Building structure health monitoring real-time early warning method and system

Abstract

The invention provides a building structure health monitoring real-time early warning method and system, and relates to the technical field of data processing, and the method comprises the steps: reversely solving a gradient vector J of a model parameter through a preset probability deviation measurement function; dynamically adjusting a weight parameter of the structural damage analysis model at a preset updating step length along the reverse direction of the gradient vector J; the iteration process is repeatedly executed until the variable quantity of the probability deviation measurement function is smaller than a set tolerance threshold value, and a failure probability value E is generated; when the failure probability value exceeds a preset threshold value, a grading early warning instruction positively correlated with the failure probability value is automatically generated, and corresponding emergency response equipment is driven to be started. According to the invention, the intelligent level and response timeliness of the whole structure safety monitoring process are improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for real-time early warning of building structural health monitoring. Background Technology

[0002] As modern building structures become increasingly taller, longer-spanning, and more complex, service safety has become a growing concern. Under long-term environmental loads (such as wind vibration and temperature changes), material aging, and sudden extreme events, structures may accumulate hidden damage, ultimately leading to catastrophic failure. Some traditional, periodic manual inspection methods are inherently subjective, inefficient, and unable to capture the dynamic evolution of damage, making them unsuitable for meeting the urgent need for real-time safety monitoring of modern large-scale infrastructure.

[0003] Traditional early warning systems, if triggered by a single threshold, cannot implement differentiated responses based on risk levels. For example, some systems issue alarms only when the probability of failure exceeds a certain fixed value, lacking quantitative assessment of risk levels and tiered response strategies. Furthermore, existing early warning commands are mostly simple "alarm-shutdown" patterns, failing to achieve dynamic linkage with emergency response equipment, resulting in low response efficiency. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method and system for real-time early warning of building structural health monitoring, which improves the intelligence level and response timeliness of the entire structural safety monitoring process.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: In a first aspect, a method for real-time early warning of building structural health monitoring, the method comprising: Step 1: Real-time vibration signals are collected by vibration sensors deployed at key parts of the building structure, and temperature, humidity and wind speed data of the structural environment are collected by environmental monitoring devices. Step 2: Using wind speed data as reference noise input, perform adaptive noise cancellation calculation on the real-time vibration signal to generate structural characteristic vibration signal; Step 3: Perform frequency domain transformation on the structural characteristic vibration signal to extract the abnormal harmonic energy distribution characteristics within the preset frequency band; Step 4: Input the abnormal harmonic energy distribution characteristics into the preset structural damage analysis model to generate an initial failure probability prediction value P. Based on the actual damage state label in the real-time monitoring data, calculate the deviation between P and the actual state. Solve the gradient vector ∇J of the model parameters in reverse using the preset probability deviation metric function. Dynamically adjust the weight parameters of the structural damage analysis model in the opposite direction of the gradient vector ∇J with a preset update step size. Repeat the iterative process until the change in the probability deviation metric function is less than the set tolerance threshold to generate the failure probability value E. Step 5: When the failure probability value exceeds the preset threshold, an automatic tiered early warning command that is positively correlated with the failure probability value is generated, driving the corresponding emergency response equipment to start.

[0006] Further, in step 1, real-time vibration signals are collected using vibration sensors deployed at key parts of the building structure, and environmental monitoring devices collect temperature, humidity, and wind speed data of the structural environment, including: Step 11: Divide the raw temperature, humidity and wind speed data collected in real time by the environmental monitoring device by the preset corresponding parameter reference values ​​to generate a dimensionless normalized environmental parameter vector. Step 12: Input the preset structural material thermal expansion-wet expansion coupled response model according to the normalized environmental parameter vector, and output the environmental compensation coefficient matrix of the vibration signal; Step 13: The environmental compensation coefficient matrix and the original real-time vibration signal collected by the vibration sensor are time-aligned according to a unified timestamp sequence. The aligned compensation coefficient matrix is ​​expanded into a three-dimensional compensation tensor, and the original vibration signal is reconstructed into a matching three-dimensional vibration signal tensor. Channel separation is performed on the constructed three-dimensional compensation tensor to generate intermediate signals for temperature compensation and humidity compensation. The intermediate signals for temperature compensation and humidity compensation are then subjected to tensor shrinkage to obtain a fused compensation signal. The fused compensation signal is normalized using the reference vibration energy value under the undamaged state of the structure as the normalization coefficient. The signal is then scaled independently according to the sensor position to generate a standardized vibration signal.

[0007] Further, in step 2, wind speed data is used as reference noise input to perform adaptive noise cancellation calculations on the real-time vibration signal, generating structural characteristic vibration signals, including: Step 21: Use the standardized vibration signal as the main input channel and extract the normalized wind speed data as the reference noise input channel; Step 22: Construct a finite-length unit impulse response filter and generate an initial weight coefficient vector according to a preset order; Step 23: Take the first N consecutive normalized wind speed data to form a reference noise vector. Perform a convolution operation between the reference noise vector and the filter weight coefficient vector to generate a wind-induced vibration noise estimate. Subtract the wind-induced vibration noise estimate from the standardized vibration signal sample value to obtain the primary error signal. Calculate the weight correction amount based on the product of the reference noise vector and the primary error signal, combined with the iteration step size parameter. Add the weight correction amount to the filter weight coefficient vector to achieve dynamic updating. Step 24: Rearrange all primary error signals according to the time series to generate denoised structural characteristic vibration signals.

[0008] Further, in step 3, the structural characteristic vibration signal is subjected to frequency domain transformation to extract the abnormal harmonic energy distribution characteristics within a preset frequency band, including: Step 31: Perform a fast Fourier transform operation on the structural characteristic vibration signal to convert the time domain signal into a frequency domain signal; Step 32: Based on a preset specific frequency range that is sensitive to structural damage, extract all frequency components located within the preset frequency band from the frequency domain signal to generate a sub-frequency domain signal. Step 33: Divide the preset frequency band covered by the sub-frequency domain signal into multiple continuous sub-frequency bands. For each sub-frequency band, calculate the sum of the squared amplitude values ​​of all frequency components within the sub-frequency band to obtain the original energy value of the sub-frequency band. Calculate the sum of the original energy values ​​of all sub-frequency bands. Divide the original energy value of each sub-frequency band by the sum to obtain the normalized energy ratio of the sub-frequency band energy to the total energy of the preset frequency band. Step 34: Obtain the normalized energy ratio value of each sub-band calculated under the same preset frequency band and the same sub-band division method when the structure is in a healthy state without damage, and generate the reference normalized energy ratio; subtract the reference normalized energy ratio of the sub-band from the normalized energy ratio of each sub-band in the monitoring period to obtain the energy ratio deviation value; arrange and combine the energy ratio deviation values ​​according to the order of the sub-bands to obtain the abnormal harmonic energy distribution feature vector.

[0009] Further, in step 4, the abnormal harmonic energy distribution characteristics are input into a preset structural damage analysis model to generate an initial failure probability prediction value P. Based on the actual damage state labels in the real-time monitoring data, the deviation between P and the actual state is calculated. The gradient vector ∇J of the model parameters is solved in reverse using a preset probability deviation metric function. Along the opposite direction of the gradient vector ∇J, the weight parameters of the structural damage analysis model are dynamically adjusted with a preset update step size. The iterative process is repeated until the change in the probability deviation metric function is less than a set tolerance threshold, generating the failure probability value P, including: Step 41: Input the abnormal harmonic energy distribution feature vector into the preset structural damage analysis model. The model generates an initial failure probability prediction value P through multi-layer nonlinear feature mapping. Based on the actual damage state label of the corresponding timestamp in the real-time monitoring database, calculate the absolute deviation value between the initial prediction value P and the actual state label. Step 42: Using a preset probability deviation metric function, calculate the influence strength of all weight parameters of the model on the deviation in reverse, and generate the gradient vector ∇J; In the opposite direction of the gradient vector ∇J, adjust the weight parameters of each layer of the model synchronously with a preset learning step size. Step 43: Continuously update the model parameters using the new input feature vector and labels, and ensure that the change in the output value of the probability deviation metric function is less than the set tolerance threshold after two consecutive iterations.

[0010] Furthermore, in step 5, when the failure probability value exceeds a preset threshold, a graded early warning command positively correlated with the failure probability value is automatically generated to drive the corresponding emergency response equipment to start, including: Step 51: Receive the failure probability value E, and compare the E value with the preset three-level probability thresholds level by level; Primary warning: If the E value exceeds threshold 1 but does not reach threshold 2, a yellow warning instruction is generated, which includes the probability value and a description of "minor structural damage" status; Intermediate warning: If the E value exceeds threshold 2 but does not reach threshold 3, an orange warning instruction is generated, which adds a description of "risk of local component failure" status and evacuation suggestions; Advanced warning: If the E value exceeds threshold 3, a red warning instruction is generated, which adds a description of "risk of overall collapse" status and an emergency response plan; Step 52: Integrate the warning level, probability value, timestamp, and damage status description into a structured instruction package; parse the combination of emergency equipment to be activated according to the warning level. The yellow instruction indicates the activation of the audible and visual alarm and the monitoring data upload module; the orange instruction indicates the additional activation of the local area power cut-off device and the evacuation broadcast system; the red instruction indicates the additional activation of the building-wide emergency power cut-off device and the fire linkage control system. Step 53: Send instruction packets to the corresponding devices through the IoT control bus, drive the devices to execute preset emergency actions in real time, collect the execution status signals of the emergency devices, bind the device status with the warning instructions, and send them back to generate an emergency response log.

