Intelligent cable monitoring method, edge computing unit and monitoring system

Abstract

This invention relates to the field of bridge cable monitoring technology. It provides an intelligent cable monitoring method, an edge computing unit, and a monitoring system to acquire monitoring data of intelligent cables. The monitoring data is then input into an AI monitoring model to obtain monitoring results. The AI ​​monitoring model determines the pruning strategy for each convolutional layer based on the hierarchical feature attributes of historical cable signals. By pruning the convolutional layers, this invention effectively reduces processing latency, decreases dependence on communication networks, and alleviates data storage and transmission pressure. Furthermore, by matching the signal feature types corresponding to different convolutional layers for hierarchical constraint pruning, this invention effectively preserves the ability to extract abnormal cable mutation features, avoids model structural collapse, and ensures the accuracy and stability of cable monitoring and identification while achieving lightweight model compression.

Description

Technical Field

[0001] This invention relates to the field of bridge cable monitoring technology, and in particular to an intelligent cable monitoring method, edge computing unit, and monitoring system. Background Technology

[0002] As the core load-bearing component of cable-stayed bridges, the health of bridge cables directly affects the overall structural safety of the bridge. Currently, smart cable technology based on fiber optic grating sensors has been widely used, enabling online monitoring of cable forces by pre-embedding sensors inside the cables.

[0003] However, existing intelligent cable monitoring systems still suffer from the following technical shortcomings in practical engineering applications: First, high data processing latency fails to meet the real-time early warning requirements for transient events. Existing systems mostly adopt a "field acquisition-remote transmission-cloud processing" model, requiring massive amounts of raw sensor data to be transmitted over long distances to cloud servers for analysis and processing. This results in long early warning response times (typically minutes), making it difficult to meet the real-time early warning needs for transient events such as sudden cable breakage, vehicle impacts, and strong winds. Second, strong dependence on communication networks limits application in remote areas. Many in-service bridges are located in remote mountainous areas or network coverage blind spots. When communication is interrupted, field monitoring data cannot be uploaded to the cloud, causing the monitoring system to fail, creating a monitoring vacuum period and posing serious safety hazards. Third, anomaly identification relies on expert experience, leading to a high false alarm rate. Fiber optic grating signals contain various interference factors (temperature drift, random vehicle load, environmental noise, etc.), making it difficult for traditional threshold alarm methods to effectively distinguish between real anomalies and interference signals, resulting in a high false alarm rate. Furthermore, manual interpretation based on waveform characteristics requires experienced experts and cannot achieve real-time on-site diagnosis. Fourth, high data transmission and storage pressure. A single cable-stayed bridge may contain hundreds of cables, each of which requires continuous data collection at frequencies of hundreds of hertz. This round-the-clock monitoring generates massive amounts of data, placing enormous pressure on transmission bandwidth and cloud storage, and resulting in high data maintenance costs. Therefore, existing monitoring methods struggle to achieve real-time on-site diagnosis of the cables. Summary of the Invention

[0004] This invention provides an intelligent cable monitoring method, an edge computing unit, and a monitoring system to address the problem that existing monitoring methods struggle to achieve real-time on-site diagnosis of cables.

[0005] A first aspect of the present invention provides an intelligent cable monitoring method, comprising: Acquire monitoring data from the smart cable; The monitoring data is input into the AI ​​monitoring model to obtain the monitoring results; Among them, the AI ​​monitoring model determines the pruning strategy of each convolutional layer based on the hierarchical feature attributes of historical cable signals.

[0006] In one possible implementation, the AI ​​monitoring model is constructed through the following steps: Collect cable tension monitoring data under various working conditions to construct a training set; A convolutional neural network for one-dimensional time-series strain signals of cables was established; Based on the hierarchical feature attributes of historical LAS signals and the L1 norm of convolutional kernel weights, all convolutional layers of the current layer of the convolutional neural network are evaluated to obtain the importance score of each convolutional layer; among them, the hierarchical feature attributes include edge mutation features and periodic semantic features. Following the order from shallow to deep layers, each convolutional layer is pruned within its layer according to its importance score. After pruning each convolutional layer, the parameters of that layer and subsequent layers are updated immediately using the training set until the overall pruning rate meets the target pruning rate. The pruned model is dynamically quantified to obtain the AI ​​monitoring model.

[0007] In one possible implementation, intra-layer pruning is performed on each convolutional layer according to its importance score, from shallowest to deepest, including: For each convolutional layer, based on the descending sorting results of the comprehensive importance scores within that layer, an independent pruning threshold and target pruning rate are set within the layer. Redundant convolutional kernels whose overall importance score is lower than the pruning threshold of that layer are removed.

[0008] In one possible implementation, for each convolutional layer, based on the descending sorting result of the comprehensive importance score within that layer, an independent pruning threshold and target pruning rate are set, including: Sort all convolutional kernels in the current convolutional layer in descending order of their overall importance score; Based on the feature extraction function and redundancy of the current convolutional layer, determine the independent target pruning rate for that layer; The pruning threshold of the convolutional layer is determined based on the alignment requirements of the output channels of the edge computing unit, the ranking results of the comprehensive importance score, and the target pruning rate.

[0009] In one possible implementation, the pruned model is dynamically quantized to obtain an AI monitoring model, including: The pruned model weights are subjected to symmetric quantization, and the activation values ​​during model runtime are subjected to window-by-window dynamic quantization to obtain the AI ​​monitoring model.

[0010] In one possible implementation, the method also includes: Receive model differential update instructions from the cloud; the differential update instructions only include local weight change information of the convolutional layer to be modified; Based on the total number of parameters of the convolutional layer to be updated, allocate a shadow buffer in the edge computing unit that is the same size as the parameters of that layer. Copy all parameters of the convolutional layer to be updated from the main model storage area to the shadow buffer; The target weights in the shadow buffer are locally modified according to the differential update instruction, and the access pointer of the main model area is pointed to the updated shadow buffer through the atomic switching operation, so that the system can immediately load the new parameters. Release the memory space occupied by the original convolutional layer.

[0011] In one possible implementation, before inputting monitoring data into the AI ​​monitoring model to obtain monitoring results, the method includes: Real-time temperature compensation is performed on the monitoring data based on the rate of change of ambient temperature and the strain-temperature compensation model.

[0012] In one possible implementation, after acquiring the monitoring data of the smart cable, the method further includes: The monitoring data is divided into multiple data segments according to a preset window size; A sliding window approach is used to analyze the real-time acquired monitoring data segment by segment to obtain the characteristic data within each sliding window; the size of the sliding window is set according to the characteristics of the cable strain signal. The diagnostic type of the monitoring results is determined based on the feature data within each sliding window.

[0013] A second aspect of the present invention provides an edge computing unit including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the intelligent cable monitoring method of the first aspect above.

[0014] A third aspect of the present invention provides an intelligent cable monitoring system, comprising an intelligent cable and an edge computing unit as described in the second aspect above.

[0015] Compared to traditional technologies, this invention provides an intelligent cable monitoring method, edge computing unit, and monitoring system to acquire monitoring data of intelligent cables. The monitoring data is then input into an AI monitoring model to obtain monitoring results. The AI ​​monitoring model determines the pruning strategy for each convolutional layer based on the hierarchical feature attributes of historical cable signals. By pruning the convolutional layers, this invention effectively reduces processing latency, decreases dependence on communication networks, and alleviates data storage and transmission pressure. Furthermore, by matching the signal feature types corresponding to different convolutional layers for hierarchical constraint pruning, this invention effectively preserves the ability to extract abnormal cable mutation features, preventing model structural collapse. This achieves lightweight model compression while ensuring the accuracy and stability of cable monitoring and identification. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the implementation of the intelligent cable monitoring method provided in this embodiment of the invention; Figure 2 This is a schematic diagram of the structure of the edge computing unit provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the intelligent cable monitoring system provided in an embodiment of the present invention. Detailed Implementation

[0017] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0018] Figure 1 This is a flowchart illustrating the implementation of the intelligent cable monitoring method provided in this embodiment of the invention. Figure 1 As shown, the intelligent cable monitoring method includes: S110, acquires monitoring data of the smart cable; S120: Input the monitoring data into the AI ​​monitoring model to obtain the monitoring results; Among them, the AI ​​monitoring model determines the pruning strategy of each convolutional layer based on the hierarchical feature attributes of historical cable signals.

