Neural network-based underground pipe gallery environment adaptive computing system

Abstract

This invention relates to the field of computer system technology and discloses an adaptive computing system for underground utility tunnels based on neural networks. The system includes an environmental feature parameter extraction unit, a topological correlation matrix calculation unit, an adaptive topological reconstruction unit, and a computational flow control unit. The system extracts the spatiotemporal dynamic feature tensor of the utility tunnel and calculates the logical topological correlation matrix by combining it with the accumulated variables of the preceding topological residuals. When the operator deviation crosses the critical variation threshold, the computation graph is decoupled in situ and a dense computational sub-network is attached. Simultaneously, the threshold is calibrated according to the reconstruction trigger frequency. Based on this, the underlying hardware addressing path is configured. This invention utilizes the in-situ topological variation of the computation graph and the threshold negative feedback mechanism to eliminate operator transmission resistance under non-stationary operating conditions and effectively suppress structural reconstruction oscillations caused by high-frequency noise.

Description

Technical Field

[0001] This invention belongs to the field of computer system technology, and in particular relates to an adaptive computing system for underground utility tunnels based on neural networks. Background Technology

[0002] Currently, fixed-topology neural network models are typically used to process sensor telemetry data streams. These static arrays allocate memory space and instruction pipelines according to preset operator connection logic, maintaining deterministic throughput and computational power consumption under stable temporal input characteristics. However, when the data stream represents complex non-stationary multi-field coupling dissipation processes or transient variability conditions, the dimension of the input feature tensor and the nonlinear spatiotemporal coupling degree drift drastically. The computing system requires network operators to exhibit synchronous dense response capabilities to maintain the continuity of the instruction stream. For transient conditions composed of multi-source heterogeneous environmental parameters within a confined space, traditional statically configured network structures exhibit computational mismatch defects. The fixed operator links cannot quickly adapt to the transiently changing data stream tensors, causing operator logic transmission resistance in the computation graph within the general-purpose processing chip. Due to the lack of topology self-evolution capability, a large number of insensitive sparse operator links continuously occupy physical storage space and cache bandwidth, causing inference instructions to queue and block in the pipeline, wasting chip hardware computing power, accumulating system inference timing delays, and bringing potential risks.

[0003] Not only does the physical architecture of utility tunnel environmental monitoring face limitations in the deployment space of monitoring components and the dulling of data acquisition functions, but the backend software control methods and data processing modes also suffer from insufficient adaptability. For example, Chinese invention patent application CN111612018A discloses a method and system for monitoring underground utility tunnels based on multi-source heterogeneous data fusion. It constructs a two-level deep model, using a first deep learning network to identify unstructured data and combining structured features with the input of a second deep learning network for comprehensive identification. However, the underlying layer still relies on a preset fixed network topology. Faced with non-stationary and variable conditions such as sudden water seepage or harmful gas leakage in the utility tunnel, the static cascaded architecture cannot decouple the computation graph in situ within the memory address space, causing a mismatch between the input feature dimension and the computing power link, triggering a deadlock in the inference instruction pipeline. Furthermore, it lacks a threshold dynamic negative feedback calibration and latency offsetting mechanism, and cannot smooth out oscillations caused by topology reconstruction or state switching under high-frequency noise, resulting in hardware register addressing conflicts and control signal phase distortion, failing to guarantee the end-to-end real-time high-precision timing control requirements of less than 12ms.

[0004] Therefore, how to enable the network computation graph to adaptively decouple based on feature variations and achieve high-precision time-series calibration while suppressing high-frequency reconstruction oscillations has become the technical problem to be solved by this invention. Summary of the Invention

[0005] To address the problems in the background technology, this technical solution provides an adaptive calculation system for underground utility tunnel environments based on neural networks, comprising:

[0006] The environmental feature parameter extraction unit acquires the original dataset of the multidimensional heterogeneous environment of the underground utility tunnel and extracts the spatiotemporal dynamic feature tensor.

[0007] The topological correlation matrix calculation unit calculates the logical topological correlation matrix between each operator node by superimposing the cumulative variables of the topological residuals of the preceding state section on the spatiotemporal dynamic feature tensor.

[0008] The adaptive topology reconstruction unit, when the deviation value of adjacent operator nodes in the logical topology correlation matrix crosses the critical mutation threshold, decouples the neural network computation graph in situ to cut off low-computing-power contribution links, and attaches a dense computational sub-network that matches the mutation trend of the spatiotemporal dynamic feature tensor to output the optimal computation graph architecture instruction; wherein, the embedded dynamic hysteresis damping module counts the number of reconstruction events of the neural network structure computation graph and calculates the reconstruction trigger frequency, and activates the threshold dynamic calibration program to increase the current critical mutation threshold when the reconstruction trigger frequency crosses the 45Hz frequency safety upper limit;

[0009] The computation flow control unit configures the hardware addressing path of the underlying general-purpose computing chip according to the optimal computation graph architecture instructions in order to output the environmental control compensation strategy.

[0010] Preferably, the adaptive topology reconstruction unit is further cascaded with a multi-scale state reprojection subsystem; the multi-scale state reprojection subsystem is used to perform matrix projection of the old topology boundary residual information before decoupling onto a low-dimensional orthogonal subspace to generate a state-preserving vector during the reconstruction transient window of the neural network computation graph; and when the input feature dimension of the dense computation subnetwork is inconsistent with the tensor dimension of the state-preserving vector, zero eigenvalues ​​are appended to the end of the state-preserving vector to complete the tensor dimension alignment, and the initial weight bias of the dense computation subnetwork is rewritten using the dimension-aligned state-preserving vector.

[0011] Preferably, the computing flow control unit further includes a heterogeneous operator interaction gap compensation subsystem, which is used to quantitatively obtain the memory addressing delay jitter caused by bus memory delay during instruction switching, and calculate a delay compensation parameter of 6.0 μs by linear product based on a delay mapping factor of 0.4 when the measured memory addressing delay jitter is 15 μs, and dynamically increase the inter-operator interaction gap of the underlying general computing chip by 6.0 μs.

[0012] Preferably, the environmental feature parameter extraction unit includes a heterogeneous data parallel acquisition module, a spatial noise reduction and conditioning module, and a temporal feature fusion module. The heterogeneous data parallel acquisition module is used to read data on harmful gas concentration, temperature and humidity, cable surface temperature, and water level from external sensors in parallel to construct a multidimensional heterogeneous environmental raw dataset. The spatial noise reduction and conditioning module is cascaded with the heterogeneous data parallel acquisition module and is used to apply a spatial adjacency matrix to the multidimensional heterogeneous environmental raw dataset for weighted filtering to generate a spatial denoising feature matrix. The temporal feature fusion module receives the spatial denoising feature matrix output by the spatial noise reduction and conditioning module and is used to input the spatial denoising feature matrix into a long short-term memory network along the time axis to extract spatiotemporal sequence associations and output a spatiotemporal dynamic feature tensor.