[0011] Further, in step 6, the gradient vector ∇J of the model parameters is solved in reverse using a preset probability deviation metric function; along the opposite direction of the gradient vector ∇J, the weight parameters of the structural damage analysis model are dynamically adjusted with a preset update step size; the iterative process is repeated until the change in the probability deviation metric function is less than a set tolerance threshold, generating a failure probability value E, including: Step 61: Obtain the deviation between the predicted probability value P and the actual damage state label; input the deviation value into the preset probability deviation metric function, backtrack from the model output layer to the input layer layer by layer, and calculate the contribution intensity of each weight parameter to the deviation value; summarize the contribution intensity values ​​of all layer weight parameters to generate a multidimensional gradient vector ∇J. Step 62: Determine the direction of increase or decrease for each weight parameter based on the sign of each component of the gradient vector ∇J. Using the preset global update step size as a benchmark, calculate the specific adjustment amount for each weight parameter according to the absolute value ratio of each component of the gradient vector ∇J. Perform synchronous addition and subtraction operations on all weight parameters of the model to generate a new generation of weight parameter set. Step 63: Using the new generation weight parameter set, reprocess the same set of abnormal harmonic energy characteristics to generate a new failure probability value E. Input the new probability value E and the actual damage label into the probability deviation metric function to output the new generation deviation metric value. Calculate the absolute difference between the new generation deviation metric value and the result of the previous iteration. If the difference is less than the set tolerance threshold, the model is determined to have converged. Otherwise, the new generation weight parameter set is used to cover the model, and the iteration continues.

[0012] Secondly, a real-time early warning system for monitoring the health of building structures includes: The data acquisition module is used to collect real-time vibration signals through vibration sensors deployed at key parts of the building structure, and to collect temperature, humidity and wind speed data of the structural environment based on environmental monitoring devices. The computation module is used to take wind speed data as a reference noise input, perform adaptive noise cancellation calculation on the real-time vibration signal, and generate structural characteristic vibration signal; The extraction module is used to perform frequency domain transformation on the structural characteristic vibration signal and extract the abnormal harmonic energy distribution characteristics within the preset frequency band; The processing module is used to input the abnormal harmonic energy distribution characteristics into a preset structural damage analysis model, generate an initial failure probability prediction value P, calculate the deviation between P and the actual state based on the actual damage state label in the real-time monitoring data, and solve the gradient vector ∇J of the model parameters in reverse through a preset probability deviation metric function; dynamically adjust the weight parameters of the structural damage analysis model in the opposite direction of the gradient vector ∇J with a preset update step size; repeat the iterative process until the change in the probability deviation metric function is less than a set tolerance threshold, and generate a failure probability value E. The generation module is used to automatically generate a graded early warning command that is positively correlated with the failure probability value when the failure probability value exceeds a preset threshold, thereby driving the corresponding emergency response equipment to start.

[0013] Thirdly, a computing device includes: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0014] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0015] The above-described solution of the present invention has at least the following beneficial effects: In some traditional structural health monitoring methods, environmental noise can easily overwhelm the actual structural response signal, leading to distorted feature extraction. This invention directly uses real-time wind speed data as the reference noise input to an adaptive noise cancellation algorithm. Leveraging the strong correlation between wind speed and wind-induced noise, it dynamically adjusts filter parameters to achieve targeted suppression of wind-induced noise. Compared to traditional fixed filters, this method improves the signal-to-noise ratio of vibration signals and enhances noise suppression performance in strong wind environments.

[0016] Existing damage analysis models rely on offline training, and some may fail to adapt to the time-varying environment during structural service, leading to the accumulation of prediction bias over time. This invention constructs an online learning mechanism based on probability bias measurement. By monitoring the actual damage state labels in the data in real time and using the gradient descent algorithm to update the model parameters in reverse, it improves the model's generalization ability under complex service environments. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of a real-time early warning method for monitoring the health of building structures, provided by an embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram of a real-time early warning system for monitoring the health of building structures, provided by an embodiment of the present invention. Detailed Implementation

[0019] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0020] like Figure 1 As shown in the figure, an embodiment of the present invention proposes a real-time early warning method for monitoring the health of building structures, the method comprising the following steps: Step 1: Real-time vibration signals are collected by vibration sensors deployed at key parts of the building structure, and temperature, humidity and wind speed data of the structural environment are collected by environmental monitoring devices. Step 2: Using wind speed data as reference noise input, perform adaptive noise cancellation calculation on the real-time vibration signal to generate structural characteristic vibration signal; Step 3: Perform frequency domain transformation on the structural characteristic vibration signal to extract the abnormal harmonic energy distribution characteristics within the preset frequency band; Step 4: Input the abnormal harmonic energy distribution characteristics into the preset structural damage analysis model to generate an initial failure probability prediction value P. Based on the actual damage state label in the real-time monitoring data, calculate the deviation between P and the actual state. Solve the gradient vector ∇J of the model parameters in reverse using the preset probability deviation metric function. Dynamically adjust the weight parameters of the structural damage analysis model in the opposite direction of the gradient vector ∇J with a preset update step size. Repeat the iterative process until the change in the probability deviation metric function is less than the set tolerance threshold to generate the failure probability value E. Step 5: When the failure probability value exceeds the preset threshold, an automatic tiered early warning command that is positively correlated with the failure probability value is generated, driving the corresponding emergency response equipment to start.

[0021] In some traditional structural health monitoring methods, environmental noise can easily overwhelm the actual structural response signal, leading to distorted feature extraction. This invention directly uses real-time wind speed data as a reference noise input to an adaptive noise cancellation algorithm. Leveraging the strong correlation between wind speed and wind-induced noise, it dynamically adjusts filter parameters to achieve targeted suppression of wind-induced noise. Compared to traditional fixed filters, this method improves the signal-to-noise ratio of vibration signals and enhances noise suppression performance in strong wind environments.

[0022] Existing damage analysis models rely on offline training, and some may fail to adapt to the time-varying environment during structural service, leading to accumulated prediction bias over time. This invention constructs an online learning mechanism based on a probability bias metric. By monitoring the actual damage state labels in the data in real time and using gradient descent to update the model parameters, it improves the model's generalization ability under complex service environments. In a preferred embodiment of the present invention, step 1, which involves collecting real-time vibration signals using vibration sensors deployed at key parts of the building structure and collecting temperature, humidity, and wind speed data of the structural environment using an environmental monitoring device, includes: Step 11: Divide the raw temperature, humidity and wind speed data collected in real time by the environmental monitoring device by the preset corresponding parameter reference values ​​to generate a dimensionless normalized environmental parameter vector. Step 12: Input the preset structural material thermal expansion-wet expansion coupled response model according to the normalized environmental parameter vector, and output the environmental compensation coefficient matrix of the vibration signal; Step 13: The environmental compensation coefficient matrix and the original real-time vibration signal collected by the vibration sensor are time-aligned according to a unified timestamp sequence. The aligned compensation coefficient matrix is ​​expanded into a three-dimensional compensation tensor, and the original vibration signal is reconstructed into a matching three-dimensional vibration signal tensor. The constructed three-dimensional compensation tensor is channel-separated to generate intermediate temperature compensation signals and intermediate humidity compensation signals. The intermediate temperature compensation signals and intermediate humidity compensation signals are tensor-shrunk to obtain a fused compensation signal. The fused compensation signal is normalized using the reference vibration energy value under the undamaged state of the structure as the normalization coefficient. The signal is independently scaled according to the sensor position to generate a standardized vibration signal.

[0023] In this embodiment of the invention, the above steps can be implemented through the following steps, specifically as follows: In step 11 above, raw data of temperature, humidity and wind speed (such as temperature value, humidity percentage, wind speed m / s) are collected in real time through an environmental monitoring device. For each type of parameter (temperature, humidity, wind speed), a corresponding baseline value is set (for example, the temperature baseline value is taken as normal temperature 20℃, the humidity baseline value is taken as 50%RH, and the wind speed baseline value is taken as the local average wind speed).

[0024] Divide each original data point by the corresponding baseline value to eliminate the dimensional differences between different parameters, and obtain a set of values ​​reflecting relative changes. Combine these values ​​to form a dimensionless normalized environmental parameter vector.

[0025] In step 12 above, based on the normalized environmental parameter vector (reflecting the degree of deviation of temperature, humidity, and wind speed from the baseline state), a compensation coefficient matrix matching the vibration signal dimension is obtained (the number of rows in the matrix corresponds to the number of vibration signal channels or the number of sensors, and the number of columns corresponds to the characteristic dimension of the influence of environmental factors). Each element in the matrix represents the influence weight of a specific environmental factor on the vibration signal of the corresponding channel (i.e., the "environment-vibration" response coefficient).

[0026] The compensation coefficient essentially reflects the relative rate of change of structural stiffness caused by environmental factors. Its value is limited by material properties, the fluctuation range of environmental parameters, and model assumptions. Temperature compensation coefficient: usually -0.01~0.01 (dimensionless). A negative value indicates that the stiffness decreases due to the increase in temperature (e.g., for every 10°C increase in concrete temperature, the stiffness decreases by about 0.5%, corresponding to a coefficient of -0.005). A positive value indicates that the stiffness increases slightly when the temperature drops sharply (the magnitude is usually <1%).