[0019] In this embodiment of the invention, the monitoring data consists of cable strain time history data and ambient temperature data obtained by a fiber optic grating sensor array embedded in the intelligent cable body or anchor through multi-point distributed acquisition using wavelength multiplexing technology; digital strain time domain signal output after demodulation and conversion by the optical signal demodulation module in the edge computing unit using a swept-frequency laser light source and a photodetector; original sensing time domain waveform data covering various typical working conditions such as normal vehicle traffic, temperature drift and sudden temperature change, wire breakage, anchor slippage, vehicle / falling object impact load, and wind-induced vibration; and auxiliary identification information such as sensor number, cable number, sampling timestamp, sampling frequency, GPS positioning location, and data integrity check code.

[0020] The AI ​​monitoring model is a lightweight diagnostic model built on a one-dimensional convolutional neural network and quantized perception. The model has built-in temperature compensation and environmental adaptive preprocessing logic, can run on an embedded chip in an edge computing unit, has a memory footprint of <512KB, and a single inference time of <10ms. It can perform end-to-end working condition classification on the input strain time history window segment and output the probability distribution of five states: normal, wire breakage, anchor slippage, impact load, and wind-induced vibration. It supports differential hot updating of the model and has high-precision identification capability for transient anomalies such as wire breakage.

[0021] The monitoring results are specifically classified as cable working condition results, including five types: normal state, abnormal wire breakage, abnormal anchor slippage, abnormal impact load, and abnormal wind-induced vibration. They include the corresponding anomaly confidence level, occurrence timestamp, cable number, and anomaly location. They are categorized into Level 1 and Level 2 warnings based on severity. The results also include pre-processed calibrated strain values, dynamic baseline values, temperature compensation values, and signal fluctuation standard deviations as auxiliary diagnostic data. These results can be directly used for local audible and visual alarms, on-site terminal push notifications, and synchronization with the cloud platform.

[0022] In some embodiments, the training process of the AI ​​monitoring model is as follows: Cable stress monitoring data under various working conditions are collected to construct a training set; a convolutional neural network is established for one-dimensional time-series strain signals of cables; based on the hierarchical feature attributes of historical cable signals and the L1 norm of the convolutional kernel weights, all convolutional layers of the current layer of the convolutional neural network are evaluated to obtain the importance score of each convolutional layer; wherein, the hierarchical feature attributes include edge mutation features and periodic semantic features; according to the importance score, each convolutional layer is pruned in an order from shallow to deep layers, and after pruning each convolutional layer, the parameters of that layer and subsequent layers are immediately updated using the training set until the overall pruning rate meets the target pruning rate; the pruned model is dynamically quantized to obtain the AI ​​monitoring model.

[0023] In this embodiment of the invention, the training steps of the AI ​​monitoring model can be as follows: Step 1: Building and deploying lightweight AI models at the edge.

[0024] 1.1 Construction of multi-condition samples.

[0025] Historical cable stress monitoring data under different working conditions were collected to construct a training sample set, including: normal vehicle traffic; temperature changes (daily temperature difference, seasonal temperature difference); cable breakage (simulated cable breakage test); anchorage slippage; impact loads (such as heavy vehicle passage, falling object impact); and wind-induced vibration. Each working condition was labeled to construct a training dataset containing no less than 100,000 samples.

[0026] 1.2 Training of a benchmark model for one-dimensional time-series signals.

[0027] Construct an initial convolutional neural network, using a one-dimensional convolutional structure to adapt to the temporal characteristics of the cable strain signal: Input layer: 1×256 (256 consecutive sampling points); Convolutional layer 1: 32 3×1 convolutional kernels, ReLU activated; Pooling layer 1: 2×1 max pooling; Convolutional layer 2: 64 3×1 convolutional kernels, ReLU activated; Pooling layer 2: 2×1 max pooling; Fully connected layer: 128 neurons, Dropout=0.5; Output layer: Softmax, number of output categories = 5 (normal, broken wire, slippage, impact, vibration).

[0028] Train until convergence using the cross-entropy loss function and the Adam optimizer.

[0029] 1.3 Gradual pruning and compression.

[0030] To deploy the model to memory-constrained embedded chips, this invention proposes a "quantization-aware progressive pruning" method to compress the baseline model, specifically including the following steps: 1.3.1 Quantitative sensitivity assessment.

[0031] Before pruning, a quantization sensitivity assessment is first performed on each convolutional kernel to identify which weights have the greatest impact on the accuracy of the quantized model. The quantization sensitivity is calculated as follows: (1) Based on the distribution of all weights w in the baseline model, calculate the quantization factor scale=max(|w|) / 127, where 127 is the effective saturation mapping upper limit of an 8-bit signed integer; (2) For each weight w, simulate the quantization process: q = round(w / scale), restore the value. =q×scale, calculate the quantization error error=|w- |; where round() is the rounding function and scale is the quantization scaling factor (scale), which represents the proportional mapping relationship between the original floating-point value and the quantized integer.

[0032] (3) Combine the gradient g recorded during training (which represents the degree of influence of the weight on the loss function, i.e., the partial derivative of the loss function with respect to the weight) to calculate the quantized contribution of each weight: contribution = error × |g|. (4) Sum the quantization contributions of all weights within the same convolution kernel to obtain the quantization sensitivity score Sensitivity(K) = Σ| - recover( )| × | |. Among them, For the first i Each weight, For the first i The gradient corresponding to each weight; The score reflects the degree of impact of the convolution kernel on the model's recognition accuracy after quantization: the higher the score, the more sensitive the convolution kernel is to quantization, and the greater the accuracy loss after quantization.

[0033] 1.3.2 Gradual pruning.

[0034] Unlike the one-time global pruning in existing technologies, this invention adopts a strategy of pruning layer by layer from shallow to deep, with each layer processed independently: (1) In order to take into account both the original size of the weights and the accuracy loss after quantization during the pruning process, this invention defines a comprehensive importance score Score(K) for each convolutional kernel.

[0035] For each convolutional layer, calculate the overall importance score of all convolutional kernels in that layer: Score(K) = ×(1-α×Sensitivity(K) / ) in, α is the L1 norm of the convolution kernel weights (reflecting the physical contribution of the convolution kernel in the current feature extraction), Sensitivity(K) is the quantization sensitivity score, and α is the balance coefficient (set to 0.5 in this invention). The maximum sensitivity score among all convolutional kernels in this layer is used for normalization.

[0036] (2) It should be noted that the one-time global pruning commonly used in the existing technology deletes convolution kernels according to the unified threshold of the entire network, which is very likely to cause the feature extraction structure of a certain layer to be excessively destroyed (for example, the shallow convolution kernels responsible for detecting high-frequency jumps of broken wires are mistakenly deleted in a large area), causing the network structure to "collapse", and subsequent fine-tuning cannot restore the accuracy.

[0037] The progressive pruning strategy employed in this invention, deeply integrated with the aforementioned comprehensive score Score(K), possesses the following three irreplaceable technical advantages: ① Layered feature protection for cable signals: Shallow convolutional kernels extract edge abrupt change information (such as broken wire steps), while deep convolutional kernels extract semantic patterns (such as vehicle periodic vibrations). Layer-by-layer pruning allows for independent calculation of Score(K) within each layer and determination of the retention threshold for that layer, avoiding the risk of widespread accidental deletion of shallow key convolutional kernels due to their generally small absolute weight values ​​under a globally uniform threshold. For key convolutional kernels responsible for extracting transient jump features of broken wires, their absolute weight values ​​are usually extremely small, but their quantization sensitivity is extremely high. In the calculation of the comprehensive importance score Score(K), high sensitivity will cause the penalty factor to significantly reduce the absolute score of the kernel, but it is precisely this characteristic that allows the system to clearly distinguish it from redundant kernels with the same weight scale but low quantization sensitivity during intra-layer ranking. With the subsequent threshold fine-tuning and hardware alignment mechanism, the system can actively relax the retention threshold in a local range, forcibly including these key convolutional kernels with small weights but high quantization sensitivity into the retention list, thereby achieving model compression while fully preserving the ability to extract abnormal abrupt change features of cables.

[0038] ② Supports differentiated compression ratio control between layers: The redundancy of different layers varies greatly. Pruning is performed using the relative ranking within each layer based on the Score(K), dynamically determining the pruning rate of each layer according to its actual situation, rather than forcibly assigning a uniform ratio. This ensures the sparsity differentiation between the first layer (waveform receiving layer) and the last layer (classification layer), keeping the accuracy loss within 1% while maintaining an inference speed of <10ms.

