A structural health monitoring method and system resistant to sensor noise interference

Abstract

This invention discloses a structural health monitoring method and system resistant to sensor noise interference, comprising: acquiring structural damage data including multiple damage scenarios; calculating the cosine similarity between the structural damage data and health signals for each damage scenario, and determining the damage type of the corresponding structure based on the calculation results; wherein the damage types include: structural damage and sensor damage; for signals containing the structural damage, locating the damage location in the corresponding structure based on a Transformer network; and outputting the damage type and location parameters of the damage detection. This invention achieves automatic structural damage detection through vibration signals, effectively managing computational costs even in the presence of faulty sensors, while simultaneously achieving high-accuracy damage detection and location.

Description

Technical Field

[0001] This invention relates to the field of engineering structure monitoring technology, and in particular to a structural health monitoring method and system that resists sensor noise interference. Background Technology

[0002] The durability of engineering structures is frequently reduced due to factors such as environmental erosion, material aging, and operational stress. This degradation weakens the structure's ability to withstand natural disasters and harsh environmental conditions, potentially leading to severe emergencies in extreme cases. Over time, minor structural degradation accumulates, increasing the risk of structural failure. Therefore, timely detection and repair of this damage are crucial for maintaining the safety and functionality of these structures. Consequently, there is an urgent need to develop real-time structural health monitoring (SHM) methods capable of continuously monitoring, detecting, and assessing structural damage.

[0003] Existing real-time structural health monitoring methods require the deployment of multiple sensors, leading to system redundancy and significant computational overhead. The computational cost of health monitoring for offshore structures is particularly high due to the massive amounts of data generated by redundant sensor systems. Offshore facilities, such as wind farms and oil and gas platforms, located in harsh marine environments, require real-time operational capabilities. This necessitates powerful computing capabilities to quickly identify and correct potential structural problems, ensuring safety and durability. Furthermore, operating costs, such as power consumption, equipment maintenance, and personnel training, highlight the necessity of improving computational efficiency to ensure economic viability and the long-term sustainability of health monitoring systems. Therefore, it is necessary to improve the accuracy of structural inspection while reducing computational resources required for health monitoring.

[0004] Therefore, existing technologies still need improvement. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a structural health monitoring method and system that is resistant to sensor noise interference, in order to address the shortcomings of existing real-time structural health monitoring methods, such as low damage location accuracy and high health monitoring computation cost.

[0006] The technical solution adopted by this invention to solve the technical problem is as follows:

[0007] In a first aspect, the present invention provides a structural health monitoring method resistant to sensor noise interference, comprising:

[0008] Acquire structural damage data encompassing multiple damage scenarios;

[0009] Calculate the cosine similarity between structural damage data and health signals for each damage scenario, and determine the damage type of the corresponding structure based on the calculation results; wherein, the damage type includes: structural damage and sensor damage;

[0010] For signals containing the structural damage, the location of the damage in the corresponding structure is located based on the Transformer network;

[0011] Outputs the damage type and location parameters of the damage detection.

[0012] In one implementation, acquiring structural damage data encompassing multiple damage scenarios includes:

[0013] The structural damage scenarios of fatigue cracks between bolt holes, bolt loosening, and / or plastic deformation of connecting components are simulated, and the acceleration signals corresponding to each scenario are recorded under white noise excitation to obtain the structural damage data.

[0014] In one implementation, the cosine similarity between structural damage data and healthy signals for each damage scenario is calculated, and the damage type of the corresponding structure is determined based on the calculation results, including:

[0015] The vibration signals in the structural damage data are segmented into vibration signal sequences of a preset length;

[0016] Perform a fast Fourier transform on the vibration signal sequence to obtain a first frequency domain dataset;

[0017] The signals collected by the structure under healthy working conditions are subjected to Fast Fourier Transform to obtain the second frequency domain dataset;

[0018] The first frequency domain dataset and the second frequency domain dataset are processed by linear projection function respectively, and the cosine similarity value between the processed first frequency domain dataset and the processed second frequency domain dataset is calculated.

[0019] The damage type of the corresponding structure is determined based on the cosine similarity value.

[0020] In one implementation, determining the damage type of the corresponding structure based on the cosine similarity value includes:

[0021] The cosine similarity value is compared with a preset threshold, and the damage type of the corresponding structure is determined based on the difference between the cosine similarity value and the preset threshold.

[0022] In one implementation, locating the damage location in the corresponding structure based on a Transformer network for a signal containing the structural damage includes:

[0023] The first frequency domain dataset is used as the input to the Transformer network;

[0024] Based on the multi-head attention mechanism, using different linear layers, weights, and biases, the vibration signals of the first frequency domain dataset are projected onto the query vector Q, the key vector K, and the value vector V, and the query vector Q, the key vector K, and the value vector V are combined to form the final attention output.

[0025] The final attention output is mapped to a continuous feature representation based on a feedforward network;

[0026] Based on the continuous feature representation, a linear projection layer is used to process the continuous feature representation and derive the damage localization result to obtain the damage set.

[0027] In one implementation, the Transformer network includes: multiple encoder layers;

[0028] Each encoder layer includes a multi-head attention sublayer and a feedforward network sublayer. In each sublayer, the output is processed by dropout, residual connections, and layer normalization.

[0029] In one implementation, the method further includes:

[0030] Based on the Bayesian optimization method, the search is performed in the hyperparameter space of the Transformer network with the goal of minimizing the loss function of the model;

[0031] Based on the search results, the optimized hyperparameters of the Transformer network are determined; wherein the optimized hyperparameters include: input sequence length, number of encoders, number of attention heads in each layer, and number of encoder layers in each encoder.

