delay max a′ Q delay (S′,a′;θ - ),y packet =r packet +γ packet max a′ Q packet (S′,a′;θ - ); r represents the immediate reward, a positive reward for successful transmission and a negative reward for packet loss; γ represents the discount factor, ranging from 0 to 1, balancing immediate and long-term benefits; θ represents the current network parameters; θ - The target network parameters are represented and updated periodically to stabilize training. The (s,a,r,s′) are stored in the experience pool through the experience replay mechanism and randomly sampled for training.

[0168] Specifically, refer to Figure 4 Dynamic routing computation relies on the STAM-MADDPG model based on spatiotemporal attention mechanism and deep Q-network model, and achieves intelligent scheduling through a three-level routing decision framework. Step S300 includes, but is not limited to, the following steps:

[0169] Step S301, Dynamic state space construction and key information extraction:

[0170] The state space monitoring unit utilizes multi-source heterogeneous data fusion sensing technology to collect parameters such as signal strength S, signal-to-noise ratio SNR, and device power E of each communication link in real time; combined with a spatial attenuation weight W of 0.75 / km. space A time penalty of 0.1 points is deducted every 1ms. time and -85dBm interference threshold T int Construct a dynamic state space; quantize the parameters into a state vector s = [S, SNR, E], where spatial attenuation causes the signal strength to change according to S. new =S d *w space Attenuation, where d is the transmission distance, and the time penalty reduces the state cost C. t =C t-1 +p*cumulative interference exceeds threshold T int The system triggers early warnings in real time; it dynamically adjusts parameter weights through a spatiotemporal attention mechanism to highlight key information; to improve the accuracy and adaptability of the spatiotemporal attention mechanism in underground power distribution room communication scenarios, it introduces multimodal information fusion, dynamic environmental perception, and network topology constraints, with attention weight α... i The calculation is as follows:

[0171] In formula (3), f(s) i α is the attention scoring function, which calculates the score by determining the correlation between each parameter and the current routing decision. i Used for weighted aggregation of state vectors, prioritizing the capture of key parameters affecting routing decisions; γ t,i β is a time-based dynamic adjustment coefficient that dynamically adjusts the attention weights of parameters at different times based on historical latency fluctuations in the communication link and the urgency of real-time transmission tasks; m,i The modality fusion weights integrate the multimodal features of 4G / 5G and power line carrier communication, and dynamically allocate weights based on link type and interference status; g(G) is the network topology constraint term, introducing the communication network topology graph G, and calculating the structural correlation between nodes through a graph convolutional network GCN; δ n,i The topology influence coefficient dynamically adjusts the influence of topology constraints based on the role of nodes in the network topology. Through multi-dimensional dynamic adjustment and topology constraints, the adaptability of the spatiotemporal attention mechanism to complex communication environments is significantly enhanced.

[0172] Step S302: Cross-modal link cooperative scheduling and strategy optimization:

[0173] The routing optimization unit abstracts communication nodes as agents, and implements cross-modal, 4G / 5G and power line carrier link coordinated scheduling based on the STAM-MADDPG model; each agent is based on local state s_i and global state s_i. global Perform action a i A deep Q-network model containing three fully connected neural networks is used to minimize latency L. delay With packet loss rate L packet The objective is to optimize the strategy; the network input is the state vector s, and the output is the Q-value Q(s,a) of each action; to enhance the optimization capability of the deep Q-network model in the communication scenario of the underground power distribution room, a multi-objective dynamic weighting, network topology awareness, and state transition prediction mechanism are introduced, and the loss function is defined as:

[0174] L = E s,a,r,s′ [w delay *(y delay -Q delay (s,a;θ)) 2 +w packet *

[0175] (y packet -Q packet (s,a;θ)) 2 (4)

[0176] In formula (4), y delay =r delay +γ delay max a′ Q delay (S′,a′;θ - ),y packet =r packet +γ packet max a′ Q packet (S',a′;θ - ), r is the immediate reward, a positive reward for successful transmission and a negative reward for packet loss; γ is the discount factor, ranging from 0 to 1, balancing immediate and long-term benefits; θ is the current network parameter. - The target network parameters are updated periodically to stabilize training. (s,a,r,s′) are stored in the experience pool through an experience replay mechanism, and training is carried out by random sampling to avoid overfitting caused by data correlation. The deep Q network model significantly improves the policy optimization capability in complex communication environments through mechanisms such as multi-objective dynamic weighting, network topology awareness, and state transition prediction.