[0039] Shallow convolutional layers directly process the 1kHz strain time history signal demodulated by the fiber grating, responsible for extracting the transient jump features caused by wire breakage ("transient jump features" refer to the abrupt amplitude change and extremely short duration (usually ≤10ms) of the strain time history formed by the instantaneous breakage of the cable wire, which is different from the smooth gradual change caused by temperature drift or vehicle load). They have a small number of effective convolutional kernels and are irreplaceable for detecting cable tension anomalies. Deep convolutional layers only perform classification decisions on the abstracted features, resulting in significant redundancy. This invention employs a layer-by-layer pruning strategy, independently evaluating each layer based on a quantitatively perceived comprehensive importance score (Score(K)). This keeps the compression rate of shallow convolutional layers between 50% and 60%, ensuring that all wire breakage sensitive kernels are retained; the compression rate of deep convolutional layers is controlled between 70% and 80%; the overall model pruning rate can ultimately reach 65%-75%, while maintaining a wire breakage detection accuracy of over 98% under these compression conditions.

[0040] ③ Seamless Coupling with Hardware Alignment Constraints: Layer-by-layer pruning provides space for "flexible adjustment of pruning thresholds" within a single layer. When the hardware constraint of "the number of channels must be a multiple of 4" needs to be met, the retention threshold based on the comprehensive importance score Score(K) can be fine-tuned within that layer (for example, when the target retention number of 13 does not meet the alignment requirement, the threshold is relaxed to retain the top 16). This achieves an optimal trade-off between accuracy and speed locally without affecting the structure of other layers. Requiring the number of convolutional kernels to be a multiple of 4 ensures that the algorithm can perfectly fill the chip's parallel computing pipeline, avoiding idle valuable computing resources. This achieves "load once, process in parallel".

[0041] Assuming a convolutional layer originally has 32 convolutional kernels, with an overall target compression rate of 60%, conventional calculations would require retaining approximately 12.8 kernels, which rounds to 13.

[0042] Traditional pruning methods simply retain the top 13 results based on the global sort. However, 13 is not a multiple of 4, which fails to meet the hardware alignment requirements of single-instruction multiple-data parallel computing on the chip. In this case, only 13 results can be forcibly retained, leading to "bubbles" in the inference pipeline and limiting inference time to 12ms, which cannot meet the real-time alert requirement of <10ms.

[0043] The layer-by-layer pruning scheme of this invention (hardware-aware): Within each layer, the system does not rigidly select the top 13, but actively relaxes the retention threshold based on the comprehensive importance score Score(K), slightly adjusting the number of retained models from 13 to 16 (a multiple of 4). At this point, although three more convolutional kernels are retained, the accuracy slightly decreases from the theoretically highest 98.7% to 98.5%, but the model structure perfectly aligns with the channel requirements of the hardware acceleration unit, and the inference time can be optimized to 9ms. After a subsequent round of real-time fine-tuning, the accuracy can quickly recover to above 98.5%.

[0044] The position marked by the "line" on the Score(K) sorting list is the threshold. To ensure that the number of items retained is a multiple of 4, this line is moved up or down slightly; this is called "threshold fine-tuning".

[0045] The edge computing box chip can process 4 numbers in parallel at a time, requiring the number of convolution kernels to ideally be a multiple of 4. Layer-by-layer pruning can flexibly adjust the number of kernels retained within each layer from 13 to 12 or 16, sacrificing 0.2% accuracy for a 33% speed improvement, ensuring that the wire break alarm is triggered within 10 milliseconds, a local fine-tuning that global pruning cannot achieve.

[0046] In some embodiments, in-layer pruning is performed on each convolutional layer according to its importance score, from shallow to deep layers. This includes: for each convolutional layer, setting an independent pruning threshold and target pruning rate based on the descending order of the overall importance score within that layer; and deleting redundant convolutional kernels whose overall importance score is lower than the pruning threshold of that layer.

[0047] (3) Steps for pruning branches layer by layer.

[0048] In this embodiment of the invention, based on quantitative sensitivity assessment, comprehensive importance score calculation, and hardware-aware constraint principles, the invention performs a progressive pruning process on the convolutional neural network from shallow to deep layers, processing each layer independently and fine-tuning each layer, specifically including: ① Evaluation and Ranking. For all convolutional kernels in the currently processed convolutional layer, a unified comprehensive importance score Score(K) is calculated. This score integrates the L1 norm of the convolutional kernel weights, the quantization sensitivity score, and the hardware alignment reward. All convolutional kernels are then sorted in descending order of Score(K) value. The higher the score, the more important the convolutional kernel is for extracting anomalous features and improving the model's recognition accuracy, and the more priority it should be given to retention.

[0049] ② Intra-layer pruning. Determine the pruning threshold based on the preset intra-layer independent target pruning rate for the current layer, delete convolutional kernels whose overall importance score (Score(K)) is lower than the threshold of the layer, and retain only high-scoring core convolutional kernels; Suppose a convolutional layer has 32 kernels, and the target pruning rate is set to 60%, meaning approximately 12.8 kernels will be retained. Calculate the Score(K) for all 32 kernels, then sort them from highest to lowest (Score(K) values: 0.95, 0.88…0.52, 0.49, 0.44, 0.41, 0.38, 0.33…). Retain 12.8 kernels, rounding down to a preliminary target of 13. "Retaining the top 13" means setting the threshold between the 13th and 14th ranked kernels, approximately between 0.44 and 0.49. Kernel with a Score(K) ≥ 0.49 are retained, while those below 0.49 are pruned.

[0050] The key steps in “threshold fine-tuning”: Option A (Tightening the threshold): Increase the threshold from between 0.44 and 0.49 to between 0.49 and 0.52. This will cause core #21 (Score=0.49) to fall below the new threshold and be pruned. Ultimately, 12 cores will be retained.

[0051] Option B (Relaxed Threshold): Lower the threshold from 0.44–0.49 to 0.33–0.38. This includes cores #21, #8, #29, and #12 (with scores from 0.49 to 0.38). Ultimately, 16 cores are retained.

[0052] Strictly ensure the following during pruning: Convolutional kernels with large weights but large errors after quantization are preferentially retained. For example, key convolutional kernels responsible for extracting transient jump features with broken fibers, and with a weight of only 0.01 but highly sensitive to quantization (although the weight is small, its role is extremely important, and its function is easily destroyed during quantization. When calculating Score(K), a "quantization sensitivity score" is added to its score. Even if its original weight is small, this addition will increase its total score, ultimately ranking it higher in the sorting, thus "forcibly retained."), will be forcibly retained due to the quantization sensitivity score. Convolutional kernels with small weights and small errors after quantization are preferentially pruned. These kernels contribute little to anomaly detection and have high redundancy. Deleting them will not cause a significant decrease in model accuracy. Suppose that in the trained model, the 5th convolutional kernel is responsible for detecting a very rare random small vibration that is neither a broken wire nor a vehicle impact. Its three weight values ​​are: [0.0003, -0.0001, 0.0004].

[0053] Small weights: Its L1 norm is only 0.0008, which is almost negligible among all weights.

[0054] Small error after quantization: When performing simulated quantization, the scale is assumed to be 0.001.

[0055] Quantification: 0.0003 / 0.001 = 0.3, rounded to the nearest integer → q = 0.

[0056] Restore: 0 × 0.001 = 0.

[0057] Error: |0.0003 - 0| = 0.0003.

[0058] After summing all weights, the quantification contribution is extremely low.

[0059] analyze: The value of this convolution kernel is almost zero, and after quantization, it is rounded to zero, resulting in negligible error. This means it contributes almost nothing to the model's output, and removing it would cause minimal loss to the model's accuracy, so it would naturally be prioritized for elimination during sorting.

[0060] Meanwhile, hardware alignment constraints are enforced during pruning, adjusting the number of retained convolutional kernels to a multiple of 4 to adapt to the SIMD parallel computing instructions of ARM chips and ensure that the inference speed meets the millisecond-level requirements.

[0061] ③ Instant fine-tuning. After completing the pruning operation of the current convolutional layer, without waiting for all layers to be pruned, the pre-constructed multi-condition training dataset is immediately used to update the parameters and fine-tune the model of the current pruned layer and all subsequent network layers. This quickly repairs the accuracy loss caused by the removal of convolutional kernels, enabling the model to quickly restore stable recognition capabilities after pruning.

[0062] Assume the baseline model has 3 convolutional layers: Convolutional layer 1: 32 convolutional kernels, responsible for extracting edge abrupt change information (such as broken wire steps).

[0063] Convolutional layer 2: 64 convolutional kernels, responsible for combining mid-layer texture features.

[0064] Convolutional layer 3: 128 convolutional kernels, responsible for abstract semantic classification.

[0065] Now, following the order from shallowest to deepest, we will first prune convolutional layer 1.

[0066] Step 1: Pruning.

[0067] Convolutional layer 1 originally had 32 kernels → sorted by Score(K) → redundant kernels were removed → 12 kernels were retained (to meet the alignment requirement of multiples of 4).

