A mechanism-loss-free turnout digital twin scene adaptive edge lightweight method

Abstract

The application discloses a mechanism-lossless turnout digital twin scene adaptive edge lightweight method, and relates to the technical field of digital twin, and the technical solution points are as follows: the calculation model and the geometric model of the turnout digital twin are divided into a mechanism key area and a non-key area and are assigned a protection weight, and MFI is defined as a unified quantitative constraint boundary; a heterogeneous coupling graph containing physical field nodes, geometric nodes and calculation nodes is constructed, mechanism constraint channel pruning and mechanism weighted edge folding simplification are performed, and constraints are transmitted bidirectionally through the heterogeneous coupling graph; the pruned calculation model is used as a student model, the full-quantity model is used as a teacher model, a joint distillation loss function is constructed, and fine-tuning training is performed; the edge terminal monitors the computing power load rate and the prediction accuracy confidence during running, and automatically switches the model configuration under the constraint that the MFI is not lower than the lower limit value. The application realizes mechanism-lossless lightweight deployment of the turnout digital twin model on the edge.

Description

Technical Field

[0001] This invention relates to the field of digital twin technology, and more specifically, to a mechanism-free adaptive edge lightweighting method for turnout digital twin scenarios. Background Technology

[0002] Digital twin technology provides a novel technical means for real-time status monitoring, fault diagnosis, and life prediction of railway turnout equipment by constructing high-fidelity virtual mapping models of physical entities. A complete turnout digital twin system typically includes two core models: a computational model, namely a physics simulation network, used to predict the distribution of multiple physics fields such as stress, vibration, and temperature fields during turnout operation in real time, supporting structural performance assessment and fault inversion; and a geometric model, namely a three-dimensional visualization mesh, used to present the three-dimensional geometry of the turnout, the spatial relationship of its components, and its deformation state on the monitoring terminal, supporting intuitive visualization interaction. However, both the computational and geometric models of high-fidelity turnout digital twins have extremely high storage and computational overhead. The number of parameters in the physics simulation network often reaches millions to tens of millions, and the number of facets in the high-precision three-dimensional geometric model can reach hundreds of thousands to millions. In railway sites, the terminal devices capable of deploying twin models are all edge devices with highly limited computing resources, storage space, and power supply, and cannot support the full high-fidelity model.

[0003] Existing techniques for lightweighting digital twin models mainly include channel pruning methods based on Taylor expansion, geometric mesh simplification methods based on improved edge folding, scene-adaptive model deployment methods, and knowledge distillation model compression methods. However, these methods suffer from the following core flaws: First, the lack of quantitative standards for mechanism fidelity means that existing methods define lossless as merely a qualitative measure of maintaining prediction accuracy without quantifiable mechanism fidelity indicators to uniformly constrain pruning intensity, geometric simplification degree, and distillation loss weights, resulting in uncontrollable lightweight boundaries. Second, computational model lightweighting and geometric model lightweighting are disconnected. Existing methods treat channel pruning of the computational network and edge folding of the 3D mesh as two independent tasks, ignoring the cross-model coupling relationship where the local mesh density of the geometric model in the turnout digital twin directly affects the accuracy of physical field calculations. Third, scene adaptation remains statically preset. Existing methods preset several fixed lightweight schemes based on equipment type, failing to perceive real-time computing load fluctuations and prediction accuracy changes during edge device operation. Fourth, knowledge distillation lacks physical mechanism constraints. General distillation methods only focus on fitting the final prediction result, without introducing physical mechanism constraints into the distillation loss function, resulting in lightweight models that may fit well at the data level but violate fundamental laws at the physical level.

[0004] Therefore, how to study and design a mechanism-independent, lossless, adaptive edge lightweighting method for turnout digital twin scenarios that can overcome the above-mentioned defects is an urgent problem to be solved. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this invention is to provide a mechanism-free adaptive edge lightweighting method for turnout digital twin scenarios. This method coordinates lightweighting computational and geometric models under the constraint of quantization mechanism fidelity, and adaptively schedules lightweighting configurations based on the real-time operating status of the edge.

[0006] The above-mentioned technical objective of the present invention is achieved through the following technical solution: A mechanism-insensitive adaptive edge lightweighting method for turnout digital twin scenarios includes the following steps: The computational and geometric models of the turnout digital twin are divided into critical and non-critical regions, and a mechanism protection weight is assigned to each region. The Mechanism Fidelity Index (MFI) is defined as a unified quantitative constraint boundary for lightweight operations. The MFI integrates three dimensions: physical field prediction error, geometric fidelity error, and coupling consistency. A lower limit value for the MFI is set, and the MFI after any lightweight operation must not be lower than the lower limit value of the MFI. A heterogeneous coupling graph containing physics nodes, geometry nodes, and computation nodes is constructed. Based on the heterogeneous coupling graph, the computation model is pruned for mechanism constraint channels, and the geometry model is simplified by weighted edge folding. Constraints are bidirectionally transmitted through the heterogeneous coupling graph: when the mesh quality decreases after geometry simplification, the importance score of the associated computation channels is corrected; after computation channels are pruned, the folding cost of the associated geometric regions is increased. Using the pruned computational model as the student model and the full high-fidelity computational model as the teacher model, a joint distillation loss function is constructed, which includes data fitting loss, physical equation residual loss, and MFI constraint loss, and the student model is fine-tuned and trained. The edge terminal monitors the computing load rate and prediction accuracy confidence during operation, and automatically switches the model configuration under the constraint that the MFI is not lower than the MFI lower limit.

[0007] Furthermore, the key areas of the mechanism include the switch rail contact section, the frog center area, the locking transmission mechanism, the slide plate contact area, and the actuating rod connection node; The non-critical area includes the equipment base, outer protective cover, and auxiliary cabinet; The mechanism protection weight of the critical mechanism region is greater than that of the non-critical mechanism region; The MFI is the weighted sum of the physical field prediction error ratio, the geometric fidelity error, and the coupling consistency.