[0177] Step S303, Millisecond-level path switching and hybrid link transmission:

[0178] When a transmission delay L is detected delay ≥30ms or packet loss rate L packet ≥10 -4 Upon activation, a millisecond-level path switching strategy is triggered, initiating route recalculation. A pre-switching buffer strategy pre-caches data to be transmitted, completing the path switch within 15ms. Real-time commands are transmitted via 4G / 5G, while historical data is transmitted back via power line carrier, achieving hybrid link transmission. After the path switch, the state space monitoring unit continuously collects data, and the route optimization unit, based on the new state update strategy, ensures end-to-end transmission latency ≤30ms and packet loss rate ≤10%. -4 And reduce system energy consumption by more than 25%.

[0179] For example, the state space monitoring unit utilizes a spatiotemporal attention mechanism, combined with multimodal information fusion, dynamic environmental perception, and network topology constraints, to calculate attention weights using formula (3). Based on historical delay fluctuations in the communication link, link type interference status, and network topology, it can dynamically adjust the level of attention paid to parameters such as signal strength, signal-to-noise ratio, and device power consumption, prioritizing the capture of key information affecting routing decisions and significantly improving the accuracy and timeliness of routing decisions. Table 2 shows the spatiotemporal attention weight data analysis table:

[0180] Table 2. Spatiotemporal Attention Weight Data Analysis Table

[0181]

[0182] The routing optimization unit is based on the STAM-MADDPG model and the improved deep Q network model. Using the loss function defined by formula (4), it aims to minimize latency and packet loss rate. Combined with multi-objective dynamic weighting, network topology awareness, and state transition prediction mechanisms, it achieves coordinated scheduling of 4G / 5G and power line carrier links. The experience playback mechanism avoids overfitting, enabling the model to quickly learn and generate optimal routing strategies in complex communication environments, significantly improving network resource utilization and transmission efficiency. The millisecond-level path switching strategy responds rapidly when transmission latency or packet loss rate exceeds the threshold. With the help of a pre-switching buffer strategy, path switching is completed within 15ms, ensuring data transmission continuity. Simultaneously, the hybrid link transmission mode rationally allocates real-time and historical data transmission paths. Combined with dynamic routing optimization, it ensures end-to-end transmission latency ≤30ms and packet loss rate ≤10%. -4 Based on this, system energy consumption is reduced by over 25%, effectively improving the reliability and economy of the communication network. Table 3 shows the data analysis of cross-modal link collaborative scheduling and strategy optimization:

[0183] Table 3. Data Analysis of Cross-Modal Link Cooperative Scheduling and Strategy Optimization

[0184]

[0185] It is understood that step S400 also includes, but is not limited to, the following steps:

[0186] Step S440: By fusing historical interference data with real-time environmental parameters using a generative adversarial network, the generator and discriminator learn adversarially to predict the interference distribution. The generator attempts to generate realistic interference samples. The discriminator distinguishes between real interference and generated samples.

[0187] In this step, the objective function for generating the adversarial network model is as follows:

[0188]

[0189] Where, max D This indicates that the discriminator D attempts to maximize its own discriminative power, min G Let V(D,G) represent the generator G's attempt to minimize the discriminative power of the discriminator, and let V(D,G) represent the optimization objective function of the generative adversarial network. Let log D(I) represent the expectation of the real disturbance data, where log D(I) follows the real data distribution P. data log D(I) represents the logarithmic probability of the discriminator D classifying the true sample I. Represents the noise vector Expectations Follows the prior distribution P noise ; This indicates that the discriminator D affects the generated sample G(P). real The logarithm of the probability complement of the judgment, P real This represents real-time environmental parameters.