[0068] The model structure now becomes: Convolutional layer 1: 12 kernels (just pruned, parameters unchanged).

[0069] Convolutional layer 2: 64 kernels (unchanged).

[0070] Convolutional layer 3: 128 kernels (unchanged).

[0071] Step 2: The accuracy drops immediately.

[0072] Because convolutional layer 1 had 20 fewer kernels, some feature extraction capabilities were instantly lost, and the overall recognition accuracy of the model dropped from 98.5% before pruning to 92.3%.

[0073] Step 3: Instant fine-tuning.

[0074] Without pruning convolutional layers 2 and 3, immediately fine-tune the model using the training set. However, the fine-tuning range is selective: The existing parameters of layers 2 and 3 are frozen (because they have not been pruned yet, they do not need to be adjusted).

[0075] Only train the weights of the remaining 12 kernels in layer 1, and allow minor adaptive adjustments to the parameters of layers 2 and 3.

[0076] ④ Iterative progression. After the current layer completes pruning and fine-tuning and converges, the same evaluation, sorting, pruning, and fine-tuning process is performed on the next convolutional layer in the order from shallow to deep layers. The above steps are repeated layer by layer until the overall model pruning rate of the entire network reaches 65%-75%, completing the entire compression process and avoiding the model structure "collapse" problem caused by traditional one-time pruning.

[0077] In some embodiments, for each convolutional layer, based on the descending sorting result of the comprehensive importance score within the layer, an independent pruning threshold and target pruning rate are set within the layer, including: sorting all convolutional kernels within the current convolutional layer in descending order according to the comprehensive importance score; determining the independent target pruning rate of the current convolutional layer based on the feature extraction function and redundancy of the current convolutional layer; and determining the pruning threshold of the convolutional layer based on the alignment requirement of the output channel number of the edge computing unit, the sorting result of the comprehensive importance score, and the target pruning rate.

[0078] In this embodiment of the invention, hardware-aware constraints are introduced during the pruning process to ensure that the pruned model structure can perfectly adapt to the instruction set optimization requirements of the target embedded chip. The specific implementation method is described in detail below: (1) The hardware constraints are clear.

[0079] For the embedded chips used in edge computing units, the supporting neural network library has a hard optimization rule: the number of convolutional kernels output by each convolutional layer must be a multiple of 4. Only when this condition is met can the chip enable the SIMD single instruction multiple data stream parallel acceleration mechanism to achieve "one-time loading, four-way parallel computing", maximize the use of the computing pipeline, avoid idle hardware resources and wasted computing cycles, and thus ensure that the inference time is strictly controlled within 10ms, meeting the real-time early warning requirements for transient anomalies such as wire breakage.

[0080] (2) Embed hardware alignment constraints into pruning decisions and comprehensive score calculation.

[0081] In the pruning decision-making stage, this invention aligns the number of convolutional kernels to a multiple of 4 as a core constraint condition and directly incorporates it into the calculation process of the comprehensive importance score of the convolutional kernels, so that the pruning algorithm actively prefers the solution that meets the hardware requirements.

[0082] Original formula for composite score: Score(K) = ×(1 α×Normalization sensitivity) The new overall score formula after incorporating hardware alignment bonuses: Score(K) = ×(1 α × normalized sensitivity) + β × alignment reward in, α is the L1 norm of the convolution kernel weights, representing the feature extraction intensity; α is the quantization sensitivity balance coefficient, fixed at 0.5; normalized sensitivity is the ratio of the quantization sensitivity score of the current convolution kernel to the maximum sensitivity score within the layer; β is the hardware alignment reward coefficient. The alignment reward is a binary reward item: if the current pruning scheme can make the number of convolution kernels in this layer a multiple of 4, then a positive reward score is given; if it cannot be aligned to a multiple of 4, no reward is given, or even a negative penalty score is imposed.

[0083] The typical range of the original overall importance score, Score_origin(K), is between 0.001 and 1.0, with a mean of 0.2 to 0.5. If the value of β exceeds 0.2, the contribution of the reward item (maximum 0.2 × 1 = 0.2) will exceed the mean of the original score. This causes low-importance convolutional kernels to be artificially inflated due to the reward score. The pruning algorithm retains a large number of redundant kernels, resulting in a decrease in model compression ratio, no significant improvement in inference speed, and an accuracy loss of more than 2.5%.

[0084] The method for determining β includes the following steps: Step 1, test β=0 on the validation set as a baseline, and record the accuracy and inference latency; Step 2, perform a grid search in the interval [0, 0.2] with a step size of 0.01, perform hardware-aware pruning for each β value, and record the accuracy retention rate and the proportion of layers that satisfy the 4-fold constraint; Step 3, select the β value that meets the accuracy retention requirement, achieves a 4-fold constraint satisfaction rate of over 90%, and is at the speed gain saturation inflection point (the inference speed improvement brought by increasing the β value is less than 5%) as the final setting. For example, in this experiment, β=0.05.

[0085] Through the above score reconstruction, the pruning algorithm will automatically prioritize pruning schemes that can make the number of retained convolutional kernels a multiple of 4 during the sorting and filtering process, thus ensuring hardware friendliness from the algorithm level.

[0086] (3) Forced correction strategy when alignment fails.

[0087] If the number of target convolutional kernels calculated based on the original pruning threshold cannot be directly aligned to a multiple of 4, this invention employs two complementary methods for forced adjustment to ensure that the final structure meets hardware acceleration requirements: Method 1: Fine-tuning the pruning threshold. Slightly increase or decrease the in-layer pruning threshold, ensuring that key high-scoring convolutional kernels are not deleted, and forcibly adjust the number of kernels retained to multiples of 4 such as 12, 16, or 20. This achieves an inference speed improvement of over 30% with minimal accuracy loss (typically less than 0.2%). See Table 1 for details.

[0088] Table 1. Convolutional Kernel Ranking Table

[0089] All "key" convolutional kernels have a score (K) ≥ 0.55, ranking in the top 10.

[0090] The 11th to 32nd ranked kernels are all ordinary convolution kernels, with a score (K) ≤ 0.47.

[0091] Whether the threshold is adjusted upwards or downwards, it must never exceed 0.55 for the 10th ranked core. Exceeding this threshold would inadvertently harm the lowest-ranked critical core (such as core #3 with a score of 0.55).

[0092] "Key high-scoring convolutional kernels" refer to the top 10 kernels with a score ≥ 0.55. As long as the threshold is not higher than 0.55, they are absolutely safe.

[0093] Scenario A: Increase the threshold (tighten, retain less).

[0094] Action: Increase the threshold from 0.37 to 0.40 (between 0.43 for 12th place and 0.39 for 13th place).

[0095] result: Core #21 (Score 0.39) is below the new threshold → it is pruned.

[0096] Keep the top 12 → 12 is a multiple of 4.

[0097] The minimum score for the critical core is 0.55, while the new threshold is only 0.40.

[0098] The threshold is far below the critical core score line → No critical cores are deleted.

[0099] Scenario B: Lower the threshold (relax the threshold, retain more). Action: Reduce the threshold from 0.37 to 0.25 (between 0.28 for the 16th place and 0.22 for the 17th place).

[0100] result: Cores #21 (0.39), #29 (0.35), #12 (0.31), and #5 (0.28) were all included in the retention scope.

[0101] Keep the top 16 → 16 is a multiple of 4.

[0102] All key cores retain their original rankings → No key cores have been deleted.

[0103] This threshold is not a subjective judgment, but the number of retentions N determined by the target pruning rate (based on functional positioning and redundancy), and the position of the "line" drawn on the Score(K) ranking table.

[0104] Method 2: Add virtual convolutional kernels. For layers that cannot be aligned through threshold fine-tuning, add a small number of virtual convolutional kernels with all zero weights at the end of the layer, or add extremely small convolutional kernels with absolute weight values ​​less than half of the current layer's quantization factor (i.e., integer zero after quantization). This ensures that the multiplication and addition results are always zero in subsequent INT8 inference, without affecting the model's original feature extraction function. It is only used to fill in the number of output channels to meet the hardware alignment requirements of SIMD parallel computing.

[0105] By introducing hardware-aware constraints, modifying the fractional formula, and correcting the forced alignment, this invention ensures that the pruned model structure is highly compatible with the chip, fully leverages the acceleration capabilities of SIMD instructions, and achieves a stable single inference time of <10ms while significantly compressing the model size.

[0106] For example, a convolutional layer originally had 32 convolutional kernels, and 60% of them need to be pruned.

[0107] Objective: The remaining quantity should ideally be a multiple of 4.

[0108] Traditional pruning: After pruning: 13 cores remain (32 × 0.4 = 12.8, rounded down to 13). 13 is not a multiple of 4. Inference time: 12ms (slower than the target of 10ms).