[0013] Preferably, the topological correlation matrix calculation unit includes a residual accumulation feedback loop and an operator correlation mapping module. The residual accumulation feedback loop is used to read the control strategy deviation value of the neural network model in the underground utility tunnel environment reasoning at the current time step, and accumulate it into the topological residual accumulation variable of the preceding state section using the first-order time discrete integral. The operator correlation mapping module is logically associated with the residual accumulation feedback loop and is used to concatenate and spatiotemporally concatenate the spatiotemporal dynamic feature tensor with the topological residual accumulation variable of the preceding state section, and input the concatenated tensor into the topological mapping dense layer, and output the logical correlation weights between each pair of operator nodes in the neural network model to form a logical topological correlation matrix.

[0014] Preferably, the adaptive topology reconstruction unit further includes a graph topology in-situ reconstruction center; when the deviation value of adjacent operator nodes in the logical topology correlation matrix crosses the critical mutation threshold, the graph topology in-situ reconstruction center decouples the neural network computation graph in-situ within the memory addressing space to disconnect the computing power link, and retrieves and dynamically attaches a dense computation subnetwork corresponding to the spatiotemporal dynamic feature tensor mutation direction from a preset heterogeneous dense computation subnetwork library.

[0015] Preferably, the environmental control compensation strategy output by the computational flow control unit is a pure digital control command stream, which includes dynamic speed regulation commands for fans, start / stop switching commands for dehumidifiers, dynamic scheduling commands for cable loads, and adjustment commands for drainage pumps. These commands are used to change the operating status variable values ​​of the corresponding external hardware within the underground utility tunnel environment.

[0016] Preferably, the adaptive topology reconstruction unit further includes a state degradation monitoring module; the state degradation monitoring module is used to perform sliding window statistics on the trigger frequency of the threshold dynamic calibration procedure per unit time, calculate the threshold calibration drift variation rate, and when the threshold calibration drift variation rate is greater than 2.5 Hz / s for three consecutive sliding windows, output an alarm signal indicating that an abnormal change has occurred in the calculation of the underground utility tunnel neural network.

[0017] Preferably, the control clock step size of the environmental feature parameter extraction unit, the topology correlation matrix calculation unit, the adaptive topology reconstruction unit, and the computational flow control unit is all locked at 50ms, and the data throughput rate of the original dataset of the multidimensional heterogeneous environment is not less than 20MB / s, so as to ensure that the end-to-end computation latency of obtaining the environmental control compensation strategy output from the original dataset of the multidimensional heterogeneous environment is less than 12ms.

[0018] Compared with existing technologies, the neural network-based adaptive calculation system for underground utility tunnel environments of this invention has the following advantages:

[0019] 1. In the adaptive computation of underground utility tunnel environment in neural network, the environmental feature parameter extraction unit collects multi-dimensional heterogeneous original telemetry datasets and extracts spatiotemporal dynamic feature tensors. The topology correlation matrix calculation unit concurrently retrieves the feature tensors and the preceding topology residual accumulation terms stored in the register. Through nonlinear cumulative effect, it smooths the monotonic mapping relationship of the operator weight matrices at each level to output the logical topology correlation matrix. When the deviation value of adjacent operator nodes crosses the critical mutation threshold, it controls the in-situ topology decoupling computation graph of the memory addressing space and cuts off the low contribution operator links. It dynamically mounts a dense computation sub-network that matches the mutation direction of the feature tensor to eliminate the transmission resistance of operators inside the network graph when there is non-stationary input.

[0020] 2. The dynamic hysteresis damping module embedded in the adaptive topology reconfiguration unit counts the number of reconfiguration events in the structural computation graph in discrete time increments to calculate the reconfiguration frequency. When the reconfiguration frequency exceeds the preset frequency safety limit, the threshold dynamic calibration program is activated. Based on the cascaded correction product relationship, the calibrated basic threshold constant is combined with the frequency increment term containing the dynamic adjustment coefficient to dynamically increase the current critical variation threshold to 1.30 to raise the variation threshold and forcibly extend the time lock protection period between adjacent reconfiguration actions, thus smoothing out the frequent reconfiguration oscillations of the network computation graph topology caused by high-frequency mutation noise.

[0021] 3. The multi-scale state reprojection subsystem of the adaptive topology reconstruction unit cascade projects the residual information of the old topology boundary before decoupling onto the low-dimensional orthogonal subspace during the transient window of network computation graph topology reconstruction to generate a state-preserving vector and extract and retain long-period trend features. When the input feature dimension of the dense computation subnetwork is inconsistent with the 96-dimensional state-preserving vector, the dimension-upgrading zero-filling module is called to append 32 zero feature values ​​to the end of the state-preserving vector to complete the tensor dimension alignment. The initial weight bias of the dense computation subnetwork is rewritten using the dimension-aligned state-preserving vector to eliminate the logical state discontinuity in the process of alternating between old and new operators. Attached Figure Description

[0022] Figure 1 This is a flowchart of the adaptive calculation process for underground utility tunnel environmental data according to the present invention;

[0023] Figure 2 This is a block diagram of the module structure of the environmental adaptive computing system of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0025] An adaptive computational system for underground utility tunnel environments based on neural networks, comprising:

[0026] The environmental feature parameter extraction unit acquires the original dataset of the multidimensional heterogeneous environment of the underground utility tunnel and extracts the spatiotemporal dynamic feature tensor.

[0027] The topological correlation matrix calculation unit calculates the logical topological correlation matrix between each operator node by superimposing the cumulative variables of the topological residuals of the preceding state section on the spatiotemporal dynamic feature tensor.

[0028] The adaptive topology reconstruction unit, when the deviation value of adjacent operator nodes in the logical topology correlation matrix crosses the critical mutation threshold, decouples the neural network computation graph in situ to cut off low-computing-power contribution links, and attaches a dense computational sub-network that matches the mutation trend of the spatiotemporal dynamic feature tensor to output the optimal computation graph architecture instruction; wherein, the embedded dynamic hysteresis damping module counts the number of reconstruction events of the neural network structure computation graph and calculates the reconstruction trigger frequency, and activates the threshold dynamic calibration program to increase the current critical mutation threshold when the reconstruction trigger frequency crosses the 45Hz frequency safety upper limit;

[0029] The computation flow control unit configures the hardware addressing path of the underlying general-purpose computing chip according to the optimal computation graph architecture instructions in order to output the environmental control compensation strategy.