[0008] Furthermore, the mechanism-constrained channel pruning specifically includes: For each convolutional channel, a Taylor number is calculated as the initial importance. A mechanism loss regularization term is introduced to correct the importance of channels connected to the mechanism key region. The correction method is to add the mechanism protection weight of the physical field region connected to the channel and the physical field prediction error increment of the physical field region after removing the channel to the initial Taylor number to obtain the corrected channel importance. The channels are sorted according to their corrected importance, and the redundant channels with the lowest contribution are removed. The pruning ratio is gradually increased from a conservative ratio. The MFI is calculated after each round of pruning. If the MFI is lower than the MFI lower limit, the result is reverted to the previous round of pruning.

[0009] Furthermore, the specific method for simplifying the mechanism-weighted edge folding is as follows: The mechanism protection weights are mapped to mesh vertices, and the edge folding cost function is improved so that the folding cost increases as the mechanism protection weights of the region where the vertex is located increase. The vertex folding cost of the mechanistic critical region is amplified to limit the mesh simplification of the mechanistic critical region, while the vertex folding cost of the non-critical region is not amplified to allow the mesh simplification of the non-critical region to be greater than that of the mechanistic critical region.

[0010] Furthermore, the bidirectional transmission of constraints through the heterogeneous coupling graph specifically refers to: After geometric simplification, the mesh quality score of each geometric node is updated. When the consistency of the face normal vector in the geometric critical region decreases beyond the preset threshold or the maximum aspect ratio of the face exceeds the preset threshold, a degradation alarm is sent to the connected physical field node and computing node, triggering the re-evaluation of the importance of the corresponding computing channel and the channel recovery operation. After calculating the network pruning, the distribution of geometric regions connected by the pruned channels is statistically analyzed. When the number of associated channels pruned in a geometric region exceeds a preset threshold, a multiplication correction factor greater than 1 is applied to the folding cost of the geometric region, so that the increase in folding cost of the geometric region in subsequent folding iterations is greater than that of other geometric regions.

[0011] Furthermore, the calculation of the residual loss of the physical equation is as follows: In the key mechanism region, the output of the student model is substituted into the stress balance equation, heat conduction equation and contact force balance condition to calculate the weighted sum of squares of the residuals, where the weight of each region is the mechanism protection weight of the region. And / or, the calculation of the MFI constraint loss is specifically as follows: when the current MFI in the distillation process is lower than the MFI lower limit, the penalty term is activated; otherwise, it is zero.

[0012] Furthermore, the method also includes: When the edge terminal starts up, it performs a hardware self-test to detect the number of CPU cores and clock speed, available memory, accelerator type and computing power, and available power supply, forming a hardware capability vector. The hardware capability vector is matched with the hardware requirement vector in the preset multi-level lightweight configuration library using cosine similarity. The configuration with the highest similarity and satisfying that the MFI is not lower than the MFI lower limit is selected as the initial loading configuration.

[0013] Furthermore, the automatic switching of model configuration under the constraint that MFI is not lower than the lower limit of MFI specifically refers to: When the computing load rate exceeds the upper limit threshold and the duration exceeds the preset time threshold, a downgrade switch is triggered. From the preset multi-level lightweight configuration library, a configuration is selected that satisfies the MFI not being lower than the lower limit of the MFI, has the largest estimated load reduction, and has a greater degree of lightweighting. When the prediction accuracy confidence level is lower than the lower limit threshold and the computing power load rate is lower than the recovery threshold, an upgrade switch is triggered. From the preset multi-level lightweight configuration library, a configuration is selected that satisfies the condition that the MFI is not lower than the lower limit value of the MFI, has the largest improvement in prediction accuracy, and has a smaller degree of lightweighting. The switching execution adopts atomic configuration replacement. First, the newly configured model weights are loaded into the spare memory slot. After verifying the integrity, the inference engine pointer is atomically switched from the current model to the new model. The inference process is not interrupted during the switching process.

[0014] Furthermore, the method also includes: The edge terminal periodically uploads runtime metric logs to the cloud. The runtime metric logs include load change trajectories, configuration switching frequency, and measured MFI values ​​under each configuration. The cloud-based system optimizes a preset multi-level lightweight configuration library based on accumulated logs, adjusts the pruning ratio, LOD accuracy, and distillation loss weight for each level, and then distributes the updated preset multi-level lightweight configuration library to the edge terminal.

[0015] Furthermore, the method also includes: Calculate the overall lightweight performance ratio, which is the ratio of MFI to the sum of the calculated parameter compression ratio and the geometric patch number compression ratio; When the overall lightweight efficiency ratio is greater than 1, it is determined to be effective lightweighting.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention provides a mechanism-free adaptive edge lightweighting method for turnout digital twin scenarios. It constructs a four-level edge lightweighting process: mechanism key area anchoring and MFI quantization, mechanism coupling perception-based collaborative lightweighting, mechanism constraint knowledge distillation, and scene self-perception online adaptive scheduling. This transforms mechanism fidelity from a vague qualitative commitment in traditional methods into a calculable and verifiable quantitative indicator. Specifically, MFI integrates three dimensions: physical field prediction error, geometric fidelity error, and coupling consistency. It sets a minimum MFI limit for any lightweighting operation, thus eliminating the risk of lightweighting damaging the core physical mechanism. Simultaneously, the heterogeneous coupling graph achieves bidirectional collaboration between computational model channel pruning and geometric model edge folding simplification, resolving the fundamental contradiction in existing technologies where computation and geometric lightweighting are separated, leading to geometric simplification compromising computational accuracy and computational pruning ignoring geometric key areas. Mechanism constraint knowledge distillation, through physical equation residual loss and MFI constraint loss, ensures that the lightweight student model conforms to basic mechanical and thermal laws even under extreme conditions. Online adaptive scheduling enables edge terminals to automatically switch model configurations based on real-time computing load and prediction accuracy, while being protected by MFI hard constraints throughout the process. This invention achieves lossless and lightweight deployment of turnout digital twin models on resource-constrained edge terminals.