[0109] This invention relates to pruning (with hardware awareness): After calculating the overall score, it was found that: Pruned to 12 cores: 98.5% accuracy, 9ms inference time. Pruned to 13 cores: 98.7% accuracy, 12ms inference time. Pruned to 16 cores: 98.3% accuracy, 9ms inference time.

[0110] The decision was as follows: 12 cores (multiples of 4) → 98.5% accuracy, 9ms time. Alternatively, 16 cores (multiples of 4) → 98.3% accuracy, 9ms time. The decision not to select 13 cores (although offering 0.2% higher accuracy, but 3ms slower) and instead reducing the number of cores by one (13 → 12) was to meet hardware alignment requirements. While this resulted in a 0.2% decrease in accuracy, it improved speed by 33%, and the accuracy can be restored to 98.5% after fine-tuning, fully meeting engineering requirements. This is an optimization strategy that trades minimal accuracy loss for significant speed gains.

[0111] In cable monitoring systems, real-time performance is the primary requirement: after a cable breakage occurs, a warning must be issued within milliseconds, otherwise the opportunity to take action may be missed; the improvement from 12ms to 9ms, although seemingly small, may determine whether transient events can be captured; ultimately: cut to 12 cores, inference time 9ms, meeting the <10ms requirement.

[0112] 1.4 Symmetric Quantization and Dynamic Quantization.

[0113] In some embodiments, the pruned model is dynamically quantized to obtain an AI monitoring model, including: applying symmetric quantization to the weights of the pruned model and applying window-by-window dynamic quantization to the activation values ​​during model operation to obtain the AI ​​monitoring model.

[0114] In this embodiment of the invention, the pruned and fine-tuned model is quantized and compressed using a combined strategy of symmetric quantization and dynamic quantization. This significantly reduces the model size while maintaining the accuracy of cable anomaly identification. Symmetric quantization of the weights is performed offline before deployment, and the quantization factor is recalculated based on the remaining weight distribution. = max(| |) / 127.

[0115] Converting the 32-bit floating-point weights to 8-bit integers reduces the storage size of each weight to 25% of its original size in the pruned model, resulting in a model size reduction of approximately 75% compared to before quantization. Simultaneously, setting the quantization factor to a power of 2 facilitates adaptation to the CMSIS-NN instruction set for accelerated shift operations.

[0116] The activation values ​​(in neural networks, activation values ​​refer to the output of each convolutional layer; they are neither the original input data nor the model's weight parameters, but rather intermediate results generated when data passes through each layer of the network. For example, in the first convolutional layer, the input is the original strain waveform (floating-point numbers), scanned by 32 convolutional kernels, producing 32 new sequences, which are the first-layer activation values; in the second convolutional layer, the input is the activation values ​​from the previous layer, scanned by 64 convolutional kernels, producing 64 sequences, which are the second-layer activation values) are dynamically quantized online during inference. The maximum activation value is calculated in real-time for each input window, dynamically determining the activation values. =max(|x|) / 127, and quantize the activation value into an 8-bit integer in real time to participate in the convolution operation. It can simultaneously adapt to normal traffic flow small amplitude signals and broken wire large amplitude sudden change signals, ensuring that both types of signals can be accurately represented within the dynamic range of 8-bit signed integers, avoiding large signal overflow or small amplitude detail loss, and effectively solving the problem that traditional fixed quantization cannot take into account both large and small signals.

[0117] The combination of weighted symmetric quantization and activation value dynamic quantization creates a significant synergistic effect. Weighted quantization compresses the model volume, while dynamic quantization improves the accuracy of wire breakage detection from 93% to 98.5% compared to static quantization. The combination of the two increases the inference speed by 3-5 times, stably meeting the real-time requirement of single inference time <10ms. In view of the characteristics of large amplitude variation, transient changes and long-term drift of cable signals, this joint strategy can ensure that large signals do not overflow and small signals are not lost, fully preserving the jump characteristics of wire breakage, and adaptive baseline changes do not require manual periodic calibration.

[0118] 1.5 Embedded chip deployment.

[0119] After quantization, the model is converted into a format supported by the embedded platform and deployed to the chip of the edge computing unit, achieving the core indicators of memory usage <512KB and single inference time <10ms. At the same time, it supports hot updates of models trained in the cloud in a differential manner. The edge end uses shadow buffer and atomic switching to complete the local upgrade, without the need for manual operation on the bridge, which is suitable for the long-term unattended operation and maintenance needs of bridges in remote mountainous areas.

[0120] In some embodiments, the method further includes: receiving a model differential update instruction sent from the cloud; wherein the differential update instruction only includes local weight change information of the convolutional layer to be modified; allocating a shadow buffer with the same size as the parameters of the convolutional layer in the edge computing unit according to the total number of parameters of the convolutional layer to be updated; copying all parameters of the convolutional layer to be updated from the main model storage area to the shadow buffer; locally modifying the target weights in the shadow buffer according to the differential update instruction, and pointing the access pointer of the main model area to the updated shadow buffer through an atomic switching operation, so that the system immediately loads the new parameters; and releasing the memory space occupied by the original convolutional layer.

[0121] In this embodiment of the invention, the model hot update adopts a differential update + in-situ modification strategy, abandoning the traditional complete double-buffered copy method, and completing the edge model upgrade in a low-traffic, highly reliable, and non-downtime manner. The system first receives the differential instruction sent from the cloud, explicitly specifying the network layer and weight parameters to be modified, then allocates a shadow buffer of matching size according to the layer, and only copies the target layer data from the main model area to the buffer. The differential parameter modification is completed in the shadow buffer, and then the pointer of the main model area is updated through atomic switching, so that the model can directly use the updated layer parameters. Finally, the memory occupied by the original layer is released, realizing lightweight and uninterrupted model iteration.

[0122] The specific steps of this hot update process are as follows: receive the differential instruction from the cloud, locate the network layer and corresponding weights to be modified; allocate a shadow buffer with the same total number of parameters as the target layer; copy only the data of this layer from the main model area to the shadow buffer; perform differential weight modification on the shadow buffer; modify the pointer of the main model area through atomic switching, and use the updated shadow buffer as the current effective layer; release the memory space occupied by the original convolutional layer, without affecting the operation of the monitoring business throughout the process.

[0123] For bridge cable monitoring scenarios, when a new type of broken wire, impact, or slippage pattern is detected, the cloud can complete the training of a new model within 24 hours. Differential commands are sent out with extremely low traffic during the low-business hours in the early morning, transmitting only a small amount of weight change information. The edge device automatically completes the update without human intervention and activates the new model the next day, continuously reducing the false alarm rate. The entire process does not require manual operation on the bridge, greatly improving the operation and maintenance efficiency of in-service bridges in remote mountainous areas, areas with poor network conditions, and inconvenient transportation.

[0124] After hot update, the edge model can maintain stable performance with memory usage of <512KB and inference time of <10ms. The new recognition rules are automatically linked with the original temperature compensation, baseline update, and dynamic threshold logic to ensure that the model accuracy and inference speed are not affected by the update, and to continuously meet the millisecond-level real-time early warning requirements for transient events such as wire breakage.

[0125] Step 2: Strain signal preprocessing based on environmental adaptive correction.

[0126] In some embodiments, before inputting monitoring data into an AI monitoring model to obtain monitoring results, the method includes: performing real-time temperature compensation on the monitoring data based on the rate of change of ambient temperature and a strain-temperature compensation model.

[0127] In this embodiment of the invention, the environmental adaptive correction step is used to solve the diagnostic benchmark offset problem caused by temperature drift and long-term time-varying effects, providing a stable and reliable data foundation for subsequent anomaly identification. The core of this step is achieved through real-time temperature compensation, dynamic updating of the long-term baseline, dynamic threshold adjustment, and an interlocking mechanism between anomaly detection and baseline update. Real-time temperature compensation relies on independently deployed stress-free temperature compensation gratings in the sensor array to obtain pure ambient temperature information, first establishing strain... Temperature compensation model = - f(T), and then complete the offline training and online fine-tuning of the compensation function through piecewise linear fitting or polynomial fitting. The compensation model is recursively corrected using stable operating conditions such as no vehicles at night and stable wind speed to offset the system deviation caused by sensor aging and residual strain from installation. For sudden temperature change scenarios with a rate of change greater than 5°C / h, the system adopts a pure differential compensation strategy, which completes rapid compensation based solely on the temperature difference ΔT and the linear coefficient γ. Furthermore, the anomaly judgment is decoupled from the baseline update path, and the compensation result of the current window does not participate in the baseline update, avoiding false alarms and baseline contamination caused by heat conduction lag.