[0030] Preferably, the adaptive topology reconstruction unit is further cascaded with a multi-scale state reprojection subsystem; the multi-scale state reprojection subsystem is used to perform matrix projection of the old topology boundary residual information before decoupling onto a low-dimensional orthogonal subspace to generate a state-preserving vector during the reconstruction transient window of the neural network computation graph; and when the input feature dimension of the dense computation subnetwork is inconsistent with the tensor dimension of the state-preserving vector, zero eigenvalues ​​are appended to the end of the state-preserving vector to complete the tensor dimension alignment, and the initial weight bias of the dense computation subnetwork is rewritten using the dimension-aligned state-preserving vector.

[0031] Preferably, the computing flow control unit further includes a heterogeneous operator interaction gap compensation subsystem, which is used to quantitatively obtain the memory addressing delay jitter caused by bus memory delay during instruction switching, and calculate a delay compensation parameter of 6.0 μs by linear product based on a delay mapping factor of 0.4 when the measured memory addressing delay jitter is 15 μs, and dynamically increase the inter-operator interaction gap of the underlying general computing chip by 6.0 μs.

[0032] Preferably, the environmental feature parameter extraction unit includes a heterogeneous data parallel acquisition module, a spatial noise reduction and conditioning module, and a temporal feature fusion module. The heterogeneous data parallel acquisition module is used to read data on harmful gas concentration, temperature and humidity, cable surface temperature, and water level from external sensors in parallel to construct a multidimensional heterogeneous environmental raw dataset. The spatial noise reduction and conditioning module is cascaded with the heterogeneous data parallel acquisition module and is used to apply a spatial adjacency matrix to the multidimensional heterogeneous environmental raw dataset for weighted filtering to generate a spatial denoising feature matrix. The temporal feature fusion module receives the spatial denoising feature matrix output by the spatial noise reduction and conditioning module and is used to input the spatial denoising feature matrix into a long short-term memory network along the time axis to extract spatiotemporal sequence associations and output a spatiotemporal dynamic feature tensor.

[0033] Preferably, the topological correlation matrix calculation unit includes a residual accumulation feedback loop and an operator correlation mapping module. The residual accumulation feedback loop is used to read the control strategy deviation value of the neural network model in the underground utility tunnel environment reasoning at the current time step, and accumulate it into the topological residual accumulation variable of the preceding state section using the first-order time discrete integral. The operator correlation mapping module is logically associated with the residual accumulation feedback loop and is used to concatenate and spatiotemporally concatenate the spatiotemporal dynamic feature tensor with the topological residual accumulation variable of the preceding state section, and input the concatenated tensor into the topological mapping dense layer, and output the logical correlation weights between each pair of operator nodes in the neural network model to form a logical topological correlation matrix.

[0034] Preferably, the adaptive topology reconstruction unit further includes a graph topology in-situ reconstruction center; when the deviation value of adjacent operator nodes in the logical topology correlation matrix crosses the critical mutation threshold, the graph topology in-situ reconstruction center decouples the neural network computation graph in-situ within the memory addressing space to disconnect the computing power link, and retrieves and dynamically attaches a dense computation subnetwork corresponding to the spatiotemporal dynamic feature tensor mutation direction from a preset heterogeneous dense computation subnetwork library.

[0035] Preferably, the environmental control compensation strategy output by the computational flow control unit is a pure digital control command stream, which includes dynamic speed regulation commands for fans, start / stop switching commands for dehumidifiers, dynamic scheduling commands for cable loads, and adjustment commands for drainage pumps. These commands are used to change the operating status variable values ​​of the corresponding external hardware within the underground utility tunnel environment.

[0036] Preferably, the adaptive topology reconstruction unit further includes a state degradation monitoring module; the state degradation monitoring module is used to perform sliding window statistics on the trigger frequency of the threshold dynamic calibration procedure per unit time, calculate the threshold calibration drift variation rate, and when the threshold calibration drift variation rate is greater than 2.5 Hz / s for three consecutive sliding windows, output an alarm signal indicating that an abnormal change has occurred in the calculation of the underground utility tunnel neural network.

[0037] Preferably, the control clock step size of the environmental feature parameter extraction unit, the topology correlation matrix calculation unit, the adaptive topology reconstruction unit, and the computational flow control unit is all locked at 50ms, and the data throughput rate of the original dataset of the multidimensional heterogeneous environment is not less than 20MB / s, so as to ensure that the end-to-end computation latency of obtaining the environmental control compensation strategy output from the original dataset of the multidimensional heterogeneous environment is less than 12ms.

[0038] Example 1: In a closed underground utility tunnel operating environment with a combination of sudden underground water seepage and hazardous gas leakage, a multidimensional heterogeneous environmental dataset is constructed using hazardous gas concentration, temperature and humidity, cable surface temperature, and accumulated water level data collected by gas sensors, temperature and humidity sensors, infrared thermometers, and water level sensors. When this dataset exhibits a non-stationary multi-field coupling dissipation state or transient variation condition over time, the dimension of the input feature tensor and the nonlinear spatiotemporal coupling degree drift. The fixed operator link cannot quickly adapt to the instantaneously changing data flow tensor, resulting in deadlock in the inference instruction pipeline and operator propagation resistance within the chip, and also causing branch conditional logic switching models. When faced with high-frequency oscillation noise, continuous switching is triggered, leading to cache invalidation and address bounce, which damages the timing phase of the control signal, causing end-to-end inference delay accumulation and deviation of the timing phase of the control signal from the preset safe range. The heterogeneous data parallel acquisition module in the environmental feature parameter extraction unit collects multi-dimensional heterogeneous environment raw datasets from external sensors with a control clock step of 50ms. The spatial domain noise reduction and conditioning module uses a spatial adjacency matrix to weight and filter the multi-dimensional heterogeneous environment raw dataset to generate a spatial denoising feature matrix. The temporal feature fusion module inputs the spatial denoising feature matrix into the long short-term memory network along the time axis to extract spatiotemporal sequence correlations and output spatiotemporal dynamic feature tensors. Correspondingly, the residual accumulation feedback loop in the topology correlation matrix calculation unit reads the control strategy deviation value of the neural network model in the underground utility tunnel environment inference at the previous time step, and uses first-order time discrete integration to accumulate the control strategy deviation value as the topology residual accumulation variable of the preceding state section. In this process, to eliminate the problem of non-closed-loop calculations caused by different physical dimensions and heterogeneous characteristics between the continuous numerical values ​​of fan speed regulation, the Boolean logic of dehumidifier start-stop, and the discharge rate adjustment of drainage pump, the residual accumulation feedback loop, after reading the control strategy deviation value, calls the dimensionless normalization operator to divide the original hardware drive deviation value of each different dimension by its corresponding rated engineering design upper limit, uniformly transforming it into a dimensionless scaling coefficient between -1.0 and 1.0, constructing a standardized generalized control error vector, and then performing a first-order time discrete integral on the error vector to generate a topological residual accumulation variable, thereby ensuring complete equivalence and self-consistency of subsequent cascade splicing operations at the mathematical structure and physical logic levels. The operator association mapping module converts the spatiotemporal dynamic feature tensor Cumulative variables of topological residuals from preceding state sections The concatenated tensors are then input into a topological mapping dense layer to calculate the pairwise logical association weights between each operator node in the neural network model, forming and outputting a logical topological association matrix. ,in For spatiotemporal dynamic feature tensors, For topological residual cumulative variables, Given the logical topological correlation matrix, the adaptive topological reconstruction unit extracts a local submatrix representing the connection weights of adjacent operator nodes in the matrix. It then calculates the absolute difference matrix between the corresponding matrix elements of the current time step and the previous time step, and obtains the Frobenius norm of this absolute difference matrix. This compresses and maps the multi-dimensional topological weight drift trend into a scalar value representing the degree of instantaneous structural anomaly, which is the adjacent operator node deviation value used as the basis for judgment.