[0032] Secondly, the present invention provides a structural health monitoring system resistant to sensor noise interference, comprising:

[0033] The data acquisition module is used to acquire structural damage data that includes multiple damage scenarios;

[0034] The damage detection module is used to calculate the cosine similarity between structural damage data and health signals in various damage scenarios, and to determine the damage type of the corresponding structure based on the calculation results; wherein, the damage type includes: structural damage and sensor damage;

[0035] The damage localization module is used to locate the damage location in the corresponding structure based on a Transformer network for a signal containing the structural damage.

[0036] The monitoring output module is used to output the damage type and location parameters of the damage detection.

[0037] Thirdly, the present invention provides a terminal, comprising: a processor and a memory, wherein the memory stores a structural health monitoring program resistant to sensor noise interference, and the structural health monitoring program resistant to sensor noise interference, when executed by the processor, is used to implement the operation of the structural health monitoring method resistant to sensor noise interference as described in the first aspect.

[0038] Fourthly, the present invention also provides a medium, which is a computer-readable storage medium, storing a structural health monitoring program resistant to sensor noise interference, wherein the structural health monitoring program resistant to sensor noise interference, when executed by a processor, is used to implement the operation of the structural health monitoring method resistant to sensor noise interference as described in the first aspect.

[0039] The present invention, by employing the above technical solution, has the following effects:

[0040] This invention employs cosine similarity to compare collected signals with health signals to determine whether a signal indicates health, structural damage, or sensor damage. For signals containing structural faults, a Transformer-based processing network is used to locate the damage in the structure. To improve detection accuracy and optimize computational resources, the loss function is modified to integrate computational power, and a Bayesian network is used to optimize hyperparameters. This invention uses vibration signals from a single location for global structural health monitoring, thereby reducing deployment and maintenance costs. Furthermore, by utilizing signals from faulty sensors, it ensures consistent structural health monitoring even when sensors are damaged. By employing cosine similarity, a Bayesian network, and a modified loss function to optimize the proposed method, a balance is achieved between computational resource utilization and detection accuracy. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0042] Figure 1 This is a flowchart of the structural health monitoring method against sensor noise interference in this invention.

[0043] Figure 2 The simulator in this invention is used to monitor steel frame structures and simulate their damage.

[0044] Figure 3This is a flowchart of the damage detection and localization process in this invention.

[0045] Figure 4 This is a schematic diagram of the cosine similarity-based defect detection in this invention.

[0046] Figure 5 This is a schematic diagram of the defect location network in this invention.

[0047] Figure 6 This is a schematic diagram of the optimal parameter search in this invention.

[0048] Figure 7 This is a schematic diagram of the vibration signal and defect signal collected by the sensor in this invention.

[0049] Figure 8 This is a schematic diagram of the confusion matrix of the detection results in this invention.

[0050] Figure 9 This is a schematic diagram of the accuracy curves of different algorithms for detecting results in this invention.

[0051] Figure 10 This is a schematic diagram of different node accuracies in this invention.

[0052] Figure 11 This is a schematic diagram of the accuracy curves for different parameter combinations in this invention.

[0053] Figure 12 This is a schematic diagram of the accuracy curve of the optimal parameter combination in this invention.

[0054] Figure 13 This is a functional schematic diagram of the terminal in one implementation of the present invention.

[0055] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0057] Exemplary methods

[0058] Vibration-based structural monitoring (SHM) tasks include detecting damage, determining its location, and assessing its nature and extent. Traditionally, these monitoring tasks were performed manually, which is not only labor-intensive but also prone to errors. With advancements in sensor technology and data mining methods, vibration-based SHM methods have gained widespread application due to their non-invasiveness, low cost, and ability to locate damage. Vibration-based SHM methods are often conceptualized as a pattern recognition problem, where damage-sensitive properties reflect the operational state of a structure. This requires deploying a network of accelerometers at strategic locations to capture vibration responses and using sophisticated algorithms to convert the data into damage indicators.

[0059] Various indicators have been developed to indicate the state of structural damage. Modal strain energy (MSE) is widely used in marine engineering and is considered a mature damage indicator. The location of damage can be determined by calculating the signal energy rate (RSE) in the acceleration signal of each element before and after damage. For example, studies have improved the Stubbs damage index within the MSE framework to identify single and multiple damage locations on offshore flare bridges. Other studies have introduced the transmodal strain energy index (CMESI), which incorporates wavelet transform (WT) to reduce the impact of environmental noise on damage location accuracy. Despite these advantages, frequency-based techniques still face challenges in accurately locating and quantifying damage. To enhance structural health monitoring, time-domain and frequency-domain signals are often combined, leveraging the advantages of both methods for more effective damage detection and analysis.

[0060] With the development of computer technology, deep learning has attracted significant attention in data analysis. Applying deep learning methods to structural damage detection and sensor fault diagnosis shows promising promise. However, distinguishing between response changes caused by sensor faults and those caused by structural damage remains a significant challenge. Furthermore, data recovery and sensor fault detection efforts may overlook damage occurring during data loss. Therefore, it is necessary to develop methods that can perform damage detection even under sensor fault conditions.

[0061] Many studies require the deployment of multiple sensors, leading to system redundancy and significant computational overhead. The computational cost of health monitoring for offshore structures is particularly high due to the massive amounts of data generated by redundant sensor systems. Offshore facilities, such as wind farms and oil and gas platforms, located in harsh marine environments, require real-time operational capabilities. This necessitates powerful computing capabilities to quickly identify and correct potential structural problems, ensuring safety and durability. Furthermore, operating costs, such as power consumption, equipment maintenance, and personnel training, highlight the necessity of improving computational efficiency to ensure economic viability and the long-term sustainability of health monitoring systems. Therefore, it is necessary to improve the accuracy of structural inspection while reducing computational resources required for health monitoring.