[0128] In this embodiment of the invention, the long-term baseline dynamic update uses strain values ​​under normal operating conditions (no anomalies, no shocks, and stable temperature) as the benchmark. Iterative updates are performed using the moving average method. The update process is initiated only when there are no anomalies and the signal is stable. A threshold δ for the amplitude of each update is set (the amplitude of a single update does not exceed 1.0). To prevent sudden data from interfering with the baseline, dynamic threshold adjustment is based on the signal standard deviation σ within the sliding window to calculate the threshold in real time: Threshold=Baseline±k×σ, where k is a multiplier factor (3~5). In scenarios with large signal fluctuations such as strong winds and dense traffic, the threshold is automatically relaxed to reduce false alarms, while in scenarios with stable signals, the threshold is tightened to improve the sensitivity of detecting minor anomalies, thus achieving adaptive anomaly discrimination.

[0129] Scene setting The theoretical strain baseline value of a certain cable under healthy conditions is 100.0 με.

[0130] Parameter settings: Baseline updated forgetting factor: α = 0.95 Single update magnitude threshold: δ=1.0 Dynamic threshold factor: k=4 Sliding window size: 256 sampling points (approximately 0.256 seconds) Timeline: 5 consecutive monitoring windows Window 1: No cars at night, everything is fine.

[0131] The collected signals are shown in Table 2: Table 2 Parameter Acquisition Table under Window 1

[0132] Step 1: Dynamic threshold calculation.

[0133] Threshold= ±k×σ=100.0±4×0.3=100.0±1.2 Normal range: [98.8, 101.2].

[0134] The current value is 100.3, which is within the normal range → No alarm will be triggered.

[0135] Step 2: Baseline update.

[0136] Calculate the ideal update amount: Δ=(1 0.95) × 100.3 = 0.05 × 100.3 = 5.015 Compared to the maximum allowed update magnitude δ=1.0: Theoretical update amount without delta limit: +5.015 (this will pull the baseline from 100.0 directly to 105.015, which is too powerful).

[0137] Actual update amount with δ limit: +1.0 (truncated).

[0138] =100.0 + 1.0 = 101.0 The baseline slowly approaches the true state, but is not entirely dominated by single-window data.

[0139] Window 2: Strong winds during the day, heavy traffic.

[0140] The collected signals are shown in Table 3: Table 3 Parameter Acquisition Table under Window 2

[0141] Step 1: Dynamic threshold calculation.

[0142] Threshold=101.0±4×6.5=101.0±26.0 Normal range: [75.0, 127.0].

[0143] The threshold has been significantly relaxed. The current value of 108.0 falls within the normal range → no alarm is triggered. Severe fluctuations caused by strong winds and heavy vehicles are correctly identified as normal operating conditions, and no false alarms are generated.

[0144] Step 2: Baseline update.

[0145] Because σ = 6.5.

[0146] σ=6.5, far exceeding the normal level, the system determines "signal instability", and baseline updates are paused. Avoid strong winds that could contaminate the baseline.

[0147] =100.0 + 1.0 = 101.0 Window 3: Broken thread in the dead of night.

[0148] The collected signals are shown in Table 4: Table 4 Parameter Acquisition Table under Window 3

[0149] Step 1: Dynamic threshold calculation.

[0150] The signal stabilizes late at night, σ decreases, and the threshold tightens. Threshold=101.0±4×0.4=101.0±1.6 Normal range: [99.4, 102.4].

[0151] The current value of 85.0 is far below the lower limit of 99.4 → triggering a Level 1 warning! Step 2: Baseline update.

[0152] The system enters "anomaly warning mode" and forcibly freezes baseline updates: =100.0 + 1.0 = 101.0 A sudden drop in cable tension (85.0) caused by a broken wire will not contaminate the baseline. If the baseline is contaminated (e.g., pulled down to 95), subsequent normal signals will be judged as "abnormally high".

[0153] Window 4: Broken wire confirmation.

[0154] Anomalies were detected in three consecutive windows (85.0, 84.8, 85.1), all of which were below the lower limit of the dynamic threshold.

[0155] The system has been upgraded to a Level 2 early warning system, triggering audible and visual alarms and sending critical alarm information via BeiDou short message service.

[0156] The baseline remains frozen.

[0157] In this embodiment of the invention, strict interlocking control between anomaly detection and baseline update is achieved through a state machine. The system includes four working states: normal, anomaly warning, anomaly confirmation, and anomaly resolution and recovery. In normal mode, baseline update and dynamic threshold adjustment are performed normally. In anomaly warning and anomaly confirmation modes, baseline update is forcibly frozen to prevent abnormal data from polluting the baseline. After the anomaly is completely resolved, the system enters the recovery period and then gradually restores the baseline update function, forming a stable, safe, and interference-resistant preprocessing closed-loop logic.

[0158] This environmentally adaptive preprocessing process works in conjunction with subsequent edge AI diagnostics and dual-mode communication modules to effectively eliminate interference factors such as temperature, load, and long-term aging. It significantly improves the identification stability of key anomalies such as wire breakage, anchor slippage, impact, and wind-induced vibration, enabling the system to maintain a low false alarm rate and high reliability in complex field environments.

[0159] Step 3: Real-time diagnosis based on the consistency between sliding windows and continuous windows.

[0160] In some embodiments, after acquiring the monitoring data of the smart cable, the method further includes: dividing the monitoring data into multiple data segments according to a preset window size; using a sliding window method to analyze the real-time acquired monitoring data segment by segment to obtain the feature data within each sliding window; wherein the size of the sliding window is set according to the characteristics of the cable strain signal; and determining the diagnostic type of the monitoring result based on the feature data within each sliding window.

[0161] In this embodiment of the invention, the real-time diagnostic step addresses the problems of high false alarm rate and inability to distinguish between transient interference and continuous anomalies in existing technologies. It achieves highly reliable anomaly identification with low false alarms through multi-scale window truncation, window-by-window CNN inference, continuous window consistency judgment, and adaptive adjustment of window parameters. The multi-scale sliding window truncation first standardizes and frames the temperature-compensated strain time history data, setting a 256-point base window and a 32-point sliding step to form overlapping segments. Simultaneously, a 64-point short window is used for rapid detection of impact loads, and a 512-point long window is used for slow temperature drift identification. The results are then fused after multi-scale parallel inference, balancing response speed and diagnostic stability.

[0162] In this embodiment of the invention, window-by-window CNN inference inputs the frame-segmented data into a deployed lightweight AI model, outputting the probability distributions of five working conditions: normal, wire breakage, anchor slippage, impact load, and wind-induced vibration. It also fully records the diagnostic results and confidence levels for each window, providing a basis for subsequent consistency determination. Continuous multi-window consistency determination uses the results of consecutive windows as the basis for early warning. A level one warning is triggered when three consecutive windows are consistent, and a level two warning is triggered when five consecutive windows are consistent. A confidence accumulation mechanism is also introduced; when the weighted cumulative confidence level exceeds 0.9, an alarm can still be effectively triggered even if the confidence levels of some individual windows are low, significantly reducing false alarms caused by random interference.

[0163] In this embodiment of the invention, the window parameters are adaptively adjusted to dynamically match the window size based on the real-time standard deviation σ of the signal. When the signal is stable, a 512-point long window is switched to improve diagnostic stability, and when the signal fluctuates violently, a 128-point short window is switched to improve the sensitivity of anomaly response. The step size is adjusted simultaneously to keep the window overlap rate at 50%~75%, ensuring that real anomalies can still be stably captured under complex conditions such as strong winds and sudden changes in traffic flow.

[0164] This step uses a complete closed loop of multi-scale, continuous judgment, and adaptive parameter tuning to filter out transient noise and lock in real faults at the algorithm level, enabling the system to maintain high accuracy in complex field environments and providing millisecond-level reliable early warning for high-risk anomalies such as wire breakage and slippage.

[0165] Step 4: Dual-mode communication and network interruption resume output.

[0166] The dual-mode communication and outage resume collaborative output steps address issues such as communication blind spot monitoring failure, data loss during network outages, and high transmission pressure. Through data hierarchical caching, intelligent communication switching, outage resume guarantee, and multi-channel early warning output, reliable transmission and early warning are achieved across all scenarios. Data hierarchical caching divides data into three levels based on importance: Level 1 data consists of key alarm information, including index number, anomaly type, timestamp, location, and confidence level; Level 2 data contains window inference results and anomaly probability distribution; and Level 3 data is the original strain time history waveform. All data is preferentially written to 64GB industrial-grade eMMC local storage, employing a cyclic overwrite strategy to retain data from the most recent 7 days. Each record includes a timestamp, index number, GPS location, and integrity check code to ensure data traceability and prevent loss.