[0017] 2. This invention achieves high-level protection for the core stress-bearing components of turnouts through the fine division of mechanistic critical regions and the construction of heterogeneous coupling graphs. Specifically, core components such as the switch rail contact section and the frog center area are designated as mechanistic critical regions and assigned high protection weights, while non-critical regions are assigned low protection weights. This ensures that lightweighting operations are strictly limited in critical regions, while allowing for greater simplification in non-critical regions, thereby maximizing the overall compression ratio while maintaining the fidelity of the core mechanism. Mechanism-constrained channel pruning, through the combination of Taylor numbers and mechanistic loss regularization terms, ensures that channels connected to high protection weight regions are retained even if their data-driven contribution is small, avoiding the risk of erroneous pruning. Mechanism-weighted edge folding simplification, by improving the folding cost function, amplifies the folding cost of vertices in critical regions while not amplifying the folding cost of vertices in non-critical regions, achieving differentiated simplification. The heterogeneous coupling graph bidirectional transitive constraint mechanism, through degradation alarms and channel pruning feedback, ensures that the importance of calculated channels can be corrected in a timely manner when the mesh quality deteriorates after geometric simplification, and that the pruning of calculated channels can protect the geometric region in reverse, forming a closed-loop collaborative protection. This invention achieves a balance between mechanism fidelity and compression efficiency.

[0018] 3. This invention embeds physical mechanism constraints directly into the knowledge distillation process by constructing a joint distillation loss function that includes data fitting loss, physical equation residual loss, and MFI constraint loss. Specifically, the physical equation residual loss forces the student model's output to satisfy the stress balance equation, heat conduction equation, and contact force balance condition during training, making the lightweight model not only approximate the teacher model in terms of data distribution but also consistent with the real physical process in terms of physical laws. The MFI constraint loss dynamically monitors the MFI value during distillation, immediately imposing a penalty if it falls below the lower limit, forcing the distillation direction to always move towards high-fidelity optimization. This invention ensures the reliability and credibility of the physical field inference results based on the lightweight model at the edge in engineering applications.

[0019] 4. This invention achieves intelligent selection of the initial lightweight configuration through hardware self-testing and cosine similarity matching during edge terminal startup, avoiding the problem of mismatch between manually preset configurations and actual hardware capabilities. During runtime, it continuously collects computing load rate and prediction accuracy confidence, and sets clear upper and lower limit thresholds and duration thresholds, making switching decisions both sensitive and resistant to interference. Both degradation and upgrade switching are hard-constrained by MFI not falling below the lower limit, ensuring the reliability of core mechanism criteria at any time. The atomic configuration replacement technology, through preloading of spare memory slots and atomic pointer switching, achieves smooth switching without interruption of inference, meeting the stringent real-time requirements of industrial sites. This invention achieves a leap from static preset to online self-sensing and adaptive operation, enabling the turnout digital twin system to automatically adjust the degree of lightweighting between peak traffic periods and idle periods, balancing real-time performance and mechanism fidelity.

[0020] 5. This invention establishes a closed-loop optimization link between edge terminals and the cloud through an edge-cloud synchronous iteration mechanism. Edge terminals periodically upload runtime metric logs to the cloud, including load change trajectories, configuration switching frequency, and measured MFI values ​​for each configuration. The cloud analyzes the actual performance of each configuration under different operating conditions based on accumulated logs, dynamically adjusts the pruning ratio, LOD accuracy, and distillation loss weight, and then distributes the updated configuration library to the edge terminals. This invention achieves self-evolution of the lightweight configuration library, enabling the system's lightweight strategy to be continuously optimized with the accumulation of runtime data, thereby continuously improving the robustness and adaptability of edge deployment.

[0021] 6. This invention proposes a comprehensive lightweight performance ratio as a unified evaluation index, combining the MFI with the sum of the compression ratio of computational parameters and the compression ratio of geometric facets, thereby realizing a quantitative comparison of different lightweighting schemes in terms of both mechanism fidelity and compression efficiency. Attached Figure Description

[0022] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart from an embodiment of the present invention; Figure 2 This is a schematic diagram of a heterogeneous coupling diagram in an embodiment of the present invention; Figure 3 This is a basic framework diagram of knowledge distillation in an embodiment of the present invention; Figure 4 This is a schematic diagram of the U-Net network architecture in an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of this invention are only for explaining this invention and are not intended to limit this invention.

[0024] Example: A mechanism-free adaptive edge lightweighting method for turnout digital twin scenarios, such as... Figure 1 As shown, it includes the following steps: S1: Divide the calculation model and geometric model of the turnout digital twin into a mechanism critical area and a non-critical area, and assign mechanism protection weights to each area. Define the mechanism fidelity index (MFI) as a unified quantitative constraint boundary for lightweight operations. The MFI integrates three dimensions: physical field prediction error, geometric fidelity error, and coupling consistency. Set a lower limit value for the MFI. The MFI after any lightweight operation must not be lower than the lower limit value of the MFI. S2: Construct a heterogeneous coupling graph containing physics nodes, geometry nodes, and computation nodes. Based on the heterogeneous coupling graph, prune the constraint channels of the computation model's execution mechanism, and simplify the weighted edge folding of the execution mechanism of the geometry model. The constraints are bidirectionally transmitted through the heterogeneous coupling graph: when the mesh quality decreases after geometry simplification, the importance score of the associated computation channel is corrected; after the computation channel is pruned, the folding cost of the associated geometric region is increased. S3: Using the pruned computational model as the student model and the full high-fidelity computational model as the teacher model, a joint distillation loss function is constructed, which includes data fitting loss, physical equation residual loss and MFI constraint loss, to fine-tune the student model. S4: When the edge terminal is running, it monitors the computing load rate and prediction accuracy confidence, and automatically switches the model configuration under the constraint that the MFI is not lower than the lower limit of the MFI.