[0039] When the logical topological correlation matrix When the deviation value of adjacent operator nodes in the adaptive topology reconstruction unit crosses the critical variation threshold, the graph topology in-situ reconstruction center in the adaptive topology reconstruction unit decouples the neural network computation graph in-situ within the memory addressing space and disconnects operator links whose computational power contribution is lower than the preset benchmark. It then retrieves and dynamically mounts spatiotemporal dynamic feature tensors from the preset heterogeneous dense computation sub-network library. A dense computational subnetwork matching the mutation trend outputs the optimal computational graph architecture instructions, and the embedded dynamic hysteresis damping module counts the number of reconstruction events in the neural network structure computational graph and calculates the reconstruction trigger frequency. When the reconstruction trigger frequency When the frequency is greater than 45Hz, the threshold dynamic calibration procedure is initiated according to the formula. Increase the current critical mutation threshold, where, This is the corrected critical mutation threshold. Based on the basic threshold constant, This is a dynamic adjustment coefficient. To reduce the reconfiguration trigger frequency, the time-lock protection period between adjacent reconfiguration actions is extended, thus reducing the reconfiguration oscillation frequency of the network computation graph topology. The 45Hz frequency safety limit and the 1.30 critical variation threshold values ​​are derived from the pre-calibration process based on the discrete control system sampling theorem and the processor thermal power boundary theory. Before system operation, the pre-calibration process measures the processor's temperature drift curve under continuous reconfiguration conditions using a temperature sensor. It is found that when the reconfiguration trigger frequency reaches 45Hz, the transient power consumption caused by physical addressing conflicts leads to the microprocessor core temperature reaching the 85℃ physical safety limit. The system is then calibrated experimentally to adjust the basic threshold constant V. base The dynamic adjustment coefficient α, used to control the mutation threshold, is set to 0.65 and calibrated to 0.02 within the engineering-restricted range of 0.01 to 0.05. When the dynamic hysteresis damping module detects that the reconstruction trigger frequency Ft reaches 45Hz, the threshold dynamic calibration program is initiated. It uses the base threshold constant 0.65 multiplied by the incremental term resulting from the linear combination of the dynamic adjustment coefficient 0.02 and the reconstruction trigger frequency 45Hz, progressively increasing the current critical mutation threshold Vt to 1.30. This raises the mutation threshold of subsequent operators and extends the time-locking protection period between adjacent reconstruction actions. The optimal computational graph architecture instruction includes operator dynamic linked lists. The reference table and the corresponding memory starting address offset are used to complete the downlink conversion closed loop from algorithm to hardware action. The computation flow control unit has an instruction parsing module, which is used to convert the high-level abstract optimal computation graph architecture instructions into specific physical control constraint actions. Based on the microprocessor bus addressing principle, the instruction parsing module directly modifies the dynamic routing pointer in the global operator control register inside the underlying general-purpose computing chip, cuts off the low computing power contribution link by setting the hardware enable position of invalid operator nodes to zero, and writes the compilation base address of the newly mounted dense computing sub-network into the chip program counter. The hardware micro-instruction level completes the addressing path reconstruction configuration.