[0062] To address the above-mentioned technical problems, this invention provides a structural health monitoring method resistant to sensor noise interference. This method uses cosine similarity to compare collected signals with health signals to determine whether the signal indicates health, structural damage, or sensor damage. For signals containing structural faults, a Transformer-based processing network is used to locate the damage in the structure. To improve detection accuracy and optimize computational resources, the loss function is modified to integrate computational power, and a Bayesian network is used to optimize hyperparameters. Therefore, this invention uses vibration signals from a single location for global structural health monitoring, thereby reducing deployment and maintenance costs. Furthermore, by utilizing signals from faulty sensors, consistent structural health monitoring is ensured even when sensors are damaged. Finally, by employing cosine similarity, a Bayesian network, and a modified loss function to optimize the proposed method, a balance is achieved between computational resource utilization and detection accuracy.

[0063] like Figure 1 As shown, this embodiment of the invention provides a structural health monitoring method resistant to sensor noise interference, comprising the following steps:

[0064] Step S100: Obtain structural damage data containing multiple damage scenarios.

[0065] In this embodiment, the method mainly uses vibration signals from a single location for global structural health monitoring, thereby reducing deployment and maintenance costs; at the same time, it utilizes signals from faulty sensors to ensure consistent structural health monitoring even when sensors are damaged; and it optimizes the proposed method by employing cosine similarity, Bayesian networks, and a modified loss function to achieve a balance between computational resource utilization and detection accuracy.

[0066] Specifically, in one implementation of this embodiment, step S100 includes the following steps:

[0067] Step S101: Simulate structural damage scenarios such as fatigue cracks between bolt holes, bolt loosening, and / or plastic deformation of connecting components, and record the acceleration signal corresponding to each scenario under white noise excitation to obtain the structural damage data.

[0068] In this embodiment, to improve the accuracy of damage detection and localization in engineering structures, a structural damage dataset is obtained through simulation. Specifically, this embodiment focuses on a steel frame structure consisting of eight main beams and twenty-five infill beams, supported by four columns. Figure 2As shown in (a), the main beam and the infill beam have lengths of 4.6 meters and 1 meter, respectively. In addition, the remaining twenty beams and two long columns have lengths of 0.77 meters and 1.65 meters, respectively. This frame structure was manufactured and constructed in a university's structural laboratory, specifically for generating a large dataset of vibration signals under various damage scenarios to serve the purpose of structural health monitoring (SHM).

[0069] To simulate structural damage, various scenarios were considered, such as fatigue cracks between bolt holes, loose bolts, or severe plastic deformation of connecting components. In this embodiment, structural damage scenarios are simulated by loosening the bolts at the connection between the infill beam and the main beam. The structural damage scenarios are as follows: Figure 2 As shown in (b). To collect the dataset, a total of 30 accelerometers were deployed, including 27 PCB model 393B04 accelerometers and 3 B&K model 8344 accelerometers, which were distributed across 30 joints on the monitoring frame structure to capture signals.

[0070] As an example, the structural damage dataset obtained through simulation in this embodiment includes thirty-one damage scenarios. The initial scenario represents an undamaged state (i.e., all joints are intact), while scenarios 2 to 31 represent damage cases for joints 1 to 30, respectively. For each scenario, acceleration signals at the connections between 30 infill beams and the main beam were recorded under white noise excitation at a sampling frequency of 1024 Hz. Signal recording lasted for 256 seconds, and each signal contained 262,144 samples. Figure 3 As shown, Figure 3 The dataset displays typical samples of both undamaged and damaged signals. More detailed information about the experimental dataset is available on the publicly available website.

[0071] In the process of simulating and acquiring structural damage datasets, the bias of fault sensors can be conceptually captured through mathematical parameters such as bias, drift, accuracy degradation, gain, and constants, as shown in Table 1. In the context of Table 1, the time series u[t] contains sensor faults derived from the sensor fault model; t represents the sampling point index; This represents the response behavior of the health sensor; while w[t] represents the introduced Gaussian white noise sequence to simulate environmental noise interference.

[0072] Table 1. Sensor Fault Models

[0073]

[0074] In practical sensor applications, positional offset can lead to quantization discrepancies between the sensor output and the actual value, manifesting as bias. Furthermore, in the event of sensor failure, the output value often remains fixed or exhibits characteristics similar to white noise. Therefore, in this embodiment, an unknown constant A is introduced to simulate a consistent value in cases of bias and complete sensor failure. Simultaneously, for drift and accuracy degradation, which can be represented as random variables, a random parameter η needs to be introduced within the range [0,1], combined with unknown constants B and C to describe these phenomena.

[0075] Furthermore, sensor gain failures can occur due to instability in the sensor power supply. The undetermined constant G in Table 1 represents the gain coefficient. It is noteworthy that an increase in the gain coefficient G is associated with a decrease in the sensor's accuracy and precision.

[0076] In this embodiment, by setting up the above sensor fault model, this model provides a framework for understanding and handling sensor faults, and can better address sensor problems in real-world applications.

[0077] like Figure 1 As shown, this embodiment of the invention provides a structural health monitoring method resistant to sensor noise interference, comprising the following steps:

[0078] Step S200: Calculate the cosine similarity between structural damage data and health signals for each damage scenario, and determine the damage type of the corresponding structure based on the calculation results; wherein, the damage type includes: structural damage and sensor damage.

[0079] In this embodiment, to utilize vibration signals to identify the presence and location of damage while balancing computational cost and accuracy, a comprehensive damage identification method is proposed, even in the event of sensor failure. For example... Figure 3 As shown, the method comprises two main components: damage detection and damage localization.

[0080] During the damage detection phase, vibration signals are compared with signals under normal operating conditions using cosine similarity. This method ensures efficient daily monitoring with minimal computational overhead, as structural damage is typically rare. Structural signals are categorized into four types: normal, sensor fault, structural damage, and structural damage with sensor fault.