[0190] Specifically, refer to Figure 5 The interference adaptive processing is based on the power domain NOMA-STBC joint anti-interference architecture, and is implemented through a three-level mechanism of signal separation, interference prediction and cross-modal cooperation. Step S400 includes, but is not limited to, the following steps:

[0191] Step S401, Signal subspace separation and anti-fading enhancement

[0192] Using an improved Singular Value Decomposition (SVD) algorithm, the mixed signal X is decomposed based on the differences in characteristics between 4G / 5G high-frequency broadband and power line carrier low-frequency narrowband: X=U∑V T Set a singularity threshold τ of 3 times, retain singularity components greater than the threshold, and convert the 4G / 5G signal X... 4G / 5G With power line carrier signal X PLCSeparation is achieved to avoid frequency band overlap interference; where U and V are unitary matrices, and Σ is a diagonal matrix containing singular values, signal separation is achieved by truncating elements less than τ in Σ; to address the multipath fading problem, a dynamic space-time block coded (STBC) matrix C is constructed to reassemble the signal; taking a 2×2 STBC matrix as an example, the signal vector s is transmitted at time t. t =

[0193] [s t,1 ,,s t,2 ] T After encoding, it becomes By adjusting matrix parameters to adapt to channel variations, a bit error rate of ≤10% is maintained even at a signal-to-noise ratio of -5dB. -3 ;

[0194] Step S402, Interference prediction and dynamic power control:

[0195] Generative adversarial networks (GANs) are used to fuse historical interference data D. hist With real-time environmental parameter P real The generator G and discriminator D use adversarial learning to predict the distribution of interference; the generator attempts to generate realistic interference samples. The discriminator distinguishes between real interference I and generated samples; the objective function is:

[0196]

[0197] After training, the model's prediction error is ≤6%, and it outputs predicted values ​​of future disturbance distribution. The transmit power P is dynamically adjusted in 2dBm steps. t When co-channel interference increases, P t+1 =P t -2; conversely, P t+1 =P t +2, reduce interference impact; based on interference intensity I level The LoRa spreading factor SF is dynamically switched; when the interference is strong, SF is increased to improve the anti-interference capability; when the interference is weak, SF is decreased to increase the transmission rate, thus balancing transmission efficiency and anti-interference performance.

[0198] Step S403, cross-modal cooperative anti-interference:

[0199] Real-time monitoring of the bit error rate (BER) and signal-to-noise ratio (SNR) performance metrics of 4G / 5G and power line carrier links; setting a threshold BER th and SNR th When any link BER>BER th or SNR <SNR thWhen the signal is received, a coordination mechanism is triggered; power domain non-orthogonal multiple access (NOMA) technology is activated, subcarrier power is allocated at a 1:3 ratio, and frequency band resources are reused; if the performance of the single-mode link continues to deteriorate, the signal is switched to another mode link for transmission, and the control strategy generated by the interference prediction module is applied to achieve dual-mode link complementarity, improving the co-channel interference suppression ratio to over 25dB, ensuring communication link reliability of 99.5%, and reducing the bit error rate to 10%. -4 the following.