[0040] During the transient window of neural network computation graph reconstruction, the multi-scale state reprojection subsystem of the adaptive topology reconstruction unit cascade projects the residual information of the old topology boundary before decoupling onto a low-dimensional orthogonal subspace to generate a state-preserving vector. When the input feature dimension of the dense computation subnetwork is inconsistent with the tensor dimension of the state-preserving vector, the dimension-upgrading zero-padding module is called to append zero eigenvalues ​​to the end of the state-preserving vector to complete tensor dimension alignment. The initial weight bias of the dense computation subnetwork is rewritten using the dimension-aligned state-preserving vector to maintain the state continuity during the alternation of old and new operators. In order to safely complete this state rewriting and subnetwork mounting within an extremely short clock step of less than 12 milliseconds, this invention implements a bottom-level... The chip's local static random access memory (SRAM) utilizes a double-buffered static memory pool. The front buffer maintains active addressing of the currently running computation graph, while the back buffer remains resident and ready in the background. Upon receiving a reconstruction instruction, the graph topology in-situ reconstruction center does not reallocate physical memory space. Instead, through an atomic pointer switch, within a compilation window of less than 0.5 milliseconds, it directly resets the register image base address of the instruction pipeline from the front buffer to the back buffer and directly overwrites the aligned state-keeping vector in the cache. This significantly eliminates bus transport overhead, thereby ensuring that the overall end-to-end inference latency remains stable at the underlying level of the computer architecture. To overcome the 12-millisecond physical limit and bridge the order-of-magnitude physical difference between the microsecond-level reconstruction transient window and the long compilation cycle of network operators, the system incorporates a physical compensation mechanism based on a double-buffered static memory pool. The front buffer maintains active addressing, while the back buffer remains resident and ready in the background. Spatial isolation and alternating material paths with double buffers offset temporal conflicts. Because the original dataset in the multidimensional heterogeneous environment suffers from strong coupling interference from multiple variables caused by external sensor aging, electromagnetic interference, and baseline data drift during long-term operation, the system incorporates an interference removal procedure to achieve true decoupling of physical variables. This interference removal procedure utilizes a spatial domain noise reduction conditioning module to process the original data through a spatial adjacency matrix. The concentrated data on harmful gas concentration and water level are processed by weighted low-pass filtering. Simultaneously, the 50Hz power frequency interference harmonic voltage fluctuation variable caused by the high-voltage cable frequency converter is collected. The interference elimination procedure uses the measured power frequency interference voltage fluctuation residual as a negative correction factor. Differential subtraction is performed at the current time step input to remove background electromagnetic noise interference from the total measured voltage, completing the logical decoupling of the physical characteristics of harmful gases from the background noise variable. The computational flow control unit receives the optimal computational graph architecture instructions to configure the hardware addressing path of the underlying general-purpose computing chip. The internal heterogeneous operator interaction gap compensation subsystem obtains the storage addressing delay jitter caused by bus storage latency during instruction switching. When the memory addressing delay jitter is measured When the delay is 15μs, it is based on the preset time delay mapping factor. Using the linear product formula Calculate the delay compensation parameters ,in, For delay compensation parameters, For time delay mapping factor, To store addressing latency jitter, the inter-operator interaction gap latching time of the underlying general-purpose computing chip is increased by 6.0 μs. An environmental control compensation strategy is implemented, outputting a pure digital control instruction stream containing commands for dynamic fan speed regulation, dehumidifier start / stop switching, cable load dynamic scheduling, and drainage pump adjustment. The end-to-end computation latency is less than 12 ms, and the operating state variable values ​​of the underlying general-purpose computing chip remain within the preset control phase range. Specifically, at the microarchitecture control level, the heterogeneous operator interaction gap compensation subsystem converts the calculated 6.0 microsecond delay compensation parameter into the corresponding clock cycle number based on the current core operating frequency of the general-purpose computing chip. The execution unit directly writes this cycle count into the pipeline pause control register inside the chip. By forcibly inserting a latching wait cycle consisting of no-operation instructions at the instruction extraction boundary between two alternating operators, bus addressing conflicts are forcibly blocked at the micro-instruction level at the hardware level, thereby eliminating timing phase lag. To control the collaborative management and resource allocation of multi-segment underground utility tunnel clusters, the system configures collaborative management and resource allocation procedures. When the test group using the method of this invention is applied to a clustered deployment environment containing M independent pipe gallery branch sections, the top-level environmental control center obtains the total global environmental control allocation task. The collaborative management and resource allocation procedure adopts the optimal allocation strategy based on the current state of each section. It reads the threshold calibration drift variation rate output by the state degradation monitoring module corresponding to each pipe gallery branch section, and uses the reciprocal of this rate as the allocation weight coefficient for each unit. The top-level environmental control center calculates the linear product of the total global environmental control allocation task and the allocation weight coefficient and distributes it. It prioritizes distributing high-load control strategy instructions to pipe gallery branch sections with lower threshold calibration drift variation rates and more stable calculation states. In large-scale application scenarios, it achieves a balanced distribution of global computing resources and control load. The pure digital control command stream output by the environmental control compensation strategy constructs a complete value closed loop from data, information, insight to action at the application layer. The fan dynamic speed regulation command and drainage pump adjustment command included in the pure digital control command stream are not a simple presentation of the original telemetry data, but are the delay compensation parameters C calculated end-to-end by the computing flow control unit. pBy associating with the hardware addressing state, it directly converts it into the specific frequency control word of the corresponding external fan drive inverter and the voltage duty cycle digital signal of the drainage pump proportional control valve. By directly changing the hardware action of the physical actuator, the safety indicators of the utility tunnel are smoothly restored. When the system executes the environmental control compensation strategy, the hardware addressing path is switched according to the optimal computational graph architecture instruction. When the underlying general-purpose computing chip processes the original dataset of the multi-dimensional heterogeneous environment, the inference instruction pipeline maintains a continuous throughput state, and the end-to-end computation latency is stable within 12ms. The concentration of harmful gases, temperature and humidity, cable surface temperature and water level in the underground utility tunnel environment are kept within the preset safety target range.

[0041] Example 2: The technical solution of this invention is operated on a pneumatically sealed underground pipe gallery test platform with sudden water seepage and gas leakage simulation functions. Its environmental feature parameter extraction unit collects multi-dimensional heterogeneous environmental raw datasets using a 50ms control clock step size, which is common in 2025. The setting of the control clock step size depends on the physical trade-off between the real-time update of the input signal and the computational power load of the operator processor. When the front-end gas sensor and water level sensor detect that the variation rate of the environmental signal is within the frequency band window of 10Hz to 50Hz, to prevent operator inference distortion caused by Nyquist sampling aliasing, the control clock step size is determined to be 50ms. Furthermore, this test platform uses a multi-channel gas generating valve array and a high-speed... The hydraulic injection pump set simulates a closed geological flow field environment, in which the resolution of harmful gas concentration measurement is not less than 0.01ppm, the sampling frequency of water level measurement is set to 100Hz, and the measurement accuracy of temperature and humidity sensors is maintained within ±0.1℃. This ensures that the original dataset of the multidimensional heterogeneous environment completely reproduces the non-stationary multi-field coupling abrupt change characteristics inside the pipe gallery at the physical input end. On this basis, in order to verify the feature perception anti-interference capability of the computing system of this invention in the actual industrial electromagnetic environment, Gaussian white noise with a signal-to-noise ratio of 20dB is actively superimposed on the external data bus to introduce 50Hz power frequency interference harmonics generated by the high-voltage frequency conversion transmission cable, thereby forming an original input data source containing background noise characteristics.

[0042] To confirm the synergistic effect of the computation graph adaptive reconstruction mechanism and the threshold damping calibration loop in solving the problems of data tensor transmission resistance and reconstruction oscillation, this experiment constructed a four-dimensional orthogonal verification system consisting of the present invention sample group, a partially missing control group one, a partially missing control group two, and an out-of-range control group one. The present invention sample group includes an in-situ computation graph topology reconstruction center, a dynamic hysteresis damping module, and a heterogeneous operator interaction gap compensation subsystem. The partially missing control group one removes the dynamic hysteresis damping module while keeping the topology of the other operators unchanged, and uses a fixed critical variation threshold of 0.65 for judgment. The partially missing control group two removes the heterogeneous operator interaction gap compensation subsystem and operates under the condition of not applying any time delay gap parameter locking compensation. The out-of-range control group one increases the step trigger frequency of external environmental abrupt stimulation to 75Hz, making it deviate from the preset frequency safety upper limit of 45Hz of the present invention system, in order to examine the nonlinear saturation failure effect at the parameter envelope boundary. The entire experimental process does not change the mapping relationship of the general computation register at the chip bottom layer.