[0081] When structural damage is detected, the damage localization phase is initiated. This phase uses a Transformer network to accurately determine the damage location. The Transformer network is optimized using a Bayesian network and a modified loss function to achieve an optimal balance between computational cost and detection accuracy.

[0082] Specifically, in one implementation of this embodiment, step S200 includes the following steps:

[0083] Step S201: The vibration signal in the structural damage data is divided into a vibration signal sequence of a preset length;

[0084] Step S202: Perform a fast Fourier transform on the vibration signal sequence to obtain a first frequency domain dataset;

[0085] Step S203: Perform a fast Fourier transform on the signals collected by the structure in a healthy working state to obtain the second frequency domain dataset;

[0086] Step S204: Perform linear projection function processing on the first frequency domain dataset and the second frequency domain dataset respectively, and calculate the cosine similarity value between the processed first frequency domain dataset and the processed second frequency domain dataset.

[0087] Step S205: Determine the damage type of the corresponding structure based on the cosine similarity value.

[0088] In one implementation of this embodiment, step S205 includes the following steps:

[0089] Step S205a: Compare the cosine similarity value with a preset threshold, and determine the damage type of the corresponding structure based on the difference between the cosine similarity value and the preset threshold.

[0090] In this embodiment, as Figure 4 As shown, in the damage detection stage, the vibration signal is first segmented into a sequence of length 512. These time-domain vibration signal sequences are denoted as V. t Then, a Fast Fourier Transform (FFT) is performed to obtain the frequency domain dataset V. f (i.e., the first frequency domain dataset). The signal collected from the structure in a healthy operating state is denoted as V. h_t The frequency domain signal after FFT processing is denoted as V. h_f (i.e., the second frequency domain dataset). V f and V h_f All structures are processed using a linear projection function, and their similarity values ​​are then calculated. The difference in similarity values ​​indicates the current operational status of the structure.

[0091] V′ f =linear f (z)=V f ×W+b (1);

[0092] V′ h_f =linear h_f (z)=V h_ f×W+b (2);

[0093]

[0094] Among them, V f V represents the frequency domain dataset obtained from damage detection; h_f V' represents the frequency domain signal collected under healthy working conditions; W and b represent the corresponding weights and biases, respectively; f V' represents the frequency domain dataset after processing by the linear projection function; h_f The frequency domain signal representing the healthy operating state after processing by the linear projection function; similarity represents the calculated cosine similarity value.

[0095] As an example, this embodiment compares the calculated cosine similarity value with a preset threshold to determine the damage type of the corresponding structure based on the difference. For instance, a calculated cosine similarity value of 95 and a preset threshold of 90-100 indicate that the current working state of the corresponding structure is normal; a calculated cosine similarity value of 85 and a preset threshold of 80-90 indicate that the current working state of the corresponding structure is sensor failure; a calculated cosine similarity value of 75 and a preset threshold of 60-80 indicate that the current working state of the corresponding structure is structural damage; and a calculated cosine similarity value below 60 and a preset threshold of 0-60 indicate that the current working state of the corresponding structure is structural damage with sensor failure. In practical applications, the cosine similarity value and the preset threshold can be set based on relevant empirical values.

[0096] This embodiment uses vibration signals from a single point to monitor the health of the entire structure. Furthermore, cosine similarity is employed for structural damage detection, reducing the computational cost and improving the computational efficiency of routine monitoring. Once damage is detected, a Transformer network can be used to locate it. Even when the signal is interfered with by faulty sensors, the Transformer network still exhibits robustness.

[0097] like Figure 1 As shown, this embodiment of the invention provides a structural health monitoring method resistant to sensor noise interference, comprising the following steps:

[0098] Step S300: For a signal containing the structural damage, locate the damage location in the corresponding structure based on the Transformer network;

[0099] Step S400: Output the damage type and location parameters of the damage detection.

[0100] In this embodiment, after comparing the structural damage data with the healthy signal using cosine similarity, for the determined signal containing the structural damage, the Transformer network designed in this embodiment can be used to locate the damage location in the corresponding structure, and output the damage type and damage location parameters of the damage detection; wherein, the Transformer network includes: multiple encoder layers; each encoder layer includes: a multi-head attention sublayer and a feedforward network sublayer, and in each sublayer, the output is processed by dropout (random deactivation), residual connection and layer normalization.

[0101] Specifically, in one implementation of this embodiment, step S300 includes the following steps:

[0102] Step S301: Use the first frequency domain dataset as the input to the Transformer network;

[0103] Step S302: Based on the multi-head attention mechanism, using different linear layers, weights and biases, the vibration signal of the first frequency domain dataset is projected onto the query vector Q, the key vector K and the value vector V, and the query vector Q, the key vector K and the value vector V are combined to form the final attention output.

[0104] Step S303: Map the final attention output to a continuous feature representation based on the feedforward network;

[0105] Step S304: Based on the continuous feature representation, the continuous feature representation is processed using a linear projection layer, and the result of damage localization is derived to obtain the damage set.

[0106] In this embodiment, when damage is detected, the Transformer network is used to identify the location of the damage, such as... Figure 5 As shown. The result V of the FFT f =x = x(x1, ..., x) n ) are used as input to the Transformer network to generate continuous feature representations z = (z1, ..., z2) n The Transformer network employs multiple encoder layers depending on the complexity of the input data. Each encoder layer consists of two sub-layers: a multi-head attention network and a feedforward network. Within each sub-layer, the output undergoes dropout, residual connections, and layer normalization.