[0025] In step S1, the critical mechanistic regions of the turnout digital twin model are first anchored. On the computational model side, the predicted output space of the physical field simulation network is divided into five critical mechanistic regions and three non-critical regions according to the turnout's physical structure. The critical mechanistic regions include the switch rail contact section area, the frog center area, the locking transmission mechanism area, the slide plate contact area, and the actuating rod connection node area. The non-critical regions include the equipment base area, the outer protective cover area, and the auxiliary cabinet area. On the geometric model side, the mesh vertices of the turnout's three-dimensional mesh model are divided into corresponding critical and non-critical mechanistic regions according to the same principle.

[0026] Next, mechanism protection weights are assigned to each region. These mechanism protection weights are denoted as... The values ​​are determined as follows: the mechanism protection weight for critical areas is greater than that for non-critical areas. In this embodiment, the mechanism protection weight for the switch rail contact section is set to 1.0, the mechanism protection weight for the frog core is set to 1.0, the mechanism protection weight for the locking transmission mechanism is set to 0.95, the mechanism protection weight for the slide plate contact area is set to 0.90, and the mechanism protection weight for the actuating rod connection node is set to 0.85; the mechanism protection weight for non-critical areas is set to 0.1 to 0.2.

[0027] Subsequently, the Mechanism Fidelity Index (MFI) is defined as the unified quantization constraint boundary for lightweight operations. The MFI integrates three dimensions: physics prediction error, geometric fidelity error, and coupling consistency. Its calculation formula is as follows: ; in, The physics prediction error ratio represents the normalized root mean square deviation between the physics predictions of the lightweight model and the full model in the key mechanism region. Geometric fidelity error, representing the ratio of the Hausdorff distance of the lightweight geometric model in the critical region of the mechanism to the turnout characteristic dimension; For coupling consistency, it represents the weight retention rate of directed edges in the computation-geometric coupling graph after lightweighting. , , These are the weights for each dimension, which are set to the default values ​​in this embodiment. , , The value of MFI ranges from 0 to 1, and the closer the value is to 1, the higher the mechanism fidelity.

[0028] Set the lower limit of MFI as In this embodiment The default value is 0.85. The MFI after any lightweight operation must not be lower than [a certain value]. .

[0029] Physical field prediction error ratio The specific calculation method is as follows: ; in, For the set of key mechanisms, The number of key regions of the mechanism; Represents a key region of a single mechanism; Mechanism protection weights for region r; This represents the number of spatial sampling points within region r; For the physical field prediction value of the lightweight model at the i-th sampling point, take the Von Mises equivalent stress value of the stress field; This represents the predicted physical field value of the full model at the i-th sampling point. This is the standard deviation of the full model's predicted values ​​within region r, used for normalization to eliminate differences in the dimensions of different physical fields.

[0030] Geometric fidelity error The calculation method is: the Hausdorff distance of the lightweight geometric model in the critical region of the mechanism divided by the turnout characteristic dimension.

[0031] Coupling Consistency The calculation method is as follows: after lightweighting, calculate the weight retention rate of directed edges in the geometric coupling graph, that is, the ratio of the sum of edge weights still retained after lightweighting to the sum of edge weights before lightweighting.

[0032] In step S2, as Figure 2 As shown, a heterogeneous coupling diagram of three types of nodes—physical field, geometry, and computation—is first constructed for the turnout. Taking the digital twin model of a No. 12 high-speed turnout under the jurisdiction of a railway bureau as an example, the physical field nodes correspond to eight physical field sub-regions, including five key mechanistic regions: the switch rail contact section region, the frog center region, the locking transmission mechanism region, the slide plate contact area region, and the action rod connection node region; and three non-key regions: the equipment base region, the outer protective cover region, and the auxiliary cabinet region. The geometry nodes correspond to eight grid sub-regions, one-to-one with the physical field nodes. The computation nodes correspond to 13 channel groups in the physical field simulation network, including a 5-layer encoder channel group, a 4-layer decoder channel group, and a 4-jump connection channel group of U-Net. The three types of nodes are connected by three types of directed edges: directed edges from geometry to physical field indicate that the geometry affects the distribution of the physical field; directed edges from physical field to computation indicate that the physical field region is predicted by a specific computation channel group; and directed edges from computation to geometry indicate that the sensitivity of the computation channel reflects the importance of the corresponding geometric region. In this embodiment, the heterogeneous coupling graph contains 21 nodes and 56 directed edges.

[0033] Next, mechanistic constraint channel pruning is performed on the computational model based on the heterogeneous coupling graph. For example... Figure 4As shown, the physics simulation network adopts a turnout multi-physics joint prediction network based on the U-Net encoder-decoder architecture. The input is the turnout geometric contour image and the working condition parameter vector, and the output is the spatial distribution map of the stress field, vibration acceleration field, and temperature field. The network consists of 5 encoder layers and 4 decoder layers. Each encoder layer has 2 convolutions, with the number of channels being 64, 128, 256, 512, and 512 respectively. Each decoder layer contains skip connections and 2 convolutions, with the number of channels being 512, 256, 128, and 64 respectively. The total number of parameters is approximately 7.8M. For each convolutional channel of U-Net, forward propagation is performed sample by sample on the validation set. The change in the first-order Taylor expansion approximate loss of the output feature map of each channel is calculated as the initial importance, denoted as the Taylor number. The validation set contains 500 pairs of turnout working conditions and physical field samples. The mean of the 500 samples is taken as the initial Taylor number for each channel.