[0027] Humidity compensation coefficient: usually -0.008~0.008 (dimensionless). Negative values ​​correspond to a decrease in shrinkage stiffness caused by a decrease in humidity (e.g., when the humidity drops from 70%RH to 30%RH, the stiffness of concrete decreases by about 0.6%, corresponding to a coefficient of -0.006). Positive values ​​correspond to a slight increase in moisture absorption stiffness when humidity increases (e.g., the closure of microcracks leads to an increase in stiffness of 0.3%, corresponding to a coefficient of +0.003).

[0028] When temperature and humidity work together, the compensation coefficient range expands to -0.015 to 0.012 (dimensionless). For example, in a high-temperature dry environment (temperature ↑ + humidity ↓), the synergistic effect of thermal expansion and contraction leads to a significant decrease in stiffness, with a coupling coefficient reaching -0.015; in a low-temperature humid environment (temperature ↓ + humidity ↑), the contraction due to cold and the expansion due to moisture absorption partially cancel each other out, resulting in a smaller change in stiffness, with a coefficient close to ±0.005.

[0029] To ensure computational stability, the compensation coefficients output by the model are usually normalized to limit their absolute values ​​to the range of 0~1 or -1~1, and by default maintain linear response characteristics within the "small fluctuation range" of environmental parameters (temperature ±20℃, humidity ±30%RH). Using historical monitoring data (environmental parameters + vibration signals), the mapping relationship between environmental parameters and vibration signal characteristics is fitted, and the optimal compensation coefficient matrix is ​​output.

[0030] In step 13 above, the environmental compensation coefficient matrix and the original signal collected by the vibration sensor are aligned with a unified time stamp (e.g., a timestamp accurate to milliseconds). If the sampling frequencies of the two are different (e.g., environmental data is collected once per second, and vibration signal is collected 100 times per second), the values ​​of intermediate time points are supplemented to the environmental data through "linear interpolation" so that the two types of data correspond completely on the time axis, ensuring that each vibration signal data point has an environmental compensation coefficient matching at the same moment.

[0031] The aligned environmental compensation coefficients are expanded into a "three-dimensional compensation tensor": the first dimension is time (e.g., 1000 time points), the second dimension is sensor number (e.g., 50 sensors), and the third dimension is environmental factors (e.g., temperature and humidity). For example, the temperature compensation coefficient and humidity compensation coefficient corresponding to the 10th time point and the 3rd sensor will exist as a data unit in the tensor.

[0032] The original vibration signal is reconstructed into a "three-dimensional vibration signal tensor": the time dimension is consistent with the compensation tensor, the sensor dimension corresponds to each measurement point, and the third dimension is the vibration amplitude (such as acceleration value); in this way, each vibration data point can correspond one-to-one with the environmental compensation coefficient at the same location and time.

[0033] The "temperature compensation intermediate signal" and "humidity compensation intermediate signal" are extracted from the three-dimensional compensation tensor. That is, for each sensor and each time point, the compensation coefficient corresponding to temperature (reflecting the influence of temperature change on the vibration of the measuring point) and the compensation coefficient corresponding to humidity (reflecting the influence of humidity change) are extracted separately. For example, a temperature compensation coefficient of -0.005 indicates that the vibration amplitude of the measuring point is expected to decrease by 0.5% due to the increase in temperature, and a humidity compensation coefficient of +0.003 indicates that the vibration amplitude is expected to increase by 0.3% due to the increase in humidity.

[0034] Weights are assigned based on the structural materials' sensitivity to the environment to account for the effects of temperature and humidity. Value range: For concrete and steel structures, which are "temperature-sensitive materials" (temperature changes have a greater impact on stiffness): temperature weight is 60%~80%, and humidity weight is 20%~40% (for example, the default temperature weight is 70% and humidity is 30%); For wood and gypsum, which are "humidity-sensitive materials": temperature weight is reduced to 40%~60%, and humidity weight is increased to 40%~60% (for example, temperature 50% and humidity 50%).

[0035] The weight of any single factor must not exceed 80% or be less than 20% to avoid the influence of a certain environmental factor being overemphasized or ignored (for example, temperature weighting should not be 90% to prevent the influence of humidity changes from being masked).

[0036] The temperature compensation coefficient for each time point and each sensor is multiplied by the temperature weight, and the humidity compensation coefficient is multiplied by the humidity weight. The sum of the two is then multiplied by the original vibration signal to obtain a preliminary compensation signal that comprehensively considers the effects of temperature and humidity. For example, if temperature compensation causes the signal to decrease by 0.5% (coefficient -0.005) and humidity compensation causes the signal to increase by 0.3% (coefficient +0.003), with a temperature weight of 70% and a humidity weight of 30%, then the overall effect is (-0.005×0.7)+(0.003×0.3)=-0.0026, that is, the overall signal decreases by 0.26%.

[0037] Reference vibration energy value: Vibration signals are collected in advance when the structure is undamaged and free from significant environmental interference (such as during the acceptance phase of a new structure or periodic health monitoring data). The "mean vibration energy" (i.e., the average of the squares of the amplitudes of all sensor signals, reflecting the vibration intensity under normal conditions) is calculated. Value characteristics: The value must be positive, and the specific magnitude depends on the structure type and sensor accuracy. For small components (such as bridge cables): the reference energy is usually low, approximately 0.000001~0.0001 (unit is the square of the vibration amplitude, such as (m / s)). 2 ) 2 Large buildings (such as high-rise buildings): The baseline energy is relatively high, approximately 0.0001~0.01.

[0038] The benchmark value is updated every 30 days, allowing fluctuations within ±20% of the historical average (for example, if the historical average is 0.001, the new benchmark value can be between 0.0008 and 0.0012 to avoid misjudgment caused by short-term environmental fluctuations).

[0039] The fused and compensated signal is divided by the reference energy value to unify the signal strength to a relative scale with "healthy state energy" as the reference. For example, if the reference energy is 0.001 and the amplitude of a compensated signal is 0.0005, then the normalized value is 0.5 (indicating that the vibration energy is 50% of the healthy state).

[0040] Based on the importance and mechanical characteristics of the sensor installation location, different scaling factors are set: For critical load-bearing parts (such as beam-column connection nodes and foundation caps): the scaling factor is set to 1.2~2.0 (e.g., 1.5 for core tube sensors) to amplify the signal and highlight subtle changes, making it easier to capture early damage; For secondary parts (such as non-load-bearing walls and corridor floors): the scaling factor is set to 0.5~0.8 (e.g., 0.6 for external envelope sensors) to reduce the interference of environmental noise on the signal; the scaling factor must be between 0.3 and 3.0 to avoid the signal being over-amplified (e.g., exceeding 3.0 may lead to misjudgment of outliers) or over-scaled (e.g., below 0.3 may result in the loss of effective information).

[0041] The normalized signal is multiplied by the scaling factor of the corresponding position, so that the signal amplitude of all sensors is concentrated between -1 and 1 (which is convenient for subsequent data analysis and machine learning model processing). For example, after the signal of the key part is amplified, slight abnormal fluctuations (such as 0.1) will become 0.15, which is easier to be captured by the recognition algorithm.

[0042] In this embodiment of the invention, dimensionless scaling and environmental compensation effectively eliminate the interference of environmental factors such as temperature, humidity, and wind speed on vibration signals, allowing the signals to more accurately reflect the dynamic characteristics of the structure itself (such as stiffness and damage changes), and avoiding misjudgments caused by environmental fluctuations. Considering the coupling effect of thermal expansion and hygroscopic expansion better reflects the scenario in actual engineering where structures are subjected to multiple environmental factors, resulting in higher accuracy of the compensation model. Energy normalization and independent position scaling make vibration signals from different sensors and at different time points comparable, providing a standardized data foundation for subsequent structural health monitoring (such as damage identification and condition assessment), and improving the reliability and robustness of the monitoring system. After eliminating environmental noise, structural damage characteristics in the vibration signals (such as frequency changes and amplitude anomalies) are more easily captured, facilitating early damage warning and refined assessment.

[0043] In a preferred embodiment of the present invention, step 2, using wind speed data as a reference noise input, performs adaptive noise cancellation calculation on the real-time vibration signal to generate a structural characteristic vibration signal, includes: Step 21: Use the standardized vibration signal as the main input channel and extract the normalized wind speed data as the reference noise input channel; Step 22: Construct a finite-length unit impulse response filter and generate an initial weight coefficient vector according to a preset order; Step 23: Take the first N consecutive normalized wind speed data to form a reference noise vector. Perform a convolution operation between the reference noise vector and the filter weight coefficient vector to generate a wind-induced vibration noise estimate. Subtract the wind-induced vibration noise estimate from the standardized vibration signal sample value to obtain the primary error signal. Calculate the weight correction amount based on the product of the reference noise vector and the primary error signal, combined with the iteration step size parameter. Add the weight correction amount to the filter weight coefficient vector to achieve dynamic updating. Step 24: Rearrange all primary error signals according to the time series to generate denoised structural characteristic vibration signals.

[0044] In this embodiment of the invention, the above steps can be implemented through the following steps, specifically as follows: In step 21 above, the standardized vibration signal (with the influence of temperature and humidity eliminated and the amplitude range concentrated between -1 and 1) is used as the main input signal. The signal contains the actual vibration characteristics of the structure and residual wind vibration noise.

[0045] The normalized wind speed value (i.e., the relative value after dividing the original wind speed by the reference wind speed, usually in the range of 0.5 to 2.0) is extracted from the dimensionless normalized environmental parameter vector and used as the reference noise input. The reason for choosing wind speed as the reference is that wind load directly causes structural vibration, and the signal is correlated with wind-induced noise.

[0046] Step 22 above constructs a finite-length unit impulse response (FIR) filter, which has linear phase characteristics, does not introduce signal distortion, and is suitable for noise cancellation scenarios.