[0107] In this embodiment, a multi-head attention mechanism plays a crucial role in the Transformer network architecture. It combines multiple independent attention functions, each with a different learned linear projection, which has been found to offer greater advantages than a single attention mechanism. This attention mechanism is widely used in sequence data processing tasks, such as natural language processing. It involves processing the input vibration signal V... f Project onto three distinct vectors: query vector Q, key vector K, and value vector V, then compute the output using the following equation:

[0108] Q,K,V=V f W Q +b Q V f W K +b K V f W V +b V (4);

[0109]

[0110] Among them, W Q b Q W represents the weight and bias corresponding to the query vector Q; K b K W represents the weights and biases corresponding to the key vector K; V b V This represents the weights and biases corresponding to the value vector V; Attention(Q,K,V) represents the attention output corresponding to the query vector Q, key vector K, and value vector V.

[0111] In this embodiment, the multi-head attention mechanism uses different linear layers and weights and biases for projection, and combines them to form the final attention output, as shown in Equations (6) and (7) below. Parameter W O W Q W V W K These represent the weights associated with each attention function, and the vectors Q, K, and V, respectively. Through these parameters, the Transformer network model can effectively integrate information from different attention heads, thereby improving its understanding of the input data.

[0112] A′=Attention multihead (Q,K,V)=Concat(h1,...h n W O (6);

[0113]

[0114] Where A' represents the final attention formed by the combination of multiple layers;

[0115] h i This indicates that multiple layers are combined to form the final attention output.

[0116] In this embodiment, layer normalization sums the calculated results x′=(x′1,...,x′) n () as input, and output As shown in equations (8), (9), and (10) below, layer normalization is used to regularize the Transformer network model, making it more stable. This approach reduces internal covariate bias, improves training efficiency and effectiveness, and thus enhances the model's robustness when processing input data.

[0117]

[0118] Where, x' N The result of layer normalization.

[0119] In this embodiment, the dropout layer (random deactivation layer), as described in Equations (11) and (12) below, aims to reduce the correlation between neuron nodes, thereby reducing overfitting, using a specified dropout probability. The Bernoulli function in Equation (11) is used to generate a random probability vector r, which consists of 0 and 1 values ​​based on a given dropout probability p. By randomly "turning off" some neurons during training, dropout helps improve the model's generalization ability, reduces dependence on specific nodes, and thus enhances the model's performance on unseen data.

[0120] r = Bernoulli(p) (11);

[0121] E t =r×x′ N (12);

[0122] The final multi-head attention network output A is calculated as shown in the following formula (13):

[0123] A = Layer norm (A′+D ropout (A′)) (13);

[0124] The subsequent sublayers include a feedforward network, which consists of two linear layers and a ReLU activation function, as shown in the following formula (14):

[0125] F′=FFN(A)=max(0,AW1+b1)W2+b2 (14);

[0126] As mentioned earlier, the final output F of the feedforward network also includes an additional dropout layer, residual connections, and layer normalization layer, as shown in the following formula (15):

[0127] F = Layer norm (F′+Dropout(F′)) (15);

[0128] After m encoder layers, the input sentence is mapped to a continuous feature representation z = (z1, ..., z2). n The Transformer network takes the input sentence x = (x1, ..., x2) as input. n Mapping to a continuous representation z = (z1, ..., z2) n ):

[0129] z = (z1, ..., z) n )=Transformer(x)x=(x1,...,x n (16);

[0130] Finally, in this embodiment, a linear projection layer is used to process the features and derive the damage localization results. The damage set is labeled as D. j .

[0131] z′ d =linear damage (z)=z×W+b (17);

[0132]

[0133] In this embodiment, a Transformer network is used to locate structural damage upon detection. Even when the signal is interfered with by faulty sensors, the Transformer network still exhibits robustness. Under challenging conditions posed by sensor failure, it can still effectively manage computational costs while achieving high-accuracy damage detection and localization. This allows it to be applied to more complex structures, enhancing robustness and scalability in practical SHM applications.

[0134] This invention provides a structural health monitoring method resistant to sensor noise interference, which further includes the following steps: based on the Bayesian optimization method, searching within the hyperparameter space of the Transformer network with the objective of minimizing the model's loss function; determining the optimized hyperparameters of the Transformer network based on the search results; wherein the optimized hyperparameters include: input sequence length, number of encoders, number of attention heads in each layer, and number of encoder layers in each encoder.

[0135] In this embodiment, the Transformer network model was optimized using Bayesian optimization. Specifically, the tree-based Parzen estimator method was used to optimize the hyperparameters of the Transformer network, including d. model encoder count n head ,num layer , representing the length of the input sequence, the number of encoders stacked in the Transformer network, the number of attention heads in each layer, and the number of encoder layers in each encoder, respectively.

[0136] In this embodiment, the objective of Transformer network optimization is to minimize the model's loss function F = f(d model encoder count n head ,num layer The search is performed within the hyperparameter space of the Transformer based on the damage localization network, where f(a, b, c, d) = f(a, b, c, d). The search method is as follows: Figure 6 As shown, an example with a single parameter is used to illustrate the point.

[0137] The first step is to randomly generate initial points (a1, b1, c1, d1), (a2, b2, c2, d2), ..., (a n b n c n d n Then evaluate the loss values ​​f(a1, b1, c1, d1), f(a2, b2, c2, d2), ..., f(a...) at these points. n b n c n d n ).

[0138] Based on the evaluation results, the 15th percentile of the loss value was calculated as the threshold γ = 0.15.

[0139]

[0140] Next, the sample is divided into two parts: D l Includes all that satisfy f(a) i b i c i d i )≤y * The sample. D g Includes all that satisfy f(a) i b i c i d i )≥y *The samples from these two sets are used to estimate two probability distributions: l(a, b, c, d) and g(a, b, c, d).