[0200] For example, this interference adaptive processing scheme comprehensively enhances communication anti-interference capabilities through a three-level collaborative mechanism. Signal separation and anti-fading enhancement achieve precise signal processing, ensuring stable transmission under low signal-to-noise ratio conditions; interference prediction and power regulation effectively predict and respond to interference; cross-mode collaboration flexibly switches when the link is abnormal, significantly enhancing communication reliability. Specific technical effects are as follows: This interference adaptive processing scheme is based on a power domain NOMA-STBC joint anti-interference architecture, significantly improving the anti-interference performance and reliability of the communication system through a three-level mechanism. Regarding signal subspace separation and anti-fading enhancement, the improved SVD algorithm achieves precise separation of 4G / 5G and power line carrier signals. Combined with a dynamic STBC matrix, the bit error rate is controlled to ≤10 in a -5dB signal-to-noise ratio environment. -3 The interference prediction and dynamic power control module utilizes generative adversarial networks to control interference prediction errors to ≤6%, and effectively reduces co-channel interference by dynamically adjusting transmit power and LoRa spreading factor. The cross-modal collaborative anti-interference mechanism monitors link performance in real time, and upon triggering, reuses frequency band resources through NOMA technology or switches links and applies interference control strategies to improve the co-channel interference suppression ratio to over 25dB, ensuring communication link reliability of 99.5% and reducing the bit error rate to 10%. -4 The following describes the stable and efficient communication transmission achieved in complex electromagnetic environments. Table 4 presents the data analysis of interference prediction and dynamic power control:

[0201] Table 4 Data Analysis Table for Interference Prediction and Dynamic Power Regulation

[0202]

[0203]

[0204] It should be noted that Deep Q-Network is abbreviated as DQN.

[0205] STAM-MADDPG stands for Spatio-Temporal Attention Mechanism-based Multi-Agent Deep Deterministic Policy Gradient; NOMA-STBC stands for Non-Orthogonal Multiple Access with Space-Time Block Coding; NSGA-Ⅲ stands for Non-dominated Sorting Genetic Algorithm III; MI stands for Mutual Information; STBC stands for Space-Time Block Coding; NOMA stands for Non-Orthogonal Multiple Access; ResNet50+CBAM stands for Residual Network 50 with Convolutional Block Attention Module; SCADA stands for Supervisory Control and Data Interaction. Acquisition refers to a monitoring and data acquisition system; InfoNCE stands for Info Noise Contrastive Estimation; CPU stands for Central Processing Unit; LoRa stands for Long Range Radio; and PLC's full English and Chinese names are as follows: English: Power Line Communication; Chinese: Power Line Carrier Communication (or simply Power Line Carrier Communication).

[0206] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved; for example, the described methods may be performed in an order different from that described. Additionally, features described with reference to certain examples may be combined in other examples.

[0207] In the description of the embodiments of this application, the terms "first," "second," "third," and "fourth" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first," "second," "third," or "fourth" may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0208] In the description of the embodiments of this application, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0209] Although embodiments of this application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A communication system for underground power distribution rooms based on heterogeneous networks and multimodal fusion, characterized in that, include: A three-dimensional heterogeneous network collaboration module includes a 4G / 5G wireless communication sublayer, a power line carrier communication sublayer, and an edge computing management sublayer. The power line carrier communication sublayer is connected to the 4G / 5G wireless communication sublayer and the edge computing management sublayer, respectively. The 4G / 5G wireless communication sublayer, the power line carrier communication sublayer, and the edge computing management sublayer achieve collaborative control through cross-layer signaling interaction. The cross-modal feature fusion module includes a multimodal encoder group, a shared feature projection layer, a Siamese network structure, and a cross-modal attention fusion module. The shared feature projection layer is connected to the multimodal encoder group and the Siamese network structure, respectively. The Siamese network structure is also connected to the cross-modal attention fusion module. The multimodal encoder group is also connected to the edge computing management sublayer. The cross-modal feature fusion module extracts and fuses multimodal data features based on an improved Siamese neural network. The dynamic routing decision engine module includes a state space monitoring unit, a routing optimization unit, and a millisecond-level path switching strategy. The state space monitoring unit is connected to the routing optimization unit and the cross-modal attention fusion module, respectively. The routing optimization unit is also connected to the millisecond-level path switching strategy. The routing optimization unit uses the STAM-MADDPG model and the deep Q network model to optimize the routing strategy in real time. An adaptive interference suppression system includes a signal subspace separation module, a cross-modal cooperative anti-interference mechanism, and an interference prediction and power control module. The interference prediction and power control module is connected to the signal subspace separation module and the cross-modal cooperative anti-interference mechanism, respectively. The cross-modal cooperative anti-interference mechanism is also connected to the signal subspace separation module and the millisecond-level path switching strategy, respectively. The signal subspace separation module, the cross-modal cooperative anti-interference mechanism, and the interference prediction and power control module combine NOMA-STBC joint anti-interference technology, generative adversarial network interference prediction, and the cross-modal cooperative mechanism to suppress complex electromagnetic interference. A cross-layer collaborative optimization module is connected to the edge computing management sub-layer and the state space monitoring unit, respectively. The cross-layer collaborative optimization module balances transmission rate, bit error rate and energy consumption based on non-dominated sorting genetic algorithm III.