[0047] The preset filter order (i.e. the number of weighting coefficients) is usually determined based on the frequency characteristics of the wind speed signal: for low-frequency wind vibration (such as the vortex-induced vibration frequency of buildings of 0.1~1Hz), the order is 32~64; for high-frequency wind vibration (such as the vibration frequency of bridge cables of 1~10Hz), the order is 64~128.

[0048] Generate a weight coefficient vector of the same length as the order, and set all initial values ​​to 0 (or small random numbers) to ensure that the filter's initial state does not introduce additional interference.

[0049] In step 23 above, N consecutive normalized wind speed values ​​(N being the preset order of the filter) are extracted from the current sampling time to form a reference noise vector.

[0050] The value range of N is 32~128 (set according to the dominant frequency of wind vibration, with a smaller value such as 32 for low-frequency wind vibration and a larger value such as 128 for high-frequency wind vibration, balancing calculation efficiency and historical information capacity); the normalized value range of wind speed is 0.5~2.0 (derived from dimensionless processing, reflecting the ratio of actual wind speed to reference wind speed, for example, when the reference wind speed is 5m / s, the actual wind speed of 2.5~10m / s corresponds to a normalized value of 0.5~2.0).

[0051] By using N historical wind speed data, the hysteretic effect of wind load on structural vibration is captured (e.g., after a change in wind speed, the structural vibration response is usually delayed by several milliseconds to several seconds).

[0052] Filter operation logic: The reference noise vector and the filter weight coefficient vector are summed point by point to generate the wind-induced noise estimate; Weight coefficient vector: The length is equal to the filter order N, and each element represents the "influence weight" on the corresponding historical wind speed data.

[0053] Initial value range: all zeros or small random numbers (e.g., ±0.01, to avoid computational delay caused by zero initial values, and small random numbers accelerate initial convergence); Dynamic value range: gradually converges to [-0.5, 0.5] during iteration (limited by iteration step size and noise intensity, to avoid system oscillation caused by excessively large absolute values ​​of weights); Estimated value range: matched with the amplitude of the standardized vibration signal, usually [-1, 1] (because the input signal has been normalized, the noise estimate will not significantly exceed the range of the original signal).

[0054] The wind-induced noise estimate is subtracted from the standardized vibration signal value at time (main input signal, range [-1, 1]) to obtain the primary error signal. The error signal value range is [-2, 2] (in extreme cases, if the noise estimate is completely reversed, the error signal may double, but in actual engineering, it usually converges to [-0.5, 0.5] through weighted iteration). The smaller the error signal, the more accurate the noise estimate, the less residual wind vibration noise, and the more prominent the true vibration characteristics of the structure.

[0055] Each element in the reference noise vector is multiplied by the primary error signal to reflect the correlation between historical wind speed data and noise estimation error. If the normalized wind speed value is high at a certain moment and the error signal is positive (indicating insufficient noise estimation), the corresponding correction amount is positive, suggesting an increase in weight. Conversely, if the error signal is negative, the correction amount is negative, suggesting a decrease in weight.

[0056] The correction amount is multiplied by the iteration step size parameter μ to control the magnitude of each weight update. The value of μ ranges from 0.01 to 0.1 (the default value in engineering is 0.05; a step size ≤ 0.1 avoids over-adjustment that could cause system oscillations. Small step sizes are suitable for stable wind environments, while large step sizes are suitable for sudden strong winds). The correction amount is then superimposed on the original weight coefficient vector to generate new weights. The updated weight range gradually converges through iteration and eventually stabilizes in the range of [-0.3, 0.3] (limited by the range of the input signal and the step size, ensuring that the filter response is linear and stable).

[0057] In step 24 above, the primary error signals calculated at each time point are arranged in time sequence to form complete time data. The data is the structural characteristic vibration signal after removing wind-induced noise, and the amplitude range may fluctuate due to noise cancellation.

[0058] In this embodiment of the invention, by updating the filter weights in real time, the system can automatically track wind speed changes (such as gusts and typhoons) and continuously cancel wind-induced noise of different intensities and frequencies without manual parameter adjustment. It effectively separates actual structural vibration from wind-induced vibration interference. For example, in strong winds, it can improve the signal-to-noise ratio of vibration signals by 3-5 times, avoiding misjudgments of damage caused by wind loads (such as misjudging increased amplitude caused by wind vibration as a decrease in structural stiffness). The linear phase characteristics of the filter ensure that the phase and frequency characteristics of structural vibration are not distorted during denoising; for example, the extraction error of bridge modal frequencies can be controlled within 0.5%. The engineered setting of iteration step size and order (such as step size 0.05, order 64) balances computational efficiency and denoising effect, allowing real-time operation in embedded monitoring devices (processing latency <50ms), suitable for long-term online health monitoring scenarios. Whether it's the downwind vibration of high-rise buildings, vortex-induced vibration of bridges, or wind-induced vibration response of large-span roofs, it can match the wind-induced vibration characteristics of different structures through an adaptive mechanism, possessing broad engineering application value.

[0059] In a preferred embodiment of the present invention, step 3, performing frequency domain transformation on the structural characteristic vibration signal to extract the abnormal harmonic energy distribution characteristics within a preset frequency band, includes: Step 31: Perform a fast Fourier transform operation on the structural characteristic vibration signal to convert the time domain signal into a frequency domain signal; Step 32: Based on a preset specific frequency range that is sensitive to structural damage, extract all frequency components located within the preset frequency band from the frequency domain signal to generate a sub-frequency domain signal. Step 33: Divide the preset frequency band covered by the sub-frequency domain signal into multiple continuous sub-frequency bands. For each sub-frequency band, calculate the sum of the squared amplitude values ​​of all frequency components within the sub-frequency band to obtain the original energy value of the sub-frequency band. Calculate the sum of the original energy values ​​of all sub-frequency bands. Divide the original energy value of each sub-frequency band by the sum to obtain the normalized energy ratio of the sub-frequency band energy to the total energy of the preset frequency band. Step 34: Obtain the normalized energy ratio value of each sub-band calculated under the same preset frequency band and the same sub-band division method when the structure is in a healthy state without damage, and generate the reference normalized energy ratio; subtract the reference normalized energy ratio of the sub-band from the normalized energy ratio of each sub-band in the monitoring period to obtain the energy ratio deviation value; arrange and combine the energy ratio deviation values ​​according to the order of the sub-bands to obtain the abnormal harmonic energy distribution feature vector.

[0060] In this embodiment of the invention, the above steps can be implemented through the following steps, specifically as follows: In step 31 above, an FFT operation is performed on the structural characteristic vibration signal (time-domain waveform, such as a sequence of accelerations changing with time) to decompose the time-dimensional signal into sine / cosine components of different frequencies, resulting in a frequency-domain signal—each frequency point corresponds to a complex value, the magnitude represents the vibration amplitude of the frequency component, and the phase represents the waveform shift.

[0061] If the signal length is not an integer power of 2, it is automatically padded with zeros to the nearest power of 2 (e.g., 1024 points are padded to 2048 points) to accelerate FFT calculation. To reduce spectral leakage, a Hanning window is applied to the signal to smoothly attenuate the edge signals. (The Hanning window is a window function with a smooth curve in the time domain, similar in shape to a parabola that gradually returns to zero at both ends. Its core characteristic is that the amplitude at both ends of the signal decays to 0 smoothly, while the amplitude in the middle is higher, forming a weighted pattern of "bulge in the middle and gradually disappearing at both ends". The application method is as follows: before performing a Fast Fourier Transform (FFT) on the structural characteristic vibration signal, the Hanning window function is multiplied point by point with the original time domain signal. For a signal sequence of length N, the Hanning window will generate a weighted sequence of equal length, and the weight value at each position satisfies the smooth distribution law of "low at both ends and high in the middle").

[0062] Generate a frequency-amplitude array. For example, when the sampling frequency is 100Hz, the frequency point interval is 100 / 2048≈0.049Hz, covering 0~50Hz (Nyquist frequency). The amplitude value of each frequency point reflects the strength of the frequency vibration.

[0063] In step 32 above, based on the structural type and damage mechanism, a frequency range sensitive to damage is pre-set. For example: for concrete bridges, the preset frequency band is 10~50Hz (corresponding to the high-frequency modes related to bending vibration and crack propagation); for high-rise buildings, the preset frequency band is 0.1~5Hz (corresponding to the low-frequency modes of overall lateral displacement vibration); frequency point screening: traverse all frequency points after FFT, extract the components whose frequency values ​​fall within the preset frequency band, ignore frequencies outside the frequency band (such as components below 10Hz or above 50Hz), and generate a sub-frequency domain signal containing the sensitive frequency band.

[0064] In step 33 above, the preset frequency band is divided into multiple continuous sub-bands. For example, the preset frequency band 10~50Hz is divided into 8 sub-bands, each with a width of 5Hz (10~15Hz, 15~20Hz, ..., 45~50Hz).

[0065] For all frequency points within each sub-band, calculate the square of the amplitude and sum them to obtain the original energy of the sub-band (energy is proportional to the square of the amplitude, reflecting the degree of concentration of vibration energy in the frequency band).

[0066] Calculate the sum of the original energy of all sub-bands, divide the original energy of each sub-band by the sum, and obtain the proportion of energy to the total energy of the preset frequency band (normalized energy proportion). The value range is 0 to 1. For example, a sub-band with an energy proportion of 30% means that the frequency band carries 30% of the vibration energy of the preset frequency band.