[0167] In this embodiment of the invention, intelligent switching of communication modes relies on a dual-mode module of 4G / 5G and BeiDou short message service to monitor network status in real time. When the 4G / 5G signal is normal, cellular network transmission is prioritized. If three consecutive heartbeat packets time out are detected, a public network interruption is determined, and the system automatically switches to BeiDou short message mode. In BeiDou mode, only level-one critical alarm information of no more than 140 bytes is sent. After the network is restored, the system automatically switches back to 4G / 5G mode, balancing transmission efficiency and emergency communication capabilities in extreme environments. The system ensures data integrity during network outages by continuously storing all data locally. After network recovery, the data is automatically uploaded in chronological order, supporting breakpoint resumption to avoid duplicate transmissions. Once uploaded, the data is marked as synchronized. If storage space is insufficient, data is overwritten according to a strategy, completely solving the problems of monitoring vacuum and data loss during network outages.

[0168] In this embodiment of the invention, the early warning output adopts a three-level linkage method of local, on-site, and cloud. The local unit triggers the audible and visual alarm at the anchor end to alert on-site personnel, and the alarm is transmitted via LoRa and Wi-Fi. Fi or Bluetooth pushes early warning information to mobile terminals and on-site monitoring centers, while simultaneously synchronizing diagnostic results and alarm records to the cloud management platform, achieving multi-channel, full-coverage early warning reach. The model deployment step converts the quantized and compressed model into an embedded platform-compatible format and deploys it to the chip of the edge computing unit, ensuring stable operation of the model in resource-constrained environments and realizing low-power, small-size, millisecond-level inference engineering applications.

[0169] This step, through a complete system of hierarchical storage, dual-mode communication, network outage resume transmission, and lightweight deployment, enables the system to work stably for a long time in complex engineering environments such as remote mountainous areas, network blind spots, and harsh outdoor conditions, significantly reducing transmission and maintenance costs and improving the practicality and reliability of the bridge cable monitoring system.

[0170] In this embodiment of the invention, the model construction in step one, the signal preprocessing in step two, the real-time diagnosis in step three, and the communication output in step four are not independent serial execution processes, but rather rely on state feedback, parameter adaptation, and cross-module collaboration mechanisms.

[0171] The first step involves reverse control of baseline updates based on diagnostic results. The trigger condition is that the CNN model in step three consecutive windows determines anomalies as either "wire breakage" or "anchor slippage." At this point, step three immediately sends a state latch signal to the environmental adaptive correction module in step two. Upon receiving the signal, step two forcibly freezes the long-term baseline dynamic update function and simultaneously latches the current temperature compensation coefficient. This prevents the baseline update algorithm from misjudging sudden drops in cable force caused by wire breakage or slippage as long-term structural relaxation, thus absorbing the abnormal data as a new normal baseline. Ultimately, this leads to missed detections of similar anomalies. Compared to the problems of blind baseline updates and the tendency for abnormal data to contaminate the baseline in traditional technologies, this invention uses reverse control of the preprocessing flow based on diagnostic results to ensure the purity and stability of the baseline from a mechanistic perspective.

[0172] The second item is the real-time adjustment of the diagnostic window scale based on the degree of signal fluctuation. When the real-time signal standard deviation σ calculated in step two undergoes a significant change, such as a sudden strong wind or a rapid change in traffic flow, step two synchronizes the σ value to the window adaptation module in step three in real time. Step three dynamically switches the CNN input window length according to a preset threshold. In strong fluctuation scenarios such as strong winds and turbulence, a short window is switched to avoid signal truncation that could cause spectral leakage and phase misalignment. In static and low-fluctuation scenarios, a long window is switched to improve noise reduction capability and diagnostic stability, so that the AI ​​model always maintains the optimal input signal-to-noise ratio, solving the defects of traditional fixed windows in complex environments, such as decreased recognition rate and insufficient sensitivity.

[0173] The third aspect involves the global scheduling of data sampling and storage strategies based on communication status. When step four detects a 4G / 5G public network interruption and automatically switches to BeiDou short message mode, it immediately sends a communication status change command to the edge computing master controller scheduler. The scheduler then controls the original data cache module from step one to enter a high compression ratio storage mode, retaining only critical data from the 1 minute before and 5 minutes after the anomaly, significantly reducing Flash write / erase frequency and storage consumption. Simultaneously, it adjusts the diagnostic reporting strategy from step three, executing only first-level critical alarm reporting and suspending the transmission of non-core data such as second-level diagnostic results. In extreme environments such as communication islands and areas without public network coverage, it prioritizes core alarm functions, power consumption, and storage lifespan, and then fully replenishes historical data through breakpoint resumption after network recovery. Unlike traditional systems where communication and data acquisition / storage are independent, and high-frequency sampling continues after network outages leading to storage exhaustion and loss of critical data, this invention achieves intelligent scheduling of all system resources based on communication status.

[0174] The fourth step is the synchronous calibration of signal preprocessing parameters during model hot updates. When the cloud-based model hot update is performed in step one, the cloud simultaneously distributes the temperature compensation polynomial coefficients obtained by fitting the training data of the new model while distributing the new model. After the edge device completes the loading of the new model, it automatically updates the temperature compensation lookup table in step two. This ensures that the data preprocessing process and the feature extraction preferences and feature space of the AI ​​model remain highly consistent, avoiding feature mismatch and accuracy reduction caused by unchanged preprocessing parameters due to model upgrades. This achieves collaborative iteration and synchronous optimization of the model and the preprocessing process.

[0175] In this embodiment of the invention, through the data interaction and control flow coordination of the above four dimensions, the invention is no longer a simple stacking of algorithm functions, but a set of intelligent systems that can operate stably for a long time in complex engineering scenarios such as unattended operation, harsh field conditions, and network fluctuations, and has stronger robustness, adaptability and reliability.

[0176] The specific implementation of this invention is illustrated using Example 1, which describes the deployment of an intelligent cable monitoring system for a double-tower cable-stayed bridge. The bridge has a main span of 580m and a total of 144 cables. Based on the stress state and risk assessment results, 24 cables with the highest stress and degradation probability were selected for intelligent upgrades. The specific implementation steps include: embedding a fiber optic grating sensor array at the anchorage end of each cable; evenly distributing three measuring points on each cable: the anchorage area, the middle section, and the end near the cable body; integrating an edge computing unit inside the protective cover at the cable anchorage end, using a combination of solar power and batteries for uninterrupted power supply; training a lightweight CNN model based on historical bridge monitoring data, and deploying it to the edge computing unit after progressive pruning, symmetrical quantization, and dynamic quantization; completing optical path connection, communication debugging, and alarm threshold parameter configuration; and continuously testing the system for 30 days, recording anomaly identification, alarm response, communication status, and system operation information throughout the process.

[0177] In this embodiment of the invention, the trial operation results show that the system stably identified two abnormal vehicle impact events with response times of 12ms and 15ms, respectively. In the artificially simulated wire breakage test, the system accurately triggered the wire breakage alarm within approximately 0.5 seconds in three consecutive windows, achieving a 100% accuracy rate. In a communication interruption scenario lasting 4 hours, the local storage module completely saved all monitoring and diagnostic data, and automatically resumed transmission after network recovery without any data loss. The overall false alarm rate of the system is less than 0.5%, far superior to the approximately 8% false alarm rate of traditional threshold alarm methods, demonstrating a significant improvement in overall performance.

[0178] Compared with existing traditional cable monitoring technologies, this invention has six outstanding advantages: First, it offers stronger real-time performance, completing data acquisition, demodulation, and AI diagnosis locally at the edge, compressing the early warning response time to the millisecond level. Pruning, quantization sensing, and dynamic quantization processes work together to improve inference speed, increasing wire breakage detection accuracy by 5%, meeting the real-time early warning needs for transient events such as wire breakage, impact, and wind vibration. Second, it has lower network dependence, supporting dual-mode communication with 4G / 5G and BeiDou short message service, ensuring reliable transmission of critical alarm information even during public network outages, completely eliminating monitoring vacuum periods. Third, it achieves higher recognition accuracy, using a lightweight one-dimensional CNN to achieve accurate classification of multiple operating conditions, effectively distinguishing between real anomalies and environmental interference, significantly reducing false alarm rates. Fourth, it reduces data transmission and storage pressure, requiring only... Upload diagnostic results and alarm information, and store raw data locally on demand, reducing data transmission volume and storage costs by more than 90%; fifth, deployment is more flexible and convenient, with highly integrated and compact edge computing units that can be directly installed inside the anchor end protective cover without altering the cable structure or requiring tensioning and releasing operations. It also supports LoRa lightweight networking and is suitable for both newly built bridges and intelligent upgrades of existing bridges; sixth, it possesses system-level intelligence, breaking through the limitations of traditional system communication modules that only serve as passive transmission channels. Driven by communication status, it achieves resource scheduling throughout the entire process of collection, storage, diagnosis, and reporting, solving industry pain points such as data congestion, storage overflow, and loss of critical information in network outage scenarios, enabling the overall system to have stronger survivability and continuous working capability in extreme environments.