2. The underground power distribution room communication system based on heterogeneous network and multimodal fusion according to claim 1, characterized in that, The 4G / 5G wireless communication sublayer adopts dynamic network slicing technology to allocate a dedicated 5MHz bandwidth for power protection signals, so that the transmission bandwidth is ≥200Mbps. The power line carrier communication sublayer optimizes OFDM modulation based on the G3-PLC standard, achieving a rate of ≥1Mbps in the 6-500kHz frequency band. The edge computing management sublayer introduces lightweight model compression technology and implements hardware-accelerated model inference based on the OpenVINO framework. Combined with a task offloading strategy, the tasks of power equipment status monitoring and fault early warning with real-time requirements of ≤5ms are completed on the edge side. Furthermore, the edge computing management sublayer automatically switches to the power line communication backup channel when the wireless link is interrupted through a cross-layer signaling interaction protocol. The 4G / 5G wireless communication sublayer is used to transmit the collected power data and control command information to the edge computing management sublayer through a dedicated network slice. The power line carrier communication sublayer is used to utilize the existing power lines in the power distribution room to provide a stable data transmission channel for near-field devices, transmit the data to the edge computing management sublayer, and serve as a backup communication path when the 4G / 5G wireless communication link is interrupted, receiving and forwarding key commands from the edge computing management sublayer. The edge computing management sublayer is used to perform localized processing and intelligent decision-making on the data from the 4G / 5G wireless communication sublayer and the power line carrier communication sublayer respectively to generate control signals, and coordinate the resource allocation of the 4G / 5G wireless communication sublayer and the channel scheduling of the power line carrier communication sublayer through cross-layer signaling interaction based on the control signals.

3. The underground power distribution room communication system based on heterogeneous network and multimodal fusion according to claim 1, characterized in that, The multimodal encoder group includes a power equipment visual encoder, a power parameter text encoder, and an environmental audio encoder. The power equipment visual encoder, the power parameter text encoder, and the environmental audio encoder are used to process image data, text data, and audio data, respectively. The shared feature projection layer constructs a projection matrix by maximizing cross-modal mutual information. The twin network structure embeds a power equipment topology graph. The cross-modal attention fusion module dynamically allocates weights based on a three-dimensional attention mechanism.

4. The underground power distribution room communication system based on heterogeneous network and multimodal fusion according to claim 1, characterized in that, The state space monitoring unit collects signal strength, signal-to-noise ratio, and device power parameters in real time through multi-source heterogeneous data fusion sensing technology; the routing optimization unit generates routing strategies based on the STAM-MADDPG model. The routing optimization unit abstracts communication nodes as intelligent agents, implements cross-modal link collaborative scheduling based on the STAM-MADDPG model, and adopts a deep Q-network model containing a three-layer fully connected neural network to minimize latency and packet loss rate. It also accelerates routing strategy optimization with the help of an experience replay mechanism. The millisecond-level path switching strategy triggers switching within 15ms through dual thresholds.