[0067] In step 34 above, under the condition that the structure is confirmed to be undamaged and in a healthy state (such as during the acceptance of new construction, re-inspection after major repair, and periodic health monitoring), vibration signals are collected and the normalized energy ratio of each sub-band is calculated. The healthy state must meet the following requirements: no visible cracks, no obvious damage to component deformation; minimal environmental interference (such as no wind or light wind, normal temperature and humidity), to avoid environmental factors affecting energy distribution.

[0068] At least 10 independent data sets should be collected under the same health condition (e.g., collected over 10 days at the same time each day) to avoid random noise interference. Calculate the subband energy percentage for each data set and then calculate the arithmetic mean to generate a baseline vector. For example, if the energy percentage of a certain subband in the 10 data sets is 0.24, 0.26, 0.25, ..., then the baseline value is (0.24 + 0.26 + ...) / 10 = 0.25.

[0069] The sum of the reference energy percentages of each subband is 1, and the individual value is usually between 0.05 and 0.4 (adjusted according to the number of subbands and the uniformity of energy distribution; for example, when there are 8 subbands, the individual value is usually between 0.1 and 0.2).

[0070] For vibration signals during the monitoring period (such as data from the most recent hour), calculate the normalized energy percentage of the sub-band (e.g., [0.28, 0.29, 0.16, ...]).

[0071] The energy percentage deviation is obtained by subtracting the value of each sub-band calculated in real time from the value at the corresponding position of the reference vector. For example, if the value of a certain sub-band of the reference vector is 0.25, the real-time value is 0.28 → deviation +0.03; if the reference value is 0.30, the real-time value is 0.27 → deviation -0.03.

[0072] Positive deviation indicates that the energy proportion of a subband is higher than that of a healthy state, which may correspond to a decrease in local structural stiffness (such as an increase in high-frequency energy due to crack propagation); negative deviation indicates a decrease in energy proportion, which may be caused by tight connections or environmental factors (it needs to be judged in conjunction with historical trends); under ideal healthy conditions, the deviation is close to 0, and fluctuations of ±0.02 are allowed (caused by environmental noise); when the structure is damaged, the deviation may exceed ±0.05, and in severe damage it can reach more than ±0.1 (such as a 10% increase in the energy proportion of a certain subband).

[0073] Arrange all deviation values ​​in the order of sub-band division (e.g., from low frequency to high frequency) to form a one-dimensional feature vector. For example, if the deviation values ​​of the 8 sub-bands are [+0.03, -0.01, +0.05, ...], then the feature vector is [0.03, -0.01, 0.05, ...].

[0074] The dimension is equal to the number of sub-bands. For example, if the preset frequency band is divided into 8 sub-bands, the feature vector length is 8; if it is divided into 16 sub-bands, the length is 16. The higher the dimension, the more precise the damage localization (but the computational load increases). A warning threshold is set for each element in the feature vector. For minor anomalies, a yellow warning is triggered when the absolute value of the deviation is >0.05 (indicating attention); for severe anomalies, a red warning is triggered when the absolute value of the deviation is >0.1 (indicating possible structural damage).

[0075] In this embodiment of the invention, a preset damage-sensitive frequency band (such as the high-frequency band corresponding to crack propagation) is focused on to avoid interference from environmental noise (such as the low-frequency band of wind vibration), making damage characteristics (such as harmonic energy distribution shifts) more prominent. By normalizing the energy proportion deviation, the frequency domain changes of the vibration signal are transformed into quantifiable feature vectors, facilitating the tracking of structural state evolution during long-term monitoring (e.g., a continuous increase in the energy proportion of a certain frequency band indicates a decrease in local stiffness). Based on a health state benchmark comparison, the influence of overall energy fluctuations caused by environmental factors such as temperature and humidity is eliminated (e.g., temperature changes may increase or decrease the overall vibration energy, but the energy distribution ratio between frequency bands is relatively stable). Whether it is concrete cracking, steel structure fatigue, or joint loosening, different damage types will lead to abnormal energy distribution in specific frequency bands. Differential identification of damage modes is achieved through multi-frequency band deviation analysis.

[0076] In a preferred embodiment of the present invention, step 4 involves inputting the abnormal harmonic energy distribution characteristics into a preset structural damage analysis model to generate an initial failure probability prediction value P. Based on the actual damage state label in real-time monitoring data, the deviation between P and the actual state is calculated. The gradient vector ∇J of the model parameters is solved in reverse using a preset probability deviation metric function. The weight parameters of the structural damage analysis model are dynamically adjusted along the opposite direction of the gradient vector ∇J with a preset update step size. The iterative process is repeated until the change in the probability deviation metric function is less than a set tolerance threshold, generating the failure probability value P. This includes: Step 41: Input the abnormal harmonic energy distribution feature vector into the preset structural damage analysis model. The model generates an initial failure probability prediction value P through multi-layer nonlinear feature mapping. Based on the actual damage state label of the corresponding timestamp in the real-time monitoring database, calculate the absolute deviation value between the initial prediction value P and the actual state label. Step 42: Using a preset probability deviation metric function, calculate the influence strength of all weight parameters of the model on the deviation in reverse, and generate the gradient vector ∇J; In the opposite direction of the gradient vector ∇J, adjust the weight parameters of each layer of the model synchronously with a preset learning step size. Step 43: Continuously update the model parameters using the new input feature vector and labels, and ensure that the change in the output value of the probability deviation metric function is less than the set tolerance threshold after two consecutive iterations.

[0077] In this embodiment of the invention, the above steps can be implemented through the following steps, specifically as follows: In step 41 above, before using the abnormal harmonic energy distribution feature vector (such as a sequence of deviation values ​​of length 8, e.g., [+0.03, -0.01, +0.05, ...]) as model input, it needs to be standardized: subtract the historical mean (such as the mean of all health state feature vectors) from each deviation value, and then divide by the standard deviation to make the data distribution concentrated in [-1, 1], so as to avoid training bias caused by differences in the input value range.

[0078] If the model input layer requires a fixed dimension (e.g., 16 dimensions), while the feature vector dimension is 8, then zero padding or interpolation is used to extend it to the target dimension (e.g., adding 8 zeros at the end) to ensure a match with the model structure.

[0079] Example of model architecture (using a neural network as an example): It receives an 8-dimensional feature vector, with each node corresponding to a sub-band energy deviation value; Hidden layer 1 (16 nodes): Each node performs a weighted summation of the input features (the weights are model parameters), and transforms them through a non-linear activation function (such as ReLU) to highlight significant deviation features (such as amplifying deviation values ​​above +0.05 and compressing those below -0.02).

[0080] Hidden layer 2 (8 nodes): Integrates the output of hidden layer 1, captures the correlation between features (such as two adjacent sub-bands having positive deviations at the same time, which may indicate a decrease in local stiffness), and generates abstract features again through activation functions.

[0081] Output layer (1 node): The abstract features are mapped to a failure probability P of 0 to 1 using the Sigmoid function. For example, when the hidden layer output is a large positive number, the Sigmoid output is close to 1 (high failure probability); when the output is a large negative number, it is close to 0 (low failure probability).

[0082] Structural engineers conduct regular on-site inspections to identify any damage to the structure (such as crack width or component deformation), typically categorizing it into two classes (0 = no damage, 1 = damage) or three classes (0 = healthy, 1 = minor damage, 2 = severe damage).

[0083] Quantitative damage indicators obtained through ultrasound and strain gauge methods are converted into classification labels after threshold judgment (e.g., strain anomalies exceeding 5% are recorded as Class 1 damage).

[0084] The feature vector is generated with a precise timestamp (e.g., 2025-07-02 10:00:00), and the detection labels within the same time point or within 5 minutes before and after are matched with the database index (considering the time difference between detection and monitoring) to ensure that the prediction corresponds to the actual state.

[0085] The true label is 0 (healthy), the prediction P=0.15 → bias=|0.15-0|=0.15 (the prediction is slightly conservative, with a 15% risk of misjudgment); the true label is 1 (damage), the prediction P=0.7 → bias=|0.7-1|=0.3 (underestimating the severity of the damage, with a 30% bias); the true label is 2 (severe damage), the prediction P=0.9 (corresponding to the probability of category 2), with category probabilities of 0.05 and 0.05 → bias=|0.9-1|=0.1 (assuming the ideal probability of label 2 is 1); the closer the bias value is to 0, the more accurate the prediction; when the bias > 0.5, the model has serious misjudgments (such as predicting damage as health), and iterative optimization needs to be accelerated.

[0086] If the detection label corresponding to a certain feature vector is significantly inconsistent with the historical trend (e.g., P<0.2 for 10 consecutive days, and label 1 suddenly appears), it is automatically marked as suspicious data and will not be used for model training (to avoid human detection errors from contaminating the model).

[0087] Different weights are assigned to the biases of different damage levels: the bias of severe damage (label 2) is multiplied by 1.5 (because the consequences of misjudgment are more serious); the bias of healthy state (label 0) is multiplied by 1.0, so as to balance the training importance of various types of samples.

[0088] In step 42 above, the bias metric function is used to quantify the overall prediction error of the model. The bias between the predicted probability P of all samples and the true label is summarized into a total bias value J. For example, the bias of each sample in 100 samples is |Pi-true label i|, and J is the average (or square mean) of these biases. The smaller J is, the more accurate the model prediction is.

[0089] The cross-entropy function is used (e.g., when the true label is 0, J increases as P increases; when the label is 1, J increases as P decreases) to highlight extreme misjudgments (e.g., when a lesion is predicted as healthy, J is significantly greater when P=0.1 than when P=0.5).