[0043] After the experiment started, the high-pressure hydraulic water injection pump unit triggered instantaneous water injection at the 120th control cycle of the time step axis, simulating an abnormal rise in the water level of the underground pipe gallery at a rate of 2.3 mm / s. At the same time, the gas generating valve array released a high concentration of mixed harmful gases, causing the local concentration of hydrogen sulfide and methane to jump from 2.15 ppm to 45.82 ppm within 3 clock cycles. The original dataset of the multidimensional heterogeneous environment with 20 dB Gaussian white noise was incorporated into the environmental feature parameter extraction unit. The spatial domain denoising conditioning module used a spatial adjacency matrix to weight and filter the original dataset of the multidimensional heterogeneous environment to generate a spatial denoising feature matrix. The electromagnetic harmonic components in the spatial denoising feature matrix decreased by 85%, effectively filtering out 50 Hz power frequency interference. The temporal feature fusion module input the spatial denoising feature matrix into the long short-term memory network, extracted the spatiotemporal sequence correlation, and output the spatiotemporal dynamic feature tensor. During this process, the residual accumulation feedback loop in the topological correlation matrix calculation unit concurrently extracts the topological residual accumulation variables of the preceding state section. The operator association mapping module will convert the spatiotemporal dynamic feature tensor Cumulative variables of topological residuals from preceding state sections The cascaded splicing tensor input topology mapping dense layer constitutes the computation and output of the logical topology correlation matrix. ,in For spatiotemporal dynamic feature tensors, For topological residual cumulative variables, The logical topological correlation matrix, as shown by the data monitoring points, indicates that the logical topological correlation matrix decreases with increasing environmental and material variability. The deviation of the weight residuals between adjacent operator nodes trended upward from 0.12 to a maximum of 0.78, accurately revealing the nonlinear mass transfer resistance of the hardware and software data flow under transient change conditions. This is evident in the logical topology correlation matrix. When the deviation value of adjacent operator nodes in the network crosses the critical variation threshold, the graph topology in-situ reconstruction center in the adaptive topology reconstruction unit decouples the neural network computation graph in-situ within the memory addressing space and disconnects low-contribution operator links with a computational power contribution of less than 5%. It then retrieves and dynamically mounts spatiotemporal dynamic feature tensors from the preset heterogeneous dense computational sub-network library. A dense computational subnetwork matching the variation trend outputs the optimal computational graph architecture instructions, and the embedded dynamic hysteresis damping module performs discrete counting per unit time on the structural computational graph reconstruction events to output the reconstruction trigger frequency. To test the system's gradient adaptability under different problem severity levels, experiments were conducted to measure the system's computational delay and addressing stability at low-level (10Hz), medium-level (30Hz), and high-level (50Hz) environmental disturbance frequencies. Under the 10Hz and 30Hz conditions, the end-to-end computational delay of the prototype was 6.2ms and 8.4ms, respectively. However, when faced with high-frequency environmental noise at 50Hz, which caused the reconstruction trigger frequency to... When the frequency exceeds the 45Hz safety limit, the system automatically initiates a dynamic threshold calibration procedure based on the formula. Increase the current critical mutation threshold, where, This is the corrected critical mutation threshold. Based on the basic threshold constant, This is a dynamic adjustment coefficient. To reconstruct the trigger frequency, the calculated The mutation threshold was increased from a basic threshold constant of 0.65 to 1.30 in a stepwise manner, forcibly extending the time lock protection period between adjacent reconstruction actions. The end-to-end computation latency of the sample group of this invention converged to 11.1ms. In contrast, the reconstruction trigger frequency of the partially missing control group continued to deteriorate to 54Hz under 50Hz noise, causing memory pipeline deadlock, and its end-to-end computation latency amplified to 43.6ms. At the same time, during the transient window of operator alternation, the memory addressing latency jitter caused by bus memory delay during instruction switching in the heterogeneous operator interaction gap compensation subsystem of the computation flow control unit was increased. When the memory addressing delay jitter is measured When the delay is 15μs, it is based on the preset time delay mapping factor. Using the linear product formula Calculate the delay compensation parameters ,in, For delay compensation parameters, For time delay mapping factor, To store addressing delay jitter, the inter-operator interaction gap latch time is dynamically increased by 6.0 μs by updating the incremental register. This results in an environmental control compensation strategy that outputs a pure digital control command stream for driving external fan speed regulation, dehumidifier start / stop, and drain pump adjustment. However, the partially missing control group two, which removed this compensation mechanism, produces a 24.5% timing phase lag distortion. Furthermore, when the external disturbance frequency is increased to an extreme value of 75 Hz in the out-of-range control group one, the critical variation threshold is limited by the trigger hardware. When the calibration calculation enters the saturation region with a maximum value of 1.85, the system's graph topology decoupling capability reaches its physical limit, and its end-to-end calculation delay jumps nonlinearly to 38.2ms. This confirms that the preset 45Hz frequency safety limit constitutes the boundary range of the entire adaptive calculation. The dynamic adjustment coefficient in this calculation formula has an engineering value range limited to 0.01 to 0.05, and is calibrated to 0.02 based on the system's anti-oscillation stability test under multi-field abrupt flow fields. If this coefficient is set below 0.01, the step feedback increment of the threshold is insufficient to effectively dampen the addressing jumps caused by high-frequency noise, leading to frequent pipeline refresh and reconstruction oscillations. If this coefficient is set above 0.05, the corrected variation threshold will enter saturation too quickly. The blocking zone causes the system to lose its topological adaptive response capability to real environmental catastrophic changes and disrupts the convergence of the hardware control loop. By weaving multi-scale state projection of boundary residuals and addressing delay feedforward hedging loops into the hardware addressing path transformation process, the inter-operator addressing conflicts and instruction timing anomalies of the underlying general-purpose computing chip are stably suppressed under the full-gradient non-stationary disturbance environment. In the 120-hour continuous operation test of the entire underground utility tunnel environmental adaptive computing system, the number of floating-point operations per second of the inference instruction pipeline is kept constant within the rated specification range. The concentration of harmful gases and water level variables of each monitoring node inside the utility tunnel are finally restored to the preset safety target baseline within 15.2 minutes after the control strategy is output, and the hardware addressing state does not produce control phase lag.

[0044] Example 3: This example combines Figures 1 to 2 This paper describes an adaptive computing system for underground utility tunnels based on neural networks. Figure 1As shown, the original dataset of the multidimensional heterogeneous environment is input into the environmental feature parameter extraction unit to extract the spatiotemporal dynamic feature tensor. The spatiotemporal dynamic feature tensor is input into the topology correlation matrix calculation unit to calculate the logical topology correlation matrix between operator nodes. The cumulative variables of the topology residuals of the preceding state sections are also input into the topology correlation matrix calculation unit to participate in the calculation of the logical topology correlation matrix between operator nodes. The logical topology correlation matrix output by the topology correlation matrix calculation unit is input into the adaptive topology reconstruction unit. The adaptive topology reconstruction unit includes an in-situ topology decoupling computation graph with a dense computational sub-network and a dynamic hysteresis damping module. When the safety limit is exceeded, the variation threshold is increased. The in-situ topology decoupling computation graph is attached to the dense computation sub-network, and the reconstruction trigger frequency is output to the dynamic hysteresis damping module. When the safety limit is exceeded, the variation threshold is increased. The dynamic hysteresis damping module is attached to the dynamic calibration critical variation threshold and outputs the in-situ topology decoupling computation graph to the dense computation sub-network. The adaptive topology reconstruction unit outputs the optimal computation graph architecture instruction to the computation flow control unit. Based on the optimal architecture instruction, the underlying hardware addressing path is configured. Based on the optimal architecture instruction, the computation flow control unit configures the underlying hardware addressing path and outputs the environmental control compensation strategy.