5. The underground power distribution room communication system based on heterogeneous network and multimodal fusion according to claim 1, characterized in that, The signal subspace separation module separates 4G / 5G and power line carrier signals based on an improved singular value decomposition algorithm. This module also constructs a dynamic space-time block coded (STBC) matrix to reassemble the signal, ensuring a bit error rate of less than or equal to 10 even at a signal-to-noise ratio of -5dB. -3 The interference prediction and power control module generates adversarial network interference predictions and dynamically adjusts the transmission power. The cross-modal cooperative anti-interference mechanism is used to trigger NOMA technology to reuse frequency band resources or switch transmission links when the performance of either 4G / 5G or power line carrier is below a threshold, thereby extending the regulation strategy generated by the interference prediction and power control module to multi-modal cooperative scenarios.

6. A communication method for underground power distribution rooms based on heterogeneous networks and multimodal intelligent fusion, applied to an underground power distribution room communication system based on heterogeneous networks and multimodal fusion, characterized in that, The underground power distribution room communication system based on heterogeneous networks and multimodal fusion includes a three-dimensional heterogeneous network collaboration module, a cross-modal feature fusion module, a dynamic routing decision engine module, an adaptive interference suppression system, and a cross-layer collaborative optimization module. The three-dimensional heterogeneous network collaboration module includes a 4G / 5G wireless communication sublayer, a power line carrier communication sublayer, and an edge computing management sublayer. The power line carrier communication sublayer is connected to the 4G / 5G wireless communication sublayer and the edge computing management sublayer, respectively. The edge computing management sublayer achieves collaborative control through cross-layer signaling interaction; the cross-modal feature fusion module includes a multimodal encoder group, a shared feature projection layer, a Siamese network structure, and a cross-modal attention fusion module. The shared feature projection layer is connected to the multimodal encoder group and the Siamese network structure, respectively. The Siamese network structure is also connected to the cross-modal attention fusion module. The multimodal encoder group is also connected to the edge computing management sublayer. The cross-modal feature fusion module extracts and fuses multimodal data features based on an improved Siamese neural network. The dynamic routing decision engine module includes a state space monitoring unit, a routing optimization unit, and a millisecond-level path switching strategy. The state space monitoring unit is connected to the routing optimization unit and the cross-modal attention fusion module. The routing optimization unit is also connected to the millisecond-level path switching strategy. The routing optimization unit uses a STAM-MADDPG model and a deep Q-network model to optimize the routing strategy in real time. The adaptive interference suppression system includes a signal subspace separation module, a cross-modal collaborative anti-interference mechanism, and an interference prediction and power control module. The interference prediction and power control module is connected to the signal subspace separation module, the cross-modal collaborative anti-interference mechanism, and the interference prediction and power control module. The cross-modal collaborative anti-interference mechanism is connected to the signal subspace separation module and the millisecond-level path switching strategy, respectively. The signal subspace separation module, the cross-modal collaborative anti-interference mechanism, and the interference prediction and power control module combine NOMA-STBC joint anti-interference technology, generative adversarial network interference prediction, and the cross-modal collaborative mechanism to suppress complex electromagnetic interference. The cross-layer collaborative optimization module is connected to the edge computing management sublayer and the state space monitoring unit, respectively. The cross-layer collaborative optimization module balances transmission rate, bit error rate, and energy consumption based on non-dominated sorting genetic algorithm III. The method includes: Cross-layer data preprocessing: In the 4G / 5G wireless communication sublayer, spatial coordinate mapping of 4G / 5G baseband signals is performed through BeiDou / GNSS dual-system positioning. In the power line carrier communication sublayer, the power line carrier signal is clock-synchronized to generate a unified data frame containing timestamps and device IDs; In the edge computing management sublayer, two types of signal data are integrated and encapsulated into standardized data frames according to dynamic slicing requirements and channel detection requirements for subsequent processing; Feature fusion coding: In the multimodal encoder group, infrared images are processed by an improved ResNet50+CBAM architecture to compress brightness to the [0.2, 0.8] range; SCADA parameters are processed by dual-channel LSTM+time window attention; and abnormal noise features are extracted by Mel spectrum-wavelet transform. In the shared feature projection layer, features are mapped through visual, text, and audio projection matrices, and weights are dynamically adjusted according to the environment. In the twin network structure, multi-scale features are captured through dilated convolution, embedded in the device topology graph, and comparative learning is performed based on the improved InfoNCE loss function. In the cross-modal attention fusion module, spatial attention is used to focus on key parts of the device, temporal attention is used to capture data 5 seconds before the fault occurs, modal attention is used to dynamically allocate weights, and fused features are output. Dynamic route calculation: The state space monitoring unit collects signal strength, signal-to-noise ratio, and device power in real time, and constructs a state vector by combining the space attenuation factor, time penalty, and interference threshold. In the routing optimization unit, the STAM-MADDPG model is used to coordinate the scheduling of 4G / 5G and power line carrier links, and the deep Q network model generates routing strategies with the goal of minimizing latency and packet loss rate. In millisecond-level path switching strategies, when the latency is ≥30ms or the packet loss rate is ≥10... -4 When the pre-switching buffer is triggered, the path reselection is completed within 15ms. Real-time commands are transmitted via 4G / 5G, while historical data is transmitted via power line carrier. Interference adaptive processing: In the signal subspace separation module, 4G / 5G and power line carrier signals are separated according to the difference in singular values ​​based on the improved singular value decomposition algorithm, and a dynamic STBC matrix is ​​constructed to resist multipath fading. In the interference prediction and power control module, the interference distribution is predicted by a generative adversarial network model, the transmit power is adjusted in 2dBm steps, and the LoRa spreading factor is dynamically switched according to the interference intensity. In the cross-modal cooperative anti-interference mechanism, when the bit error rate of a certain link is >10 -4 If the signal-to-noise ratio is less than the threshold, NOMA technology is activated to reuse frequency band resources, or the connection is switched to a backup link. Cross-layer collaborative optimization: Design a weighted optimization function to dynamically adjust the weights; allocate the proportion of 4G / 5G and power line carrier transmission based on non-dominated sorting genetic algorithm III; when the 4G / 5G load is ≥70%, divert 30% of non-real-time services to power line carrier and improve spectrum efficiency through NOMA technology. For device-level power consumption control, the transmission power is reduced by 50% when the battery level is below 20%, and the sleep cycle is extended to 500ms; For network-level power consumption control, when the server CPU utilization is greater than or equal to 80%, the server CPU utilization is reduced by 0.3GHz for every 10% increase, and tasks are migrated to low-load servers. The indicators are guaranteed through a constraint processing mechanism, and the solution set is updated using an elite retention strategy guided by reference points.