[0179] In some embodiments, the edge computing unit includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the smart cable monitoring method as described in the above embodiments.

[0180] Figure 2 This is a schematic diagram of the edge computing unit provided in an embodiment of the present invention. Figure 3 This is a structural schematic diagram of the intelligent cable monitoring system provided in an embodiment of the present invention. Figure 2 and Figure 3 As shown, the intelligent cable monitoring system includes an intelligent cable and an edge computing unit as described in the above embodiment. The intelligent cable monitoring system of the present invention includes the following components: 1. Intelligent cable body 31.

[0181] Cable body: composed of high-strength steel wire or fiber-reinforced composite materials; Anchors: Located at both ends of the cable, used to fix the cable and transmit cable force; Fiber Bragg grating sensor array: Embedded in the cable or anchorage, it is deployed at multiple points to collect parameters such as strain and temperature of the cable in real time. The sensor uses wavelength multiplexing technology, and multiple grating measurement points can be connected in series on a single optical fiber.

[0182] 2. Edge computing unit 32.

[0183] Integrated into the protective cover at the end of the anchor, it boasts a high protection rating (IP67 and above) and is suitable for harsh outdoor environments. Its internal components include the following modules: (1) Optical signal demodulation module 21.

[0184] A swept-frequency laser source and a photodetector are used to convert the wavelength variation of fiber optic grating reflection into digital time-history strain data. Supports multi-channel synchronous acquisition, with a sampling frequency up to 1kHz; Built-in temperature compensation algorithm to eliminate the effect of temperature on wavelength drift.

[0185] (2) Lightweight AI diagnostic module 22.

[0186] Deploy a pruned and quantized convolutional neural network model; Real-time analysis of time-history strain data is achieved to identify abnormal modes such as normal working conditions, wire breakage, anchor slippage, and impact loads. The model adopts a one-dimensional convolutional structure. The input is a strain time history segment (such as 256 sampling points) extracted by a sliding window, and the output is the probability distribution of various working conditions. The inference time is less than 10ms, which meets the real-time requirements.

[0187] (3) Local storage module 23.

[0188] It uses industrial-grade eMMC storage chips with a capacity of no less than 64GB; Used to cache raw data, diagnostic results, model parameters, and runtime logs; It supports local data storage during network outages and automatic resume transmission after network recovery.

[0189] (4) Communication module 24.

[0190] Supports dual-mode communication between 4G / 5G cellular networks and BeiDou short message service; In areas with signal coverage, prioritize using 4G / 5G to upload diagnostic results and alarm information; In areas without public network signals, critical alarm information (such as "broken wire alarm + retrieval number + timestamp") is sent via BeiDou short message service to ensure emergency communication in extreme situations.

[0191] 3. Early warning output unit 25.

[0192] Audible and visual alarm: installed at the anchoring end, it emits an audible and visual alarm locally to alert personnel on site; Wireless transmission module: pushes warning information to mobile terminals (such as mobile phones, tablets) or on-site monitoring centers via LoRa, Wi-Fi or Bluetooth.

[0193] The input of the optical signal demodulation module is connected to the fiber Bragg grating sensor array, and the output is connected to the lightweight AI diagnostic module. It sends the demodulated digital strain time history data to the AI ​​diagnostic module in real time for anomaly identification. The lightweight AI diagnostic module is connected to the local storage module and the early warning output unit. It writes the diagnostic results, anomaly categories, and confidence levels to the local storage module and sends alarm commands to the early warning output unit. The local storage module is connected to the communication module and transmits the hierarchically cached data and information stored during network outages to the communication module. The communication module is bidirectionally connected to the lightweight AI diagnostic module and the local storage module. It receives network status signals and performs global scheduling of data acquisition and storage strategies. At the same time, it uploads key alarms and diagnostic results to the cloud or on-site monitoring center. The early warning output unit is connected to the communication module and pushes local early warning information to mobile terminals and monitoring platforms simultaneously. All modules are uniformly powered and scheduled by the edge computing unit and the main control chip.

[0194] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. Unless otherwise specified or in conflict with logic, the terminology and / or descriptions between different embodiments are consistent and can be referenced interchangeably. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0195] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A smart cable monitoring method, characterized in that, include: Acquire monitoring data from the smart cable; The monitoring data is input into the AI ​​monitoring model to obtain the monitoring results; The AI ​​monitoring model determines the pruning strategy for each convolutional layer based on the hierarchical feature attributes of historical cable signals.

2. The intelligent cable monitoring method of claim 1, wherein, The AI ​​monitoring model is constructed through the following steps: Collect cable tension monitoring data under various working conditions to construct a training set; A convolutional neural network was established for one-dimensional time-series strain signals of cables; Based on the hierarchical feature attributes of historical LAS signals and the L1 norm of the convolutional kernel weights, all convolutional layers of the current layer of the convolutional neural network are evaluated to obtain the importance score of each convolutional layer; wherein, the hierarchical feature attributes include edge abrupt change features and periodic semantic features. Following the order from shallow to deep layers, each convolutional layer is pruned within its layer according to the importance score. After pruning each convolutional layer, the parameters of that layer and subsequent layers are immediately updated using the training set until the overall pruning rate meets the target pruning rate. The pruned model is dynamically quantified to obtain the AI ​​monitoring model.

3. The intelligent cable monitoring method of claim 2, wherein, Following the order from shallow to deep layers, and based on the importance scores, each convolutional layer undergoes intra-layer pruning, including: For each convolutional layer, based on the descending sorting results of the comprehensive importance scores within that layer, an independent pruning threshold and target pruning rate are set within the layer. Redundant convolutional kernels whose overall importance score is lower than the pruning threshold of that layer are removed.

4. The intelligent cable monitoring method of claim 3, wherein, For each convolutional layer, based on the descending ranking of the comprehensive importance scores within that layer, an independent pruning threshold and target pruning rate are set, including: Sort all convolutional kernels in the current convolutional layer in descending order of their overall importance score; Based on the feature extraction function and redundancy of the current convolutional layer, determine the independent target pruning rate for that layer; The pruning threshold of the convolutional layer is determined based on the alignment requirements of the output channel number of the edge computing unit, the ranking result of the comprehensive importance score, and the target pruning rate.

5. The intelligent cable monitoring method of claim 1, wherein, The pruned model is dynamically quantified to obtain the AI ​​monitoring model, including: The pruned model weights are subjected to symmetric quantization, and the activation values ​​during model runtime are subjected to window-by-window dynamic quantization to obtain the AI ​​monitoring model.

6. The intelligent cable monitoring method of claim 1, wherein, The method further includes: Receive model differential update instructions sent from the cloud; wherein, the differential update instructions only include local weight change information of the convolutional layer to be modified; Based on the total number of parameters of the convolutional layer to be updated, allocate a shadow buffer in the edge computing unit that is the same size as the parameters of that layer. Copy all parameters of the convolutional layer to be updated from the main model storage area to the shadow buffer; The target weights in the shadow buffer are locally modified according to the differential update instruction, and the access pointer of the main model area is pointed to the updated shadow buffer through the atomic switching operation, so that the system can immediately load the new parameters. Release the memory space occupied by the original convolutional layer.

7. The intelligent cable monitoring method of claim 1, wherein, Before inputting the monitoring data into the AI ​​monitoring model to obtain the monitoring results, the method includes: Real-time temperature compensation is performed on the monitoring data based on the rate of change of ambient temperature and the strain-temperature compensation model.

8. The intelligent cable monitoring method of claim 1, wherein, After acquiring the monitoring data of the smart cable, the method further includes: The monitoring data is divided into multiple data segments according to a preset window size; A sliding window approach is used to analyze the real-time acquired monitoring data segment by segment to obtain the characteristic data within each sliding window; the size of the sliding window is set according to the characteristics of the cable strain signal. The diagnostic type of the monitoring results is determined based on the feature data within each sliding window.

9. An edge computing unit, characterized by The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor, when executing the computer program, implements the steps of the intelligent cable monitoring method according to any one of claims 1 to 8.

10. A smart cable monitoring system, characterized by, It includes a smart cable and an edge computing unit as shown in claim 9 above.

Images