[0141] l(a,b,c,d)=p(a,b,c,d0|f(a,b,c,d)≤y * ;

[0142] g(a,b,c,d)=p(a,b,c,d)|f(a,b,c,d)≥y * ;

[0143] To select the next evaluation point, you can maximize the following ratios:

[0144]

[0145] In this embodiment, the loss function is modified to integrate computing power, and a Bayesian network is used to optimize hyperparameters, thereby improving detection accuracy and optimizing computing resources.

[0146] To verify the structural health monitoring method against sensor noise interference provided in this embodiment, further explanation is given below through experiments.

[0147] The experimental and validation work of the proposed vibration-based damage identification method in this embodiment includes three parts: data preparation, evaluation matrix, and damage detection algorithm.

[0148] 1) Data preparation:

[0149] This embodiment aims to develop an automated damage detection method based on vibration signals that can operate effectively even in the event of sensor failure. Therefore, the training dataset must include vibration signals from both functional and defective sensors. Furthermore, to enhance the robustness of the method, different degrees of sensor failure must be considered.

[0150] As previously mentioned, in this embodiment, the Transformer network is designed to identify the damage state of the frame structure, assuming the sensors are in optimal condition. To simulate sensor failure, five different types of sensor failures are introduced by manipulating the health sensor data, each with a different failure magnitude. The degree of sensor failure is based on the mean and standard deviation (δ) of the input signal, as shown in Table 2. Figure 7 As shown, Figure 7 The image shows health signals and fault sensor signals.

[0151] Table 2 Mathematical Models for Different Defects

[0152]

[0153] 2) Evaluation Matrix:

[0154] Damage identification and sensor fault diagnosis can be modeled as a classification problem using deep learning techniques. This means classifying vibration signals into the locations of damage. As a deep learning-based classification task, the effectiveness of the proposed method is evaluated using an annotated dataset, and performance is measured using common metrics, primarily accuracy. The accuracy metric is calculated using the following formula:

[0155]

[0156] True positive (TP), true negative (TN), false positive (FP), and false negative (FN) are the number of correctly classified positive samples, correctly classified negative samples, misclassified positive samples, and misclassified negative samples, respectively.

[0157] 3) Damage detection algorithm:

[0158] To identify structural damage using both vibration signals and fault sensors, and to balance computational cost and accuracy, the damage detection algorithm consists of two parts: first, detecting the presence of damage, and then locating the damage. The presence detection part uses a lightweight cosine similarity network, suitable for low-cost daily monitoring. The location detection part employs a Transformer-based network to locate the damage, while also optimizing it with a Bayesian network to achieve a balance between computational cost and accuracy.

[0159] 3.1 Presence detection:

[0160] First, the vibration signal is processed to calculate a cosine similarity value with the healthy state signal, thereby detecting the presence of damage. Both signals are initially processed through two linear projection layers before similarity is calculated. This process is divided into two parts: the first part distinguishes the signal as healthy or faulty; the second part further classifies fault signals as faulty sensors, structural faults, or structural faults with sensor-related faults.

[0161] like Figure 8 As shown, Figure 8 The confusion matrix for these two parts is shown in the figure. Figure 8 In the middle (a), the test results of the first step of the confusion matrix are shown. Figure 8 (b) shows the test results of the second step of the confusion matrix. In the test experiment, the first part was tested 574 times and the second part was tested 104 times in this embodiment. There were only 9 prediction errors, which shows that cosine similarity comparison can be effectively used for daily structural health monitoring.

[0162] 3.2 Positioning Detection:

[0163] In this phase, the goal is to accurately locate damage using vibration signals. To reduce redundancy in the sensor system, signals from individual nodes are used to monitor the entire frame structure. First, the superiority of the Transformer network over other algorithms is evaluated. Next, the strength of the Transformer network at different nodes is assessed to establish its robustness across multiple nodes. Finally, a Bayesian network is employed to further optimize the Transformer network, achieving the optimal balance between computational cost and accuracy. This methodology ensures that the system is both efficient and effective in detecting structural damage.

[0164] 3.3 Hyperparameter Settings:

[0165] To optimize the performance and efficiency of the localization and detection algorithm, the training hyperparameters were set as follows: batch size of 8192, Adam optimizer used, and initial learning rate of 0.0001. Furthermore, a cosine annealing learning rate decay strategy was employed, gradually reducing the learning rate as training progressed. The dataset was divided into an 80% training set and a 20% validation set, providing a robust framework for model training and validation.

[0166] 3.4 Loss Calculation:

[0167] For loss calculation, the CrossEntropyLoss function is used to quantify the difference between the predicted label and the true input label. To enhance the optimization of computational cost, time is integrated as a cost metric in the loss function, as described in Equation (20). This loss calculation ensures that the model not only makes accurate predictions but also runs within acceptable time constraints, achieving a balance between computational efficiency and algorithm performance.

[0168]

[0169] Loss=CrossEntropyLoss+Time (20).

[0170] 3.5 Algorithm Evaluation:

[0171] Initially, various algorithms were evaluated to determine their effectiveness in detecting structural damage within the frame structure. For controllable comparisons, specific nodes (such as node 8) were designated for global damage detection, thus allowing for a focused evaluation of the performance of each algorithm. Figure 9 As shown, based on the results in Table 3 and Figure 9 According to the description, the Transformer network achieves an accuracy of 75%, which is at least 18% higher than other evaluated algorithms such as convolutional networks (Conv) and recurrent neural networks (RNN).

[0172] However, the inherent complexity of Transformer networks negatively impacts their accuracy. To address this issue, Bayesian optimization is employed to fine-tune the hyperparameters of the Transformer network. This strategic optimization effectively balances computational cost with detection accuracy, improving network performance while maintaining efficiency in resource utilization.