7. The underground power distribution room communication system based on heterogeneous network and multimodal fusion according to claim 6, characterized in that, The formula for calculating the attention weight in the modal attention dynamic allocation weight is as follows: Among them, e t =score(h t ,q) represents the attention score, h t Let α represent the hidden state of the LSTM at time t, q represent the query vector, and α represent the hidden state of the LSTM at time t. t The attention weight at time t is used to weight and sum the data at different times to obtain the feature representation; β t This represents the spatiotemporal coupling coefficient, reflecting the differences in the importance of data at different times, and is dynamically adjusted based on the dynamic changing trends of the power system; w m Represents modal interaction weights, measuring the degree of correlation between multimodal data, and is dynamically learned through a cross-modal attention mechanism; ρ t,m This represents the equipment topology constraint terms, introduces a power equipment topology relationship graph, and quantifies the correlation between the current data and the fault propagation path; This represents modal feature enhancement terms, which enhance the characteristics of different modal data. Wherein, s(x i ,x j ) represents sample x i and x j The similarity score; τ represents the temperature parameter, used to control the difficulty of contrastive learning, by minimizing the InfoNCE loss function, bringing positive sample pairs closer and pushing negative sample pairs further apart; φ i,i+ The weights represent the topological constraints of the devices. Based on the power equipment topology graph, the topological correlation between device nodes in a positive sample pair is calculated using a graph neural network. i,j ε represents the negative sample association penalty term, which incorporates prior knowledge in the power field to introduce a dynamic penalty mechanism for negative sample pairs; ε represents the smoothing factor to ensure the stability of the loss function calculation.