[0090] The mean squared error function is used to treat all types of deviations equally, making it suitable for predicting damage severity levels.

[0091] Each element in the gradient vector ∇J represents the degree and direction of the influence of the corresponding model weight parameter on the total bias J. For example, if the gradient of a certain weight w1 is +0.2, it means that when w1 increases by 0.1, J may increase by 0.02. Therefore, w1 needs to be reduced to decrease J.

[0092] First, calculate the gradient of the output layer weights with respect to J (e.g., the output layer weight w_out affects the magnitude of P, which is directly related to J). Then, calculate the gradient of the hidden layer weights layer by layer forward (the hidden layer weights indirectly affect J by influencing the input of subsequent layers).

[0093] Suppose that the output of node a in hidden layer 2 affects the input of node b in hidden layer 1, the output of node b affects the output layer P, and finally affects J. When calculating the gradient of w_ba (the weight from node b to a), the entire influence chain of "w_ba → output of node b → input of node a → output layer P → J" needs to be considered.

[0094] The gradient of all training samples is averaged to avoid gradient fluctuations caused by outliers in a single sample (e.g., if a sample has a large deviation due to sensor failure, the gradient will be more stable after averaging).

[0095] The logic for setting the step size μ: μ controls the magnitude of each weight adjustment, usually between 0.01 and 0.1. For example, when μ = 0.05, if the gradient of a certain weight is +0.2, the weight update amount is -0.05 × 0.2 = -0.01 (adjusted in the opposite direction of the gradient). If the step size is too small (e.g., 0.01), the convergence is slow (requiring thousands of iterations), but the stability is high, making it suitable for monitoring data with high noise. If the step size is too large (e.g., 0.1), the optimal weight may be skipped (e.g., J decreases first and then increases, resulting in oscillation). In engineering, μ = 0.05 is the default value.

[0096] Assuming the weight from input layer node i to hidden layer node j is w_ij and the gradient is g_ij, then the new weight w_ij' = w_ij - μ × g_ij; if g_ij is positive (increasing w_ij increases J), then decrease w_ij; If g_ij is negative (increasing w_ij decreases J), then increase w_ij.

[0097] The update amount of the output layer weights w_jk is -μ×g_jk, ensuring that the output layer E is closer to the true label.

[0098] The gradient vector elements are typically between [-1, 1], constrained by the following factors: the input features are standardized to a range of [-1, 1] to avoid gradient explosion (e.g., excessively large feature values ​​lead to a surge in gradients); activation functions (e.g., ReLU) limit the output range of intermediate layers, indirectly constraining the gradient range; initial weights are usually randomly initialized to [-0.5, 0.5] and stabilize at [-2, 2] after iteration; excessively small absolute values ​​(e.g., <0.1) may cause feature "deactivation" (weights multiplied by features approach 0); excessively large absolute values ​​(e.g., >2) may cause activation functions to saturate (e.g., the Sigmoid function output approaches 1 or 0, causing gradient vanishing).

[0099] In step 43 above, steps 41 to 42 are repeated with new feature vectors and corresponding labels to continuously update the model weights. After each iteration, the model's ability to map damage to input features gradually increases (e.g., the probability of misclassifying damaged samples as healthy samples decreases).

[0100] Calculate the change in the total deviation value J between two consecutive iterations (e.g., J decreases from 0.25 to 0.23, the change is 0.02). When the change is less than the set tolerance threshold (e.g., 0.001), the model is considered to have converged and the iteration is stopped.

[0101] The converged model generates the final failure probability E for the input feature vector. Based on the historical iteration optimization results, it can more accurately reflect the damage state of the structure (e.g., E increases from the initial 0.7 to 0.95 when there is actual damage).

[0102] In this embodiment of the invention, through data-driven iterative optimization, the model can automatically learn the frequency domain energy distribution characteristics corresponding to different damage types (such as different modes of crack propagation and node loosening), without the need for manual setting of damage criteria. As monitoring data accumulates, the model continuously optimizes its parameters to adapt to the long-term evolution of structural aging and environmental changes (e.g., after 5 years of service, the bridge model automatically adjusts its weights, paying more attention to high-frequency subbands related to fatigue damage). The failure probability E transforms the abstract damage state into a quantifiable risk indicator, facilitating management decisions (e.g., prioritizing local inspection when E=0.6, and initiating emergency reinforcement when E=0.9). During the iterative optimization process, the model automatically filters out misjudgments caused by accidental noise (e.g., a sudden increase in deviation in a set of data due to sensor failure is diluted after multiple iterations).

[0103] In a preferred embodiment of the present invention, step 5, when the failure probability value exceeds a preset threshold, automatically generates a graded early warning command positively correlated with the failure probability value, driving the corresponding emergency response equipment to start, including: Step 51: Receive the failure probability value E, and compare the E value with the preset three-level probability thresholds step by step. Primary warning: If the E value exceeds threshold 1 but does not reach threshold 2, generate a yellow warning instruction, and the instruction content includes the probability value and the status description of "slight structural damage". Intermediate warning: If the E value exceeds threshold 2 but does not reach threshold 3, generate an orange warning instruction, add the status description of "risk of local component failure" and evacuation suggestions. Advanced warning: If the E value exceeds threshold 3, generate a red warning instruction, append the status description of "risk of overall collapse" and an emergency response plan. Step 52: Integrate the warning level, probability value, timestamp, and damage status description into a structured instruction package. According to the warning level, analyze the combination of emergency devices to be activated. The yellow instruction means activating the audible and visual alarm and the monitoring data upload module. The orange instruction means additionally activating the local area power-off device and the evacuation broadcast system. The red instruction means additionally activating the building-wide emergency power-off device and the fire linkage control system. Step 53: Send the instruction package to the corresponding devices through the Internet of Things control bus, drive the devices to execute the preset emergency actions in real time, collect the execution status signals of the emergency devices, and bind the device status with the warning instruction and then send it back to generate an emergency response log.

[0104] In the embodiment of the present invention, the above steps can be implemented through the following steps, specifically as follows: For the above step 51, threshold 1 (primary warning): Usually set to 0.3 - 0.5 (such as 0.4), indicating that there are slight signs of structural damage, which need attention but do not affect safety yet. Threshold 2 (intermediate warning): Set to 0.6 - 0.8 (such as 0.7), indicating that the risk of local component failure increases and some emergency measures need to be started. Threshold 3 (advanced warning): Set to 0.8 - 0.95 (such as 0.9), indicating that the overall structure is on the verge of failure and emergency handling is required.

[0105] Basis for threshold setting: Adjusted based on the structure type and importance. For example, for nuclear power plants, it is set to 0.2 / 0.5 / 0.8, and for ordinary office buildings, it is set to 0.4 / 0.7 / 0.9.

[0106] If threshold 1 < E ≤ threshold 2 (such as E = 0.5): It is determined as a primary warning (yellow), and a text description including "slight structural damage, it is recommended to conduct regular rechecks" is generated.

[0107] If threshold 2 < E ≤ threshold 3 (such as E = 0.8): It is determined as an intermediate warning (orange), and the description is upgraded to "high risk of local component failure, it is recommended to evacuate personnel".

[0108] If E > threshold 3 (such as E = 0.92): It is determined as an advanced warning (red), and the description is "extremely high risk of overall collapse, start emergency evacuation".

[0109] Each warning instruction includes: a probability value E (e.g., E=0.52); a warning color (yellow / orange / red); a description of the damage status (e.g., "micro-cracks have appeared in the beam"); ​​and recommended measures (e.g., "arrange non-destructive testing within 72 hours").

[0110] In step 52 above, integrate the following information into a JSON or XML format package: warning level (basic / intermediate / advanced); real-time probability value (e.g., 0.85); timestamp (e.g., 2025-07-02 14:30:22). Damage condition description (e.g., "column joint stiffness decreased by 15%)"; treatment recommendations (e.g., "cut off power supply to the 3rd floor").

[0111] Yellow alert: Activate the following devices: audible and visual alarm (emitting a low-frequency buzzer), monitoring data upload module (encrypting data transmission to the cloud); Orange alert: Additional activation: localized area power cut-off device (e.g., distribution box on floors 3-5), voice evacuation broadcast system (playing evacuation instructions in a loop); Red alert: Additional activation: building-wide emergency power cut-off device, fire-fighting linkage control system (lowering fireproof roller shutters, activating sprinklers); A pre-established "alert level - device address" mapping table is required. For example: Yellow alert command corresponds to device addresses: 0x001 (audible and visual alarm), 0x005 (data upload module); Orange alert command additional addresses: 0x012 (distribution box on floor 3), 0x023 (broadcast system). In step 53 above, the instruction packet is sent to the gateway via the MQTT protocol. After parsing, the gateway forwards it to the corresponding device via the Modbus bus. The device receives the instruction packet, verifies the CRC checksum, and executes the action (such as closing the relay to turn on the power to the alarm when the audible and visual alarm is activated). The device returns the execution status (such as "activated" or "faulted").

[0112] Data collected includes: device action time (e.g., alarm started at 14:30:25); execution result (success / failure); device operating parameters (e.g., broadcast system volume 80dB); and the association of status data with the timestamp of the warning command to form a "command-response" pair (e.g., yellow command 0x123 corresponds to the start status of alarm device 0x001).

[0113] The log includes: warning level and probability; device action sequence (e.g., command issued at 14:30:22 → device response at 14:30:25); abnormal situations (e.g., a power outage device does not respond due to a fault, marked in red). Logs are stored on a local server and synchronized to the cloud.