[0045] like Figure 2 As shown, the environmental feature parameter extraction unit is connected to the topology correlation matrix calculation unit, the topology correlation matrix calculation unit is connected to the adaptive topology reconstruction unit, and the adaptive topology reconstruction unit is connected to the computational flow control unit.

[0046] Example 4: In the long-term continuous operation environment of underground utility tunnels, when the ambient temperature experiences seasonal drift, the humidity-sensitive components undergo physical aging, and the sensor signal acquisition bus is exposed to a humid electromagnetic flow field for a long time, resulting in baseline data drift, the baseline of the original dataset of the multidimensional heterogeneous environment slowly and continuously decays. The monotonic causal mapping between the input variables of the neural network model and the physical channel results in long-term loss of timeliness. Consequently, the accuracy of the optimal computational graph architecture instruction retrieval under transient and sudden operating conditions decreases, causing erroneous divergence in the calculation of operator interaction gaps. The hardware addressing path of the underlying general-purpose computing chip fails, causing the environmental control compensation strategy to deviate from the preset protection range.

[0047] To address the uncertainty caused by long-period baseline data drift in network computation graph reconstruction, the adaptive topology reconstruction unit incorporates a sliding time window with a fixed length of 30 days. Within this window, a timeliness guarantee procedure performs daily polling analysis on the historical environmental dataset, and initiates an online parameter baseline update procedure when the average reconstruction trigger frequency of the preceding period exceeds 40Hz. The graph topology in-situ reconstruction center utilizes a pre-defined heterogeneous dense computational sub-network library with a network topology architecture comprising three cascaded graph convolutional layers and self-attention layers. When the logical topology correlation matrix... When the deviation value of adjacent operator nodes in the graph crosses the critical variation threshold, the graph topology in-situ reconstruction center obtains the current spatiotemporal dynamic feature tensor. The generated mutation trend vector is used to calculate the cosine similarity between the mutation trend vector and the topological feature vector of the candidate subnetwork in the heterogeneous dense computing subnetwork library. The candidate subnetwork with the highest cosine similarity is selected as the attached dense computing subnetwork, and the optimal computing graph architecture instruction is output in the address space. In order to transform the irregular nonlinear graph topology model structure in the heterogeneous dense computing subnetwork library into a feature vector that can be quantized and compared, each candidate subnetwork structure in the library is pre-parsed into a standardized node adjacency matrix and operator operation list, and then globally pooled and mapped through an offline trained graph structure embedding network to compress and flatten its static connection topology into a fixed 128-dimensional topological feature vector. This allows it to directly perform dot product and cosine similarity measurement with the 128-dimensional mutation trend vector extracted from the spatiotemporal dynamic feature tensor in the same dimensional space. At the same time, the multi-scale state reprojection subsystem in the adaptive topology reconstruction unit retrieves the fixed multi-dimensional orthogonal subspace projection matrix in the base address of the memory. Matrix multiplication is used to accumulate the topological residual variables of the preceding state section. Projection matrix of multidimensional orthogonal subspace Multiply, according to the formula The state-preserving vector is calculated. ,in, Preserve the vector for the state. For multidimensional orthogonal subspace projection matrix, For the topology residual accumulation variable, the composite dimension features are reduced in dimensionality and constrained to the static space of the weight register directly addressed by the high-speed bus. Furthermore, when the dimensionality-upgrading zero-filling module adjusts the tensor dimension of the state-preserving vector to match the input feature dimension of the dense computational subnetwork, the adaptive topology reconstruction unit converts this state-preserving vector... The weight register of the underlying general-purpose computing chip is written to realize the continuous state transition of the hardware and software architecture under long-period baseline drift conditions. The multidimensional orthogonal subspace projection matrix used here is to collect residual historical samples under non-stationary conditions in advance and construct a feature matrix. The core principal component orthogonal basis vectors are obtained by performing singular value decomposition and normalized by the Schmitt orthogonalization method before being offline and solidified in the base address of high-speed memory. When the dimension of the cumulative variable of the topological residual of the previous state section changes abruptly during operation, the projection unit performs dynamic clipping or zero matrix padding on the row and column index of the projection matrix according to the current boundary size, so as to ensure that the mathematical dimensions of the two are matched and accurately mapped in real time during matrix multiplication operations.

[0048] The adaptive topology reconstruction unit applies a multi-dimensional orthogonal subspace projection matrix and a long-period parameter reference online update procedure. The input tensor offset caused by multi-field coupling drift is canceled in place at the initial end of the time step axis. The updated weight register value controls the cumulative divergence rate of operator cascade inference to within 0.3%. The end-to-end computation latency remains stable at 11.4ms after running for 180 days. The addressing path of the underlying general-purpose computing chip does not experience addressing interruption due to baseline failure. The response curves of temperature and humidity variables inside the underground utility tunnel smoothly match the preset control phase target.

[0049] Example 5: When the system faces the on-site deployment of a new utility tunnel section, due to the differences in cable routing characteristics in different physical fields, the input original signal and the preset benchmark of the model have nonlinear channel resistance, resulting in bias pseudo-features. During the system startup transient, the environmental feature parameter extraction unit receives 100 stable signals of control clock steps from each sensor to construct a calibration sequence, which is then processed by the zero-point compensation subsystem according to the calibration formula. The cumulative topological residuals of the preceding state section were calculated. The initial value, where, These are the initial values ​​for the cumulative variables of the topological residuals. This is the bias correction factor. The original telemetry voltage vector is used to complete the in-situ locking of the baseline, so that the adaptive computing system can control the data flow error within the register safety boundary before generating transient adjustment actions.

[0050] After completing the topological residual accumulation variables After initial value calculation, the temporal feature fusion module in the environmental feature parameter extraction unit uses an adaptive sampling rate matching operator to calibrate the spatiotemporal dynamic feature tensor. With improved time resolution, when the residual value of the power frequency interference voltage fluctuation input from the external bus is greater than 12mV, the time window length of the adaptive feature acquisition is shortened from 500ms to 200ms, and the sampling point density is increased to offset the timing phase deviation of the addressing path caused by sudden operating conditions. The end-to-end calculation delay is controlled within 12ms, and the lock-up time of the operator interaction gap of the underlying general-purpose computing chip is stably maintained within the preset safety index range.