[0034] A mechanism loss regularization term is introduced to correct the importance of channels connected to the mechanism critical region. The correction formula is: ; in, The corrected channel importance, The initial Taylor number, The mechanistic constraint strength coefficient is set to 0.3 by default in this embodiment. Let K be the set of physical field regions connected by channel k in the heterogeneous coupling diagram. This is to remove the physical field prediction error increment in region r after removing channel k. Through this correction, channels connected to regions with high mechanism protection weights, even with low initial Taylor numbers, will gain higher post-correction importance due to compensation from the mechanism loss regularization term, and thus be retained. For example, in this embodiment, channel number 187 in the encoder's third layer initially ranks in the bottom 30% of Taylor numbers and is a pruning candidate. However, because this channel is connected to the stress field prediction node in the frog core region in the heterogeneous coupling diagram, its removal causes the stress prediction error in the frog core region to increase from 3.2% to 8.7%, and the physical field prediction error increment to be 5.5%. After correction... Its ranking jumped to the top 50% and was ultimately retained.

[0035] Based on the revised channel importance The channels with the lowest contribution were sorted and removed. The pruning ratio was gradually increased from a conservative 15%, and the MFI was calculated after each round of pruning. In this embodiment, the MFI after 15% pruning was 0.93, which is greater than 0.85. After further increasing to 25%, the MFI was 0.89, which is greater than 0.85. After further increasing to 35%, the MFI was 0.86, which is greater than 0.85. After 40% pruning, the MFI was 0.82, which is less than 0.85. Therefore, the pruning was reduced to 35%. The final pruning ratio was 34.7%, the number of parameters decreased from 7.82M to 3.56M, the compression ratio was 45.5%, and the MFI was 0.865.

[0036] Next, a mechanism-weighted edge folding simplification was performed on the geometric model. The original turnout 3D mesh model has approximately 450,000 faces, and the OBJ format file size is approximately 22MB. Mechanism protection weights are mapped to mesh vertices, and the edge folding cost function is improved so that the folding cost increases with the increase of the mechanism protection weight in the region where the vertex is located. The improved edge folding cost formula is: ; in, The cost of the improved edge folding The folding cost is a standard quadratic error metric. To protect the strength coefficient, it is set to 2.0 by default in this embodiment. Let e ​​be the two endpoints. and A higher mechanistic protection weight is applied. The vertex folding cost in the mechanistic critical region is amplified to limit mesh simplification in that region, while the vertex folding cost in non-critical regions is not amplified to allow for greater mesh simplification in non-critical regions compared to the mechanistic critical region. Three levels of LOD accuracy models are generated through iterative edge folding: LOD1 is a high-precision model with the number of faces reduced to 180,000, a folding rate of less than 5% in the mechanistic critical region, and a folding rate of approximately 70% in the non-critical region. The accuracy is 0.96; LOD2 is a medium-precision model, with the number of faces reduced to 80,000, the folding rate of the critical region is less than 12%, and the folding rate of the non-critical region is about 85%. The value is 0.91; LOD3 is a low-precision model with the number of faces reduced to 30,000, the folding rate of the critical region is less than 20%, and the folding rate of the non-critical region is about 95%. It is 0.85.

[0037] Finally, bidirectional constraint transfer is achieved through a heterogeneous coupling graph. After geometric simplification, the mesh quality score of each geometric node is updated. The mesh quality score includes patch normal consistency and patch aspect ratio. When the patch normal consistency decreases by more than 20% or the maximum patch aspect ratio exceeds 10:1, a degradation alarm is sent to the connected physics and computation nodes, triggering a reassessment of the importance of the corresponding computation channel and a channel recovery operation. In this embodiment, during LOD2 simplification, the patch normal consistency of the variable cross-section mesh region of the switch rail increases from 3.2 degrees before simplification to 4.8 degrees after simplification, an increase of 50% exceeding the 20% threshold, and the maximum patch aspect ratio reaches 12:1, exceeding the 10:1 threshold, triggering a degradation alarm. The heterogeneous coupling graph transmits this alarm to the computation nodes connected to this geometric node, namely the decoder's Layer 1 and Layer 2 channel groups, triggering a round of channel importance reassessment for this region. The channel importance of decoder Layer 2 channels 312 and 401 is corrected in the reassessment. The value is improved, and the mesh is restored from being to be removed to being retained, thus eliminating the risk of the geometric mesh being oversimplified at the variable cross section of the switch rail, which would lead to a deterioration in the accuracy of subsequent stress calculations based on the simplified geometry.

[0038] On the other hand, after computational network pruning, the distribution of geometric regions connected by pruned channels is statistically analyzed. When the number of associated channels pruned in a certain geometric region exceeds a preset threshold, a multiplication correction factor greater than 1 is applied to the folding cost of that geometric region, making the increase in folding cost of that geometric region greater than that of other geometric regions in subsequent folding iterations. In this embodiment, the preset threshold is set to 30% of the total number of associated channels in that geometric region. Through this mechanism, geometric regions with a large number of pruned channels receive stronger protection in subsequent edge folding iterations, avoiding the risk of oversimplification of geometric critical regions due to computational pruning.

[0039] In step S3, as Figure 3 As shown, the pruned computational model is used as the student model, and the full high-fidelity computational model is used as the teacher model. A joint distillation loss function, including data fitting loss, physical equation residual loss, and MFI constraint loss, is constructed to fine-tune the student model. The teacher model is a full U-Net physics simulation network with 7.82M parameters; the student model is a pruned U-Net network with 3.56M parameters. The teacher model parameters are frozen, while the student model parameters are trainable.