[0173] Table 3: Structure and Performance of Different Algorithms

[0174]

[0175]

[0176] The spatial strength of the algorithm was rigorously evaluated within the optimally tuned Transformer network. This was achieved through damage detection using vibration signals collected from various nodes of the structure. The damage detection accuracy at each node was meticulously measured, allowing for a comprehensive comparison of the Transformer network's effectiveness in locating damage at different locations within the structure. This detailed evaluation helps verify the network's robustness and reliability in real-world scenarios.

[0177] Table 4: Structure and performance of different nodes

[0178] Node number accuracy Node number accuracy 3 0.82 15 0.844 5 0.85 18 0.82 8 0.75 23 0.817 10 0.76 25 0.87 13 0.815 28 0.87 14 0.83 30 0.88 8+fft 0.98

[0179] Clearly, the Transformer network can flexibly handle vibration signals from various nodes in a frame structure. (See Table 4 and...) Figure 10 As shown, the significant difference between the highest and lowest accuracies is approximately 13%. Based on this insight, the network was optimized by incorporating a Fast Fourier Transform (FFT) into the input data. This optimization resulted in an impressive 98% accuracy, even when processing previously identified poorly performing node data.

[0180] To balance the computational cost and accuracy of structural health monitoring, the Transformer network was further optimized using Bayesian optimization, employing a modified loss function that incorporates computation time as an optimization factor. Bayesian optimization is an effective method for finding the extrema of computationally expensive functions that lack a closed-form expression. It is particularly suitable for computationally expensive functions where derivatives are difficult to evaluate or functions are non-convex. Bayesian optimization utilizes a surrogate function to approximate the function to be optimized based on continuous observations. Each observation corresponds to a training run where hyperparameters are selected by the optimization process. The surrogate function fits the target function, and the function determines the next observation point in the search space, balancing exploration and utilization. The hyperparameter space is shown in Table 5, and the optimal hyperparameters are determined using a tree-structured Parzen estimator method. Figure 11 As shown, these parameters and their associated loss values ​​are... Figure 11 The results were visualized, with the best recorded loss value being 0.155081. Furthermore, the accuracy of the optimized network after hyperparameter tuning was 96.2%, as shown in the image. Figure 12 As shown, this high level of accuracy highlights the effectiveness of the Bayesian optimization process in improving network performance, even under challenging conditions where sensors are inaccurate.

[0181] Table 5: Hyperparameters of Transformer Networks

[0182] Hyperparameters Parameter space Optimal parameters <![CDATA[d model ]]> 128,256,512 512 <![CDATA[encoder count ]]> 1,2,3 1 <![CDATA[n head ]]> 2,4,8 2 <![CDATA[num layer ]]> 1,2,3 1

[0183] In this embodiment, to reduce the computational cost of routine monitoring, cosine similarity calculation is performed between real-time vibration signals and signals from healthy states to verify the presence of damage. Furthermore, vibration signals from a single point are used to monitor the health of the entire structure. Cosine similarity calculation is more efficient in routine monitoring. Once damage is detected, a Transformer network is used to locate it. Even when the signal is interfered with by faulty sensors, the Transformer network still exhibits robustness. In addition, this embodiment also uses a Bayesian network and a modified loss function that considers computational cost to achieve a balance between computational and detection accuracy.

[0184] The method proposed in this embodiment was trained on steel frame structures under different damage scenarios, and four fault models were used to simulate the interference caused by fault sensors.

[0185] This embodiment achieves the following technical effects through the above technical solution:

[0186] This embodiment uses cosine similarity to compare the collected signals with health signals to determine whether a signal indicates health, structural damage, or sensor damage. For signals containing structural faults, a Transformer-based processing network is used to locate the damage in the structure. To improve detection accuracy and optimize computational resources, the loss function is modified to integrate computational power, and a Bayesian network is used to optimize hyperparameters. This embodiment uses vibration signals from a single location for global structural health monitoring, thereby reducing deployment and maintenance costs. Furthermore, by utilizing signals from faulty sensors, consistent structural health monitoring is ensured even when sensors are damaged. By employing cosine similarity, a Bayesian network, and a modified loss function to optimize the proposed method, a balance is achieved between computational resource utilization and detection accuracy.

[0187] Exemplary device

[0188] Based on the above embodiments, the present invention also provides a structural health monitoring system resistant to sensor noise interference, comprising:

[0189] The data acquisition module is used to acquire structural damage data that includes multiple damage scenarios;

[0190] The damage detection module is used to calculate the cosine similarity between structural damage data and health signals in various damage scenarios, and to determine the damage type of the corresponding structure based on the calculation results; wherein, the damage type includes: structural damage and sensor damage;

[0191] The damage localization module is used to locate the damage location in the corresponding structure based on a Transformer network for a signal containing the structural damage.

[0192] The monitoring output module is used to output the damage type and location parameters of the damage detection.

[0193] This embodiment achieves the following technical effects through the above technical solution:

[0194] This embodiment uses cosine similarity to compare the collected signals with health signals to determine whether a signal indicates health, structural damage, or sensor damage. For signals containing structural faults, a Transformer-based processing network is used to locate the damage in the structure. To improve detection accuracy and optimize computational resources, the loss function is modified to integrate computational power, and a Bayesian network is used to optimize hyperparameters. This embodiment uses vibration signals from a single location for global structural health monitoring, thereby reducing deployment and maintenance costs. Furthermore, by utilizing signals from faulty sensors, consistent structural health monitoring is ensured even when sensors are damaged. By employing cosine similarity, a Bayesian network, and a modified loss function to optimize the proposed method, a balance is achieved between computational resource utilization and detection accuracy.

[0195] Based on the above embodiments, the present invention also provides a terminal, the principle block diagram of which can be as follows: Figure 13 As shown.