[0051] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit of this application and the scope of protection of this invention, and all of these forms are within the protection scope of this application.

Claims

1. An adaptive computational system for underground utility tunnel environments based on neural networks, characterized in that, include: The environmental feature parameter extraction unit acquires the original dataset of the multidimensional heterogeneous environment of the underground utility tunnel and extracts the spatiotemporal dynamic feature tensor. The topological correlation matrix calculation unit calculates the logical topological correlation matrix between each operator node by superimposing the cumulative variables of the topological residuals of the preceding state section on the spatiotemporal dynamic feature tensor. The adaptive topology reconstruction unit, when the deviation value of adjacent operator nodes in the logical topology correlation matrix crosses the critical mutation threshold, decouples the neural network computation graph in situ to cut off low-computing-power contribution links, and attaches a dense computational sub-network that matches the mutation trend of the spatiotemporal dynamic feature tensor to output the optimal computation graph architecture instruction; wherein, the embedded dynamic hysteresis damping module counts the number of reconstruction events of the neural network structure computation graph and calculates the reconstruction trigger frequency, and activates the threshold dynamic calibration program to increase the current critical mutation threshold when the reconstruction trigger frequency crosses the 45Hz frequency safety upper limit; The computation flow control unit configures the hardware addressing path of the underlying general-purpose computing chip according to the optimal computation graph architecture instructions in order to output the environmental control compensation strategy.

2. The adaptive calculation system for underground utility tunnel environment based on neural networks according to claim 1, characterized in that, The adaptive topology reconstruction unit is also cascaded with a multi-scale state reprojection subsystem. The multi-scale state reprojection subsystem is used to perform matrix projection of the old topology boundary residual information before decoupling onto a low-dimensional orthogonal subspace to generate a state-preserving vector during the reconstruction transient window of the neural network computation graph. When the input feature dimension of the dense computation subnetwork is inconsistent with the tensor dimension of the state-preserving vector, zero eigenvalues ​​are appended to the end of the state-preserving vector to complete the tensor dimension alignment, and the initial weight bias of the dense computation subnetwork is rewritten using the dimension-aligned state-preserving vector.

3. The adaptive calculation system for underground utility tunnel environment based on neural networks according to claim 1, characterized in that, The compute flow control unit also includes a heterogeneous operator interaction gap compensation subsystem, which is used to quantitatively obtain the memory addressing delay jitter caused by bus memory delay during instruction switching. When the measured memory addressing delay jitter is 15μs, a delay compensation parameter of 6.0μs is calculated by linear product based on a delay mapping factor of 0.4, and the dynamic locking of the operator interaction gap of the underlying general-purpose computing chip is increased by 6.0μs.

4. The adaptive calculation system for underground utility tunnel environment based on neural networks according to claim 1, characterized in that, The environmental feature parameter extraction unit includes a heterogeneous data parallel acquisition module, a spatial noise reduction and conditioning module, and a temporal feature fusion module. The heterogeneous data parallel acquisition module is used to read data such as harmful gas concentration, temperature and humidity, cable surface temperature, and water level from external sensors in parallel to construct a multidimensional heterogeneous environment raw dataset. The spatial noise reduction and conditioning module is cascaded with the heterogeneous data parallel acquisition module and is used to apply a spatial adjacency matrix to the multidimensional heterogeneous environment raw dataset for weighted filtering to generate a spatial noise reduction feature matrix. The temporal feature fusion module receives the spatial denoising feature matrix output by the spatial domain denoising conditioning module. It is used to input the spatial denoising feature matrix into the long short-term memory network along the time axis, extract the spatiotemporal sequence association, and output the spatiotemporal dynamic feature tensor.

5. The adaptive calculation system for underground utility tunnel environment based on neural networks according to claim 1, characterized in that, The topology correlation matrix calculation unit includes a residual accumulation feedback loop and an operator correlation mapping module. The residual accumulation feedback loop is used to read the control strategy deviation value of the neural network model in the underground utility tunnel environment inference at the current time step, and accumulate it into the topology residual accumulation variable of the preceding state section using the first-order time discrete integral. The operator correlation mapping module is logically associated with the residual accumulation feedback loop and is used to concatenate and spatiotemporally concatenate the spatiotemporal dynamic feature tensor with the topology residual accumulation variable of the preceding state section. The concatenated tensor is then input into the topology mapping dense layer, and the logical correlation weights between each pair of operator nodes in the neural network model are output to form a logical topology correlation matrix.

6. The adaptive calculation system for underground utility tunnel environment based on neural networks according to claim 1, characterized in that, The adaptive topology reconstruction unit also includes a graph topology in-situ reconstruction center. When the deviation value of adjacent operator nodes in the logical topology correlation matrix crosses the critical mutation threshold, the graph topology in-situ reconstruction center decouples the neural network computation graph in the memory addressing space to disconnect the computing power link, and retrieves and dynamically attaches the dense computation subnetwork corresponding to the mutation direction of the spatiotemporal dynamic feature tensor from the preset heterogeneous dense computation subnetwork library.

7. The adaptive calculation system for underground utility tunnel environment based on neural networks according to claim 1, characterized in that, The environmental control compensation strategy output by the computational flow control unit is a pure digital control command stream. The pure digital control command stream includes dynamic speed regulation commands for fans, start / stop switching commands for dehumidifiers, dynamic scheduling commands for cable loads, and adjustment commands for drainage pumps. These commands are used to change the operating status variable values ​​of the corresponding external hardware within the underground utility tunnel environment.

8. The adaptive calculation system for underground utility tunnel environment based on neural networks according to claim 1, characterized in that, The adaptive topology reconstruction unit also includes a state degradation monitoring module; the state degradation monitoring module is used to perform sliding window statistics on the trigger frequency of the threshold dynamic calibration procedure per unit time, calculate the threshold calibration drift variation rate, and output an alarm signal indicating that an abnormal change has occurred in the calculation of the underground utility tunnel neural network when the threshold calibration drift variation rate is greater than 2.5 Hz / s for three consecutive sliding windows.

9. The adaptive calculation system for underground utility tunnel environment based on neural networks according to claim 1, characterized in that, The control clock steps of the environmental feature parameter extraction unit, the topology correlation matrix calculation unit, the adaptive topology reconstruction unit, and the computation flow control unit are all locked at 50ms, and the data throughput rate of the original dataset of the multidimensional heterogeneous environment is not less than 20MB / s, so as to ensure that the end-to-end computation latency of obtaining the environmental control compensation strategy output from the original dataset of the multidimensional heterogeneous environment is less than 12ms.

Images