[0040] The formula for calculating the combined distillation loss function is: ; in, For the combined distillation loss function, For data fitting loss, For the residual loss of the physical equation, For MFI-constrained loss, The physical loss weight is set to 0.2 by default in this embodiment. The MFI constraint weight is set to 1.0 by default in this embodiment.

[0041] Data fitting loss The calculation formula is: ; Where MSE represents the mean squared error. This is the predicted output of the student model given input x. This represents the teacher model's predicted output under the same input. The first term is used to make the student model's prediction approximate the teacher model's prediction. KL represents the KL divergence, used to measure the difference between two probability distributions, and softmax is the normalization exponential function. This is the temperature coefficient, which is set to 3.0 by default in this embodiment to soften the output distribution. The KL divergence weight is set to 0.5 by default in this embodiment.

[0042] Physical equation residual loss The calculation formula is: ; in, For the set of key mechanisms, This describes the residuals of the student model output after substituting the governing physics equations into the critical region r of the mechanism. The governing physics equations consist of three equations. The first is the stress equilibrium equation, applicable to the static stress field, expressed as... ,in For stress tensor, For physical density, The first is the divergence operator. The second is the heat conduction equation, applicable to the temperature field, expressed as... ,in For material density, Let T be the specific heat capacity, t be the temperature, t be the time, and k be the thermal conductivity. The third condition is the contact force equilibrium condition, applicable to contact force fields, expressed as: ,in For the total contact force, For normal contact force, The friction force is used. An automatic differential tool is employed to calculate the partial derivatives of the student model output with respect to each spatial coordinate, which are then substituted into the above equations to calculate the residuals. In this embodiment, the residuals of the stress balance equation and the residuals of the heat conduction equation are calculated for three key mechanistic regions: the switch rail close-fitting section, the frog center, and the locking transmission mechanism.

[0043] MFI constraint loss The calculation formula is: ; in, This represents the MFI value of the current student model during the distillation process. When Below When this term is activated, it generates a penalty gradient, forcing the distillation process to increase the MFI; when When the value is not less than 0.85, this item is zero and no additional constraint is generated.

[0044] Distillation training employed the Adam optimizer with a learning rate of 0.0001, a batch size of 16, and 100 training epochs. Before distillation, the MFI of the student model after pruning was 0.865; after distillation, the MFI of the student model increased to 0.892, and the mean residual of the physical equation decreased from 0.042 before distillation to 0.021, a reduction of 50%, verifying the effectiveness of mechanistic constraint distillation. Through mechanistic constraint knowledge distillation, the lightweight student model not only mimics the output of the teacher model but is also constrained by the rigid constraints of physical laws, ensuring that the prediction results of the student model conform to basic mechanical and thermal laws even under extreme conditions.

[0045] In step S4, a hardware self-test and initial matching are first performed upon edge terminal startup. The edge terminal is a trackside monitoring gateway equipped with an ARM Cortex-A78AE octa-core processor with a clock speed of 2.0GHz, 8GB of memory, and an integrated NPU accelerator with a computing power of 6 TOPS. Upon terminal startup, the number of CPU cores and clock speed, available memory, accelerator type and computing power, and current available power supply are automatically detected, and these parameters are combined to form a hardware capability vector. Multiple lightweight configuration libraries are pre-configured in the cloud. Each configuration includes the corresponding computational model pruning ratio, geometric model LOD accuracy level, and distillation loss weight, along with the required hardware requirement vector. The hardware capability vector is matched with the hardware requirement vector of each configuration using cosine similarity. The configuration with the highest cosine similarity and an MFI not lower than the MFI lower limit of 0.85 is selected as the initial loading configuration. In this embodiment, the initial loading configuration is an LOD2 geometric model paired with a 35% pruned computational model, with 3.56M parameters, a single inference time of approximately 85ms, and an MFI of 0.892.

[0046] Once the edge terminal enters operational mode, two key metrics are continuously collected: computing load rate and prediction accuracy confidence. The computing load rate is the moving average of the CPU or GPU utilization of the current inference thread, with a sliding window of 60 seconds. The prediction accuracy confidence is the normalized value of the prediction variance of the most recent N inference results. The upper limit threshold for the computing load rate is set to 80%, and the recovery threshold is set to 50%; the lower limit threshold for prediction accuracy confidence is 0.7; and the duration threshold is set to 30 seconds to prevent accidental switching due to momentary jitter.

[0047] When the computing load exceeds 80% and lasts for more than 30 seconds, a degradation switchover decision is triggered. The system scans all configurations that meet the MFI requirement of at least 0.85 from a pre-set library of multiple lightweight configurations, calculates the estimated load reduction of each configuration relative to the current configuration, and selects the configuration with the largest estimated load reduction and the highest level of lightweighting as the target configuration. The formula for calculating the estimated load reduction is as follows: ; in, To estimate the extent of load reduction, This represents the current computing load rate. The number of parameters in the current computational model. The target computational model has a set of parameters. In this embodiment, during the morning rush hour when traffic is heavy, the computing load rate rises to 87% and remains there for more than 30 seconds, triggering a downgrade switchover. The selected target configuration is a LOD3 geometric model with a 40% pruned computational model, 2.95M parameters, an estimated inference time of 65ms, and an MFI of 0.855. The switchover is performed using atomic configuration replacement: first, the weights of the newly configured model are loaded into the spare memory slot, which takes about 60ms; then, the integrity of the loading is verified; after the verification passes, the inference engine pointer is atomically switched from the current model to the new model, with the pointer switching taking less than 5ms. The entire switchover process does not interrupt inference. After the downgrade, the computing load rate drops to 62%, and the MFI decreases from 0.892 to 0.855, but is still higher than the lower limit of 0.85.