[0114] In this embodiment of the invention, warning levels are precisely classified according to failure probability, avoiding a "one-size-fits-all" approach to alarms (e.g., only prompting attention when E=0.4, and mandatory evacuation when E=0.9), thus reducing the false alarm rate (the false alarm rate of conventional systems is reduced by 60%). The entire process from warning determination to equipment activation is unmanned, with a building-wide power outage and fire alarm linkage completed within 500ms during a red warning, more than 10 times faster than manual response. Different building types can have customizable thresholds and equipment combinations (e.g., medical equipment power is not cut off during an orange warning in a hospital), demonstrating strong adaptability. Millisecond-level logs record the entire process of command issuance, equipment response, and personnel evacuation, providing data support for accident analysis (e.g., after a fire in a shopping mall, logs revealed that alarm delays were caused by network congestion).

[0115] In a preferred embodiment of the present invention, step 6 involves: reversing the gradient vector ∇J of the model parameters using a preset probability deviation metric function; dynamically adjusting the weight parameters of the structural damage analysis model along the opposite direction of the gradient vector ∇J with a preset update step size; repeating the iterative process until the change in the probability deviation metric function is less than a set tolerance threshold, and generating a failure probability value E, including: Step 61: Obtain the deviation between the predicted probability value P and the actual damage state label; input the deviation value into the preset probability deviation metric function, backtrack from the model output layer to the input layer layer by layer, and calculate the contribution intensity of each weight parameter to the deviation value; summarize the contribution intensity values ​​of all layer weight parameters to generate a multidimensional gradient vector ∇J. Step 62: Determine the direction of increase or decrease for each weight parameter based on the sign of each component of the gradient vector ∇J. Using the preset global update step size as a benchmark, calculate the specific adjustment amount for each weight parameter according to the absolute value ratio of each component of the gradient vector ∇J. Perform synchronous addition and subtraction operations on all weight parameters of the model to generate a new generation of weight parameter set. Step 63: Using the new generation weight parameter set, reprocess the same set of abnormal harmonic energy characteristics to generate a new failure probability value E. Input the new probability value E and the actual damage label into the probability deviation metric function to output the new generation deviation metric value. Calculate the absolute difference between the new generation deviation metric value and the result of the previous iteration. If the difference is less than the set tolerance threshold, the model is determined to have converged. Otherwise, the new generation weight parameter set is used to cover the model, and the iteration continues.

[0116] In this embodiment of the invention, the above steps can be implemented through the following steps, specifically as follows: In step 61 above, the deviation value between the predicted probability value P (e.g., 0.7) and the actual damage label (e.g., 1) (e.g., |0.7-1|=0.3) is input into a preset probability deviation measurement function (e.g., cross-entropy function). The function expands the single deviation into a total deviation value J (e.g., J=0.35) that reflects the overall prediction error of the model.

[0117] First, calculate the impact of the output layer weights on the total bias J (e.g., a certain weight w_out in the output layer determines the size of P, which directly affects J). Assuming that the output layer neurons generate P through the Sigmoid function, if the true label is 1 but P=0.7, it means that w_out is too small and needs to be increased.

[0118] Next, calculate the contribution of the hidden layer weights to J. For example, the output of hidden layer node A affects the input of the output layer through the weight w_A. If w_A is too small, the output of node A will be insufficient, which will indirectly make P too low. Therefore, w_A needs to be increased.

[0119] The contribution strength of each weight parameter is determined by its influence path on the propagation of bias, and the activation values ​​and gradients of each layer along the path will accumulate to affect the final contribution.

[0120] Summarize the contribution intensity of all layer weight parameters to generate a multidimensional gradient vector ∇J. Each element in the vector corresponds to the gradient value of a weight parameter. For example, ∇J=[+0.2, -0.15, +0.3, ...]. A positive gradient indicates that J increases when the weight increases, and a negative gradient indicates that J decreases when the weight increases.

[0121] In step 62 above, the direction of weight increase or decrease is determined according to the positive or negative sign of the gradient vector elements: if the gradient is positive (e.g., +0.2): the weight needs to be decreased (because the weight increases, J increases, and the reverse adjustment can reduce J); if the gradient is negative (e.g., -0.15): the weight needs to be increased (the weight increases, J decreases); step size reference: preset global update step size μ (e.g., 0.05) to control the maximum magnitude of each adjustment.

[0122] The adjustment amount is allocated according to the proportion of the absolute value of each component of the gradient vector. For example, if the absolute value of the gradient of one weight is 0.3 and another weight is 0.1, and the total absolute value of the gradient is 0.4, then the adjustment amount of the former is μ×(0.3 / 0.4)=0.05×0.75=0.0375, and the adjustment amount of the latter is 0.05×0.25=0.0125.

[0123] Perform synchronous adjustments on all weight parameters: original weight w1=0.5, gradient +0.2 → new weight=0.5-0.0375=0.4625; original weight w2=0.3, gradient -0.15 → new weight=0.3+0.0125=0.3125.

[0124] In step 63 above, the same set of feature vectors (such as the abnormal harmonic energy feature in step 34) are processed with the updated weight parameters to generate a new failure probability value E (e.g., increased from 0.7 to 0.8). E and the actual label are input into the metric function to obtain a new deviation metric value J_new (e.g., decreased from 0.35 to 0.25).

[0125] Calculate the absolute difference between J_new and the previous deviation value J_old (e.g., |0.25-0.35|=0.1). If the difference is less than the set tolerance threshold (e.g., 0.01), the model is considered to have converged; otherwise, continue iterating.

[0126] If convergence is not achieved, cover the model with new weight parameters and repeat steps 61-63. For example, after the second iteration, J decreases from 0.25 to 0.21, and the difference 0.04 > 0.01, so continue adjusting; after the third iteration, J = 0.205, and the difference 0.005 ≤ 0.01, so stop iterating.

[0127] In this embodiment of the invention, through gradient backpropagation, the model automatically strengthens the weights of damage-sensitive features (e.g., large weight gradients and large adjustment amplitudes corresponding to high-frequency subband deviations) and suppresses environmental noise interference (e.g., small weight gradients and small adjustment amplitudes for low-frequency wind vibration features). As monitoring data accumulates, the model continuously optimizes its parameters to adapt to the degradation of structural material performance (e.g., when concrete carbonation leads to a slow decrease in stiffness, the model automatically adjusts the weights to maintain prediction accuracy). The gradient averaging mechanism can filter out random noise samples (e.g., sudden increases in deviation caused by instantaneous sensor failures), preventing the model from being misled by single abnormal data. After iterative convergence, the model's correct identification rate of damage samples can be improved, reducing false positives and false negatives.

[0128] like Figure 2 As shown, embodiments of the present invention also provide a real-time early warning system for monitoring the health of building structures, comprising: The data acquisition module is used to collect real-time vibration signals through vibration sensors deployed at key parts of the building structure, and to collect temperature, humidity and wind speed data of the structural environment based on environmental monitoring devices. The computation module is used to take wind speed data as a reference noise input, perform adaptive noise cancellation calculation on the real-time vibration signal, and generate structural characteristic vibration signal; The extraction module is used to perform frequency domain transformation on the structural characteristic vibration signal and extract the abnormal harmonic energy distribution characteristics within the preset frequency band; The processing module is used to input the abnormal harmonic energy distribution characteristics into a preset structural damage analysis model, generate an initial failure probability prediction value P, calculate the deviation between P and the actual state based on the actual damage state label in the real-time monitoring data, and solve the gradient vector ∇J of the model parameters in reverse through a preset probability deviation metric function; dynamically adjust the weight parameters of the structural damage analysis model in the opposite direction of the gradient vector ∇J with a preset update step size; repeat the iterative process until the change in the probability deviation metric function is less than a set tolerance threshold, and generate a failure probability value E. The generation module is used to automatically generate a graded early warning command that is positively correlated with the failure probability value when the failure probability value exceeds a preset threshold, thereby driving the corresponding emergency response equipment to start. It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0129] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0130] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0131] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for real-time early warning of building structural health monitoring, characterized in that, The method includes: Step 1: Real-time vibration signals are collected by vibration sensors deployed at key parts of the building structure, and temperature, humidity and wind speed data of the structural environment are collected by environmental monitoring devices. Step 2: Using wind speed data as reference noise input, perform adaptive noise cancellation calculation on the real-time vibration signal to generate structural characteristic vibration signal; Step 3: Perform frequency domain transformation on the structural characteristic vibration signal to extract the abnormal harmonic energy distribution characteristics within the preset frequency band; Step 4: Input the abnormal harmonic energy distribution characteristics into the preset structural damage analysis model to generate an initial failure probability prediction value P. Based on the actual damage state label in the real-time monitoring data, calculate the deviation between P and the actual state. Solve the gradient vector ∇J of the model parameters in reverse using the preset probability deviation metric function. Dynamically adjust the weight parameters of the structural damage analysis model in the opposite direction of the gradient vector ∇J with a preset update step size. Repeat the iterative process until the change in the probability deviation metric function is less than the set tolerance threshold to generate the failure probability value E. Step 5: When the failure probability value exceeds the preset threshold, an automatic tiered early warning command that is positively correlated with the failure probability value is generated, driving the corresponding emergency response equipment to start.