8. The communication method for underground power distribution rooms based on heterogeneous networks and multimodal intelligent fusion according to claim 6, characterized in that, After the state space monitoring unit collects signal strength, signal-to-noise ratio, and device power in real time, and constructs a state vector by combining spatial attenuation factor, time penalty, and interference threshold, the process includes: The attention weight α is dynamically adjusted by a spatiotemporal attention mechanism, incorporating multimodal information fusion, dynamic environment perception, and network topology constraints. i The calculation formula is as follows: Where, f(s) i ) represents the attention scoring function, which obtains a score by calculating the correlation between each parameter and the current routing decision; α i Used for weighted aggregation of state vectors, prioritizing the capture of key parameters affecting routing decisions; γ t,i b represents the time-based dynamic adjustment coefficient, which dynamically adjusts the attention weights of parameters at different times based on historical latency fluctuations in the communication link and the urgency of real-time transmission tasks; β m,i The modality fusion weights represent the multimodal features of 4G / 5G and power line carrier communication, dynamically allocating weights based on link type and interference status; g(G) represents the network topology constraint term, introducing the communication network topology graph G, and calculating the structural relationships between nodes through a graph convolutional network; δ n,i This represents the topology influence coefficient, which dynamically adjusts the influence of topology constraints based on the role of a node in the network topology.

9. The underground power distribution room communication system based on heterogeneous network and multimodal fusion according to claim 6, characterized in that, The method of coordinating 4G / 5G and power line carrier links through the STAM-MADDPG model, and generating routing strategies with the deep Q network model aiming to minimize latency and packet loss rate, includes: The Deep Q-Network model learns the optimal routing strategy through an improved Deep Q-Network loss function, balancing immediate rewards and long-term benefits. The formula for calculating the improved Deep Q-Network loss function is as follows: Among them, y delay =r delay +γ delay max a′ Q delay (S′,a';θ - ),y packet =r packet +γ packet max a' Q packet (S',a';θ - ); r represents the immediate reward, a positive reward for successful transmission and a negative reward for packet loss; γ represents the discount factor, ranging from 0 to 1, balancing immediate and long-term benefits; θ represents the current network parameters; θ - The target network parameters are represented and updated periodically to stabilize training. The (s,a,r,s′) are stored in the experience pool through the experience replay mechanism and randomly sampled for training.

10. The underground power distribution room communication system based on heterogeneous network and multimodal fusion according to claim 6, characterized in that, The interference adaptive processing also includes: By leveraging generative adversarial networks to fuse historical interference data with real-time environmental parameters, the generator and discriminator learn adversarially to predict interference distribution. The generator attempts to produce realistic interference samples. The discriminator distinguishes between real interference and generated samples. The objective function of the generative adversarial network model is as follows: Where, max D This indicates that the discriminator D attempts to maximize its own discriminative power, min G Let V(D,G) represent the generator G's attempt to minimize the discriminative power of the discriminator, and let V(D,G) represent the optimization objective function of the generative adversarial network. Let logD(I) represent the expectation of the real disturbance data, where logD(I) follows the real data distribution P. data logD(I) represents the logarithmic probability of the discriminator D classifying the true sample I. This represents the expression for the noise vector log(1-D(G(P)). real The expectation of log(1-D(G(P)) real )) Follows the prior distribution P noise ;log(1-D(G(P) real )) indicates that the discriminator D evaluates the generated sample G(P) real The logarithm of the probability complement of the judgment, P real This represents real-time environmental parameters.

Images