[0048] An upgrade switch is triggered when the prediction accuracy confidence level falls below 0.7 and the computing load rate falls below 50%. All configurations with an MFI (Mean Factor Index) of at least 0.85 are scanned from a pre-defined library of lightweight configurations. The estimated accuracy improvement of each configuration relative to the current configuration is calculated, and the configuration with the largest estimated accuracy improvement and the smallest lightweight level is selected as the target configuration. In this embodiment, after the morning peak, the computing load rate drops to 38%, and the prediction accuracy confidence level remains at 0.88. Since the current MFI is close to the lower limit and the load is relaxed, an upgrade switch is triggered. The target configuration is a LOD2 geometric model paired with a 35% pruned computation model, with an MFI of 0.892. After the switch, the computing load rate recovers to 48%, both within the normal range.

[0049] The edge terminal periodically uploads runtime metric logs to the cloud, once daily during the low-load period in the early morning. The runtime metric logs include load change trajectories, configuration switching frequency, and measured MFI values ​​for each configuration. Based on the accumulated logs, the cloud optimizes a pre-set multi-tiered lightweight configuration library, adjusting the pruning ratio, LOD accuracy, and distillation loss weight for each tier. In this embodiment, cloud log analysis revealed that the measured MFI value for LOD3 with 40% pruning in a high-temperature environment (rail temperature exceeding 50 degrees Celsius) was 0.84, slightly lower than the offline evaluation value of 0.855. This is mainly due to material parameter drift at high temperatures, leading to increased residuals in the physical equations. Accordingly, the cloud lowered the high-temperature MFI evaluation value for this configuration to 0.84 and added a trigger rule to the configuration library: when the ambient temperature exceeds 45 degrees Celsius, the MFI evaluation value for LOD3 with 40% pruning is 0.84; if the lower limit of MFI is 0.85, the configuration is unusable. The updated configuration library is then distributed to the edge terminal, preventing misselection of this configuration in subsequent high-temperature scenarios.

[0050] In some preferred examples, the present invention also uses the overall lightweight performance ratio as the final evaluation index, and its calculation formula is as follows: ; in, To achieve a comprehensive lightweight performance ratio, To calculate the compression ratio of the model parameters, To reduce the number of parameters in a lightweight computational model, The number of parameters for the full computation model; The compression ratio of the number of facets in the geometric model. The number of faces in a lightweight geometric model. This represents the number of faces in the full geometric model. The larger the value, the higher the fidelity of the mechanism obtained per unit compression cost, and the better the overall lightweight effect. When A value greater than 1 is considered effective lightweighting, meaning that a mechanistic fidelity exceeding 0.85 is maintained with less than half the model volume. In this embodiment, a 35% pruning is combined with a LOD2 configuration. The calculation is 0.892 divided by the sum of 0.455 and 0.178, resulting in 1.41, which is greater than 1, thus indicating effective lightweighting.

[0051] Working Principle: This invention constructs a four-level edge lightweighting process: anchoring of key mechanistic regions and MFI quantization, mechanism coupling perception-based collaborative lightweighting, mechanism constraint knowledge distillation, and scene self-aware online adaptive scheduling. This transforms the vague qualitative commitment of mechanism fidelity in traditional methods into a calculable and verifiable quantitative indicator. Specifically, MFI integrates three dimensions: physics prediction error, geometric fidelity error, and coupling consistency. It sets a minimum MFI limit for any lightweighting operation, fundamentally eliminating the risk of lightweighting damaging core physical mechanisms. Simultaneously, the heterogeneous coupling graph achieves bidirectional collaboration between computational model channel pruning and geometric model edge folding simplification, resolving the fundamental contradiction in existing technologies where computation and geometric lightweighting are disconnected, leading to geometric simplification compromising computational accuracy and computational pruning ignoring key geometric regions. Mechanism constraint knowledge distillation, through physical equation residual loss and MFI constraint loss, ensures that the lightweight student model conforms to basic mechanical and thermal laws even under extreme conditions. Online adaptive scheduling enables edge terminals to automatically switch model configurations based on real-time computing load and prediction accuracy, while remaining protected by MFI hard constraints throughout the process. This invention enables the lossless and lightweight deployment of digital twin models of turnouts on resource-constrained edge terminals.

[0052] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0053] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0054] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0055] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0056] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A mechanism-free, lossless adaptive edge lightweighting method for turnout digital twin scenarios, characterized in that, Includes the following steps: The computational and geometric models of the turnout digital twin are divided into critical and non-critical regions, and a mechanism protection weight is assigned to each region. The Mechanism Fidelity Index (MFI) is defined as a unified quantitative constraint boundary for lightweight operations. The MFI integrates three dimensions: physical field prediction error, geometric fidelity error, and coupling consistency. A lower limit value for the MFI is set, and the MFI after any lightweight operation must not be lower than the lower limit value of the MFI. A heterogeneous coupling graph containing physical field nodes, geometric nodes, and computational nodes is constructed. Based on the heterogeneous coupling graph, the constraint channels of the computational model are pruned, and the weighted edges of the geometric model are simplified by folding. Constraints are bidirectionally transmitted through the heterogeneous coupling graph: when the mesh quality decreases after geometric simplification, the importance score of the associated computational channels is corrected. After calculating the channel pruning, feedback is provided to improve the folding cost of the associated geometric region; Using the pruned computational model as the student model and the full high-fidelity computational model as the teacher model, a joint distillation loss function is constructed, which includes data fitting loss, physical equation residual loss, and MFI constraint loss, and the student model is fine-tuned and trained. The edge terminal monitors the computing load rate and prediction accuracy confidence during operation, and automatically switches the model configuration under the constraint that the MFI is not lower than the MFI lower limit.