2. The method for real-time early warning of building structural health monitoring according to claim 1, characterized in that, Step 1: Real-time vibration signals are collected using vibration sensors deployed at key parts of the building structure. Environmental monitoring devices are used to collect temperature, humidity, and wind speed data of the structural environment, including: Step 11: Divide the raw temperature, humidity and wind speed data collected in real time by the environmental monitoring device by the preset corresponding parameter reference values ​​to generate a dimensionless normalized environmental parameter vector. Step 12: Input the preset structural material thermal expansion-wet expansion coupled response model according to the normalized environmental parameter vector, and output the environmental compensation coefficient matrix of the vibration signal; Step 13: The environmental compensation coefficient matrix and the original real-time vibration signal collected by the vibration sensor are time-aligned according to a unified timestamp sequence. The aligned compensation coefficient matrix is ​​expanded into a three-dimensional compensation tensor, and the original vibration signal is reconstructed into a matching three-dimensional vibration signal tensor. Channel separation is performed on the constructed three-dimensional compensation tensor to generate intermediate signals for temperature compensation and humidity compensation. The intermediate signals for temperature compensation and humidity compensation are then subjected to tensor shrinkage to obtain a fused compensation signal. The fused compensation signal is normalized using the reference vibration energy value under the undamaged state of the structure as the normalization coefficient. The signal is then scaled independently according to the sensor position to generate a standardized vibration signal.

3. The method for real-time early warning of building structural health monitoring according to claim 2, characterized in that, Step 2: Using wind speed data as reference noise input, perform adaptive noise cancellation calculations on the real-time vibration signal to generate structural characteristic vibration signals, including: Step 21: Use the standardized vibration signal as the main input channel and extract the normalized wind speed data as the reference noise input channel; Step 22: Construct a finite-length unit impulse response filter and generate an initial weight coefficient vector according to a preset order; Step 23: Take the first N consecutive normalized wind speed data to form a reference noise vector. Perform a convolution operation between the reference noise vector and the filter weight coefficient vector to generate a wind-induced vibration noise estimate. Subtract the wind-induced vibration noise estimate from the standardized vibration signal sample value to obtain the primary error signal. Calculate the weight correction amount based on the product of the reference noise vector and the primary error signal, combined with the iteration step size parameter. Add the weight correction amount to the filter weight coefficient vector to achieve dynamic updating. Step 24: Rearrange all primary error signals according to the time series to generate denoised structural characteristic vibration signals.

4. The method for real-time early warning of building structural health monitoring according to claim 3, characterized in that, Step 3: Perform frequency domain transformation on the structural characteristic vibration signal to extract the abnormal harmonic energy distribution characteristics within the preset frequency band, including: Step 31: Perform a fast Fourier transform operation on the structural characteristic vibration signal to convert the time domain signal into a frequency domain signal; Step 32: Based on a preset specific frequency range that is sensitive to structural damage, extract all frequency components located within the preset frequency band from the frequency domain signal to generate a sub-frequency domain signal. Step 33: Divide the preset frequency band covered by the sub-frequency domain signal into multiple continuous sub-frequency bands. For each sub-frequency band, calculate the sum of the squared amplitude values ​​of all frequency components within the sub-frequency band to obtain the original energy value of the sub-frequency band. Calculate the sum of the original energy values ​​of all sub-frequency bands. Divide the original energy value of each sub-frequency band by the sum to obtain the normalized energy ratio of the sub-frequency band energy to the total energy of the preset frequency band. Step 34: Obtain the normalized energy ratio value of each sub-band calculated under the same preset frequency band and the same sub-band division method when the structure is in a healthy state without damage, and generate the reference normalized energy ratio; subtract the reference normalized energy ratio of the sub-band from the normalized energy ratio of each sub-band in the monitoring period to obtain the energy ratio deviation value; arrange and combine the energy ratio deviation values ​​according to the order of the sub-bands to obtain the abnormal harmonic energy distribution feature vector.

5. The method for real-time early warning of building structural health monitoring according to claim 4, characterized in that, Step 4: Input the abnormal harmonic energy distribution characteristics into the preset structural damage analysis model to generate an initial failure probability prediction value P. Based on the actual damage state label in the real-time monitoring data, calculate the deviation between P and the actual state. Solve the gradient vector ∇J of the model parameters in reverse through the preset probability deviation measurement function. In the opposite direction of the gradient vector ∇J, dynamically adjust the weight parameters of the structural damage analysis model with a preset update step size. Repeat the iterative process until the change in the probability deviation metric function is less than a set tolerance threshold, generating a failure probability value E, including: Step 41: Input the abnormal harmonic energy distribution feature vector into the preset structural damage analysis model. The model generates an initial failure probability prediction value P through multi-layer nonlinear feature mapping. Based on the actual damage state label of the corresponding timestamp in the real-time monitoring database, calculate the absolute deviation value between the initial prediction value P and the actual state label. Step 42: Using a preset probability deviation metric function, calculate the influence strength of all weight parameters of the model on the deviation in reverse, and generate the gradient vector ∇J; In the opposite direction of the gradient vector ∇J, adjust the weight parameters of each layer of the model synchronously with a preset learning step size. Step 43: Continuously update the model parameters using the new input feature vector and labels, and ensure that the change in the output value of the probability deviation metric function is less than the set tolerance threshold after two consecutive iterations.

6. The method for real-time early warning of building structural health monitoring according to claim 5, characterized in that, Step 5: When the failure probability value exceeds a preset threshold, an automatically generated tiered early warning command positively correlated with the failure probability value is generated to drive the corresponding emergency response equipment to start, including: Step 51: Receive the failure probability value E, and compare the P value with the preset three-level probability thresholds level by level; Primary warning: If the E value exceeds threshold 1 but does not reach threshold 2, a yellow warning instruction is generated, which includes the probability value and a description of "minor structural damage"; Intermediate warning: If the P value exceeds threshold 2 but does not reach threshold 3, an orange warning instruction is generated, which adds a description of "risk of local component failure" and evacuation suggestions; Advanced warning: If the E value exceeds threshold 3, a red warning instruction is generated, which adds a description of "risk of overall collapse" and an emergency response plan. Step 52: Integrate the warning level, probability value, timestamp, and damage status description into a structured instruction package; parse the combination of emergency equipment to be activated according to the warning level. The yellow instruction indicates the activation of the audible and visual alarm and the monitoring data upload module; the orange instruction indicates the additional activation of the local area power cut-off device and the evacuation broadcast system; the red instruction indicates the additional activation of the building-wide emergency power cut-off device and the fire linkage control system. Step 53: Send instruction packets to the corresponding devices through the IoT control bus, drive the devices to execute preset emergency actions in real time, collect the execution status signals of the emergency devices, bind the device status with the warning instructions, and send them back to generate an emergency response log.

7. The method for real-time early warning of building structural health monitoring according to claim 6, characterized in that, Step 6: Solve the gradient vector ∇J of the model parameters in reverse using the preset probability deviation metric function; dynamically adjust the weight parameters of the structural damage analysis model in the opposite direction of the gradient vector ∇J with a preset update step size. Repeat the iterative process until the change in the probability deviation metric function is less than a set tolerance threshold, generating a failure probability value E, including: Step 61: Obtain the deviation between the predicted probability value P and the actual damage state label; input the deviation value into the preset probability deviation metric function, backtrack from the model output layer to the input layer layer by layer, and calculate the contribution intensity of each weight parameter to the deviation value; summarize the contribution intensity values ​​of all layer weight parameters to generate a multidimensional gradient vector ∇J. Step 62: Determine the direction of increase or decrease for each weight parameter based on the sign of each component of the gradient vector ∇J. Using the preset global update step size as a benchmark, calculate the specific adjustment amount for each weight parameter according to the absolute value ratio of each component of the gradient vector ∇J. Perform synchronous addition and subtraction operations on all weight parameters of the model to generate a new generation of weight parameter set. Step 63: Using the new generation weight parameter set, reprocess the same set of abnormal harmonic energy characteristics to generate a new failure probability value E. Input the new probability value E and the actual damage label into the probability deviation metric function to output the new generation deviation metric value. Calculate the absolute difference between the new generation deviation metric value and the result of the previous iteration. If the difference is less than the set tolerance threshold, the model is determined to have converged. Otherwise, the new generation weight parameter set is used to cover the model, and the iteration continues.

8. A real-time early warning system for monitoring the health of building structures, wherein the system implements the method as described in any one of claims 1 to 7, characterized in that, include: The data acquisition module is used to collect real-time vibration signals through vibration sensors deployed at key parts of the building structure, and to collect temperature, humidity and wind speed data of the structural environment based on environmental monitoring devices. The computation module is used to take wind speed data as a reference noise input, perform adaptive noise cancellation calculation on the real-time vibration signal, and generate structural characteristic vibration signal; The extraction module is used to perform frequency domain transformation on the structural characteristic vibration signal and extract the abnormal harmonic energy distribution characteristics within the preset frequency band; The processing module is used to input the abnormal harmonic energy distribution characteristics into a preset structural damage analysis model, generate an initial failure probability prediction value P, calculate the deviation between P and the actual state based on the actual damage state label in the real-time monitoring data, and solve the gradient vector ∇J of the model parameters in reverse through a preset probability deviation metric function; dynamically adjust the weight parameters of the structural damage analysis model in the opposite direction of the gradient vector ∇J with a preset update step size; repeat the iterative process until the change in the probability deviation metric function is less than a set tolerance threshold, and generate a failure probability value E. The generation module is used to automatically generate a graded early warning command that is positively correlated with the failure probability value when the failure probability value exceeds a preset threshold, thereby driving the corresponding emergency response equipment to start.

9. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to perform the method as described in any one of claims 1 to 7.

Images