[0196] The terminal includes: a processor, a memory, an interface, a display screen, and a communication module connected via a system bus; wherein, the processor of the terminal provides computing and control capabilities; the memory of the terminal includes a storage medium and internal memory; the storage medium stores the operating system and computer programs; the internal memory provides an environment for the operation of the operating system and computer programs in the storage medium; the interface is used to connect to external devices; the display screen is used to display relevant information; and the communication module is used to communicate with a cloud server or other devices.

[0197] When executed by a processor, this computer program is used to implement a structural health monitoring method that is resistant to sensor noise interference.

[0198] It will be understood by those skilled in the art that Figure 13The schematic diagram shown is merely a partial structural diagram related to the present invention and does not constitute a limitation on the terminal to which the present invention is applied. A specific terminal may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0199] In one embodiment, a terminal is provided, comprising: a processor and a memory, the memory storing a sensor noise-resistant structural health monitoring program, which, when executed by the processor, is used to implement the operation of the sensor noise-resistant structural health monitoring method described above.

[0200] In one embodiment, a storage medium is provided, wherein the storage medium stores a sensor noise-resistant structural health monitoring program, which, when executed by a processor, is used to implement the operation of the sensor noise-resistant structural health monitoring method described above.

[0201] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, database, or other media used in the embodiments provided by this invention can include both non-volatile and volatile memory.

[0202] In summary, this invention provides a structural health monitoring method and system resistant to sensor noise interference, comprising: row cosine similarity comparison; determining the damage type of the corresponding structure based on the comparison result; wherein the damage type includes: structural damage and sensor damage; for signals containing the structural damage, locating the damage location in the corresponding structure based on a Transformer network; and outputting the damage type and location parameters of the damage detection. This invention achieves automatic structural damage detection through vibration signals, effectively managing computational costs even in the presence of faulty sensors, while simultaneously achieving high-accuracy damage detection and location.

[0203] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A structural health monitoring method resistant to sensor noise interference, characterized in that, include: Acquire structural damage data encompassing multiple damage scenarios; Calculate the cosine similarity between structural damage data and health signals for each damage scenario, and determine the damage type of the corresponding structure based on the calculation results; wherein, the damage type includes: structural damage and sensor damage; For signals containing the structural damage, the location of the damage in the corresponding structure is located based on the Transformer network; Outputs the damage type and location parameters of the damage detection; The acquisition of structural damage data encompassing multiple damage scenarios includes: The structural damage scenarios of fatigue cracks between bolt holes, bolt loosening and / or plastic deformation of connecting components are simulated, and the acceleration signals corresponding to each scenario are recorded under white noise excitation to obtain the structural damage data. Calculate the cosine similarity between structural damage data and healthy signals for each damage scenario, and determine the damage type of the corresponding structure based on the calculation results, including: The vibration signals in the structural damage data are segmented into vibration signal sequences of a preset length; Perform a fast Fourier transform on the vibration signal sequence to obtain a first frequency domain dataset; The signals collected by the structure under healthy working conditions are subjected to Fast Fourier Transform to obtain the second frequency domain dataset; The first frequency domain dataset and the second frequency domain dataset are processed by linear projection function respectively, and the cosine similarity value between the processed first frequency domain dataset and the processed second frequency domain dataset is calculated. The damage type of the corresponding structure is determined based on the cosine similarity value; The step of determining the damage type of the corresponding structure based on the cosine similarity value includes: The cosine similarity value is compared with a preset threshold, and the damage type of the corresponding structure is determined based on the difference between the cosine similarity value and the preset threshold. The step of locating the damage location in the corresponding structure based on a Transformer network for a signal containing the structural damage includes: The first frequency domain dataset is used as the input to the Transformer network; Based on the multi-head attention mechanism, using different linear layers, weights, and biases, the vibration signals of the first frequency domain dataset are projected onto the query vector Q, the key vector K, and the value vector V, and the query vector Q, the key vector K, and the value vector V are combined to form the final attention output. The final attention output is mapped to a continuous feature representation based on a feedforward network; Based on the continuous feature representation, a linear projection layer is used to process the continuous feature representation and derive the damage localization result to obtain the damage set.

2. The structural health monitoring method against sensor noise interference according to claim 1, characterized in that, The Transformer network includes: multiple encoder layers; Each encoder layer includes a multi-head attention sublayer and a feedforward network sublayer. In each sublayer, the output is processed by dropout, residual connections, and layer normalization.

3. The structural health monitoring method against sensor noise interference according to claim 1, characterized in that, The method further includes: Based on the Bayesian optimization method, the search is performed in the hyperparameter space of the Transformer network with the goal of minimizing the loss function of the model; Based on the search results, the optimized hyperparameters of the Transformer network are determined; wherein the optimized hyperparameters include: input sequence length, number of encoders, number of attention heads in each layer, and number of encoder layers in each encoder.

4. A structural health monitoring system resistant to sensor noise interference, used to implement the structural health monitoring method resistant to sensor noise interference as described in any one of claims 1-3, characterized in that, include: The data acquisition module is used to acquire structural damage data that includes multiple damage scenarios; The damage detection module is used to calculate the cosine similarity between structural damage data and health signals in various damage scenarios, and to determine the damage type of the corresponding structure based on the calculation results; wherein, the damage type includes: structural damage and sensor damage; The damage localization module is used to locate the damage location in the corresponding structure based on a Transformer network for a signal containing the structural damage. The monitoring output module is used to output the damage type and location parameters of the damage detection.

5. A terminal, characterized in that, include: The processor and memory, wherein the memory stores a structural health monitoring program resistant to sensor noise interference, and the structural health monitoring program resistant to sensor noise interference, when executed by the processor, is used to implement the operation of the structural health monitoring method resistant to sensor noise interference as described in any one of claims 1-3.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a structural health monitoring program resistant to sensor noise interference, which, when executed by a processor, is used to implement the operation of the structural health monitoring method resistant to sensor noise interference as described in any one of claims 1-3.

Images