2. The mechanism-insensitive adaptive edge lightweighting method for turnout digital twin scenarios according to claim 1, characterized in that, The key areas of the mechanism include the switch rail close-fitting section, the frog core area, the locking transmission mechanism, the slide plate contact area, and the actuating rod connection node; The non-critical area includes the equipment base, outer protective cover, and auxiliary cabinet; The mechanism protection weight of the critical mechanism region is greater than that of the non-critical mechanism region; The MFI is the weighted sum of the physical field prediction error ratio, the geometric fidelity error, and the coupling consistency.

3. The mechanism-insensitive adaptive edge lightweighting method for turnout digital twin scenarios according to claim 1, characterized in that, The mechanism-constrained channel pruning specifically refers to: For each convolutional channel, a Taylor number is calculated as the initial importance. A mechanism loss regularization term is introduced to correct the importance of channels connected to the mechanism key region. The correction method is to add the mechanism protection weight of the physical field region connected to the channel and the physical field prediction error increment of the physical field region after removing the channel to the initial Taylor number to obtain the corrected channel importance. The channels are sorted according to their corrected importance, and the redundant channels with the lowest contribution are removed. The pruning ratio is gradually increased from a conservative ratio. The MFI is calculated after each round of pruning. If the MFI is lower than the MFI lower limit, the result is reverted to the previous round of pruning.

4. The mechanism-insensitive adaptive edge lightweighting method for turnout digital twin scenarios according to claim 1, characterized in that, The specific method for simplifying the weighted edge folding mechanism is as follows: The mechanism protection weights are mapped to mesh vertices, and the edge folding cost function is improved so that the folding cost increases as the mechanism protection weights of the region where the vertex is located increase. The vertex folding cost of the mechanistic critical region is amplified to limit the mesh simplification of the mechanistic critical region, while the vertex folding cost of the non-critical region is not amplified to allow the mesh simplification of the non-critical region to be greater than that of the mechanistic critical region.

5. The mechanism-insensitive adaptive edge lightweighting method for turnout digital twin scenarios according to claim 1, characterized in that, The bidirectional transmission of constraints through the heterogeneous coupling graph specifically refers to: After geometric simplification, the mesh quality score of each geometric node is updated. When the consistency of the face normal vector in the geometric critical region decreases beyond the preset threshold or the maximum aspect ratio of the face exceeds the preset threshold, a degradation alarm is sent to the connected physical field node and computing node, triggering the re-evaluation of the importance of the corresponding computing channel and the channel recovery operation. After calculating the network pruning, the distribution of geometric regions connected by the pruned channels is statistically analyzed. When the number of associated channels pruned in a geometric region exceeds a preset threshold, a multiplication correction factor greater than 1 is applied to the folding cost of the geometric region, so that the increase in folding cost of the geometric region in subsequent folding iterations is greater than that of other geometric regions.

6. The mechanism-free adaptive edge lightweighting method for turnout digital twin scenarios according to claim 1, characterized in that, The calculation of the residual loss of the physical equation is as follows: In the key mechanism region, the output of the student model is substituted into the stress balance equation, heat conduction equation and contact force balance condition to calculate the weighted sum of squares of the residuals, where the weight of each region is the mechanism protection weight of the region. And / or, the calculation of the MFI constraint loss is specifically as follows: when the current MFI in the distillation process is lower than the MFI lower limit, the penalty term is activated; otherwise, it is zero.

7. The mechanism-insensitive adaptive edge lightweighting method for turnout digital twin scenarios according to claim 1, characterized in that, The method also includes: When the edge terminal starts up, it performs a hardware self-test to detect the number of CPU cores and clock speed, available memory, accelerator type and computing power, and available power supply, forming a hardware capability vector. The hardware capability vector is matched with the hardware requirement vector in the preset multi-level lightweight configuration library using cosine similarity. The configuration with the highest similarity and satisfying that the MFI is not lower than the MFI lower limit is selected as the initial loading configuration.

8. The mechanism-insensitive adaptive edge lightweighting method for turnout digital twin scenarios according to claim 1, characterized in that, The automatic switching of model configuration under the constraint that MFI is not lower than the lower limit of MFI specifically refers to: When the computing load rate exceeds the upper limit threshold and the duration exceeds the preset time threshold, a downgrade switch is triggered. From the preset multi-level lightweight configuration library, a configuration is selected that satisfies the MFI not being lower than the lower limit of the MFI, has the largest estimated load reduction, and has a greater degree of lightweighting. When the prediction accuracy confidence level is lower than the lower limit threshold and the computing power load rate is lower than the recovery threshold, an upgrade switch is triggered. From the preset multi-level lightweight configuration library, a configuration is selected that satisfies the condition that the MFI is not lower than the lower limit value of the MFI, has the largest improvement in prediction accuracy, and has a smaller degree of lightweighting. The switching execution adopts atomic configuration replacement. First, the newly configured model weights are loaded into the spare memory slot. After verifying the integrity, the inference engine pointer is atomically switched from the current model to the new model. The inference process is not interrupted during the switching process.

9. The mechanism-insensitive adaptive edge lightweighting method for turnout digital twin scenarios according to claim 1, characterized in that, The method also includes: The edge terminal periodically uploads runtime metric logs to the cloud. The runtime metric logs include load change trajectories, configuration switching frequency, and measured MFI values ​​under each configuration. The cloud-based system optimizes a preset multi-level lightweight configuration library based on accumulated logs, adjusts the pruning ratio, LOD accuracy, and distillation loss weight for each level, and then distributes the updated preset multi-level lightweight configuration library to the edge terminal.

10. A mechanism-free, lossless adaptive edge lightweighting method for turnout digital twin scenarios according to claim 1, characterized in that, The method also includes: Calculate the overall lightweight performance ratio, which is the ratio of MFI to the sum of the calculated parameter compression ratio and the geometric patch number compression ratio; When the overall lightweight efficiency ratio is greater than 1, it is determined to be effective lightweighting.

Images