Early failure warning method for bridge erecting machine

Abstract

The application provides a kind of early fault warning method of bridge machine, for the technical problems that weak fault characteristics are easy to be covered by strong noise under complex operation condition, traditional model generalization ability is weak and early warning threshold is rigid, the application first collects multiple source operation data and carries out denoising reconstruction using variational mode decomposition algorithm, realizes weak fault signal enhancement;Secondly, extract multi-dimensional statistical features and use local preserving projection algorithm to reduce dimension, obtain low-dimensional sensitive feature vector;Then construct adaptive kernel spectrum clustering model, use data-driven scale factor and introduce mahalanobis structure penalty similarity matrix to identify high-precision early weak fault state;Finally, combined with the mechanism of non-linear fatigue damage of equipment and real-time working condition variable updates dynamic warning threshold, and then trigger multi-level response to prevent false alarm.The application greatly improves the purity of feature extraction and cross-condition generalization ability in strong noise environment, realizes low false alarm, zero false alarm and accurate active early warning.

Description

Technical Field

[0001] This invention relates to the field of bridge erecting machines, and in particular to a method for early fault warning of bridge erecting machines. Background Technology

[0002] As a core large-scale engineering equipment in the construction of highways, railways, and urban integrated transportation hubs, the safety of bridge erecting machines directly affects the progress of the entire project and the safety of personnel and property. In actual construction operations, bridge erecting machines typically face extremely complex working environments, including heavy-load lifting, frequent luffing, joint operations with other machine types, and severe external weather conditions. Under long-term heavy loads and alternating stress, the key load-bearing structural components, drive mechanisms, and connecting parts of the bridge erecting machine inevitably develop fatigue damage and micro-cracks. However, in the early stages of bridge erecting machine operation, these weak fault signals are often completely drowned out by the massive background noise of mechanical operation and complex environmental interference signals, exhibiting extremely low signal-to-noise ratio characteristics. Traditional signal processing methods mostly rely on single-dimensional time-domain or frequency-domain analysis, which is prone to failure in strong noise environments, making the extraction of early weak fault features like finding a needle in a haystack, and difficult to achieve effective identification.

[0003] Meanwhile, existing bridge erecting machine fault early warning models generally suffer from inherent defects such as poor adaptability across different machine models and working conditions. The data distribution characteristics of different bridge erecting machine models, or even the same equipment at different construction stages, vary greatly. Traditional data-driven models often fall into the trap of overfitting, and once the operating scenario changes, the model's recognition accuracy drops drastically, significantly increasing the cost of on-site redeployment and debugging. Furthermore, traditional early warning systems mostly use fixed alarm thresholds set by expert experience. This rigid threshold setting mechanism is completely detached from the dynamic stress state of the bridge erecting machine, which changes in real time with external working conditions such as lifting load, operating frequency, and wind load. When the equipment is under high load or severe working conditions, fixed thresholds are prone to frequent false alarms, interfering with the normal judgment of on-site personnel; while when the equipment is in the later stages of fatigue accumulation, rigid thresholds may lead to missed detections of fatal faults, failing to truly fulfill the safety backup role of the early warning system. Summary of the Invention

[0004] The main objective of this invention is to provide an early fault warning method for bridge erecting machines. This application addresses the technical problems of early weak faults of bridge erecting machines being difficult to extract and identify due to strong noise masking them under complex operating conditions, as well as the poor cross-condition adaptability and frequent false alarms and missed alarms caused by the weak model generalization ability and rigid warning thresholds of existing early warning systems.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: an early fault warning method for bridge erecting machines, comprising the following steps: S1. Real-time acquisition of multi-source operating data of key parts of the bridge erecting machine through sensor array, and decomposition and enhancement processing of the acquired signals using mode decomposition algorithm to extract weak fault components hidden in strong noise background. S2. Extract physical quantities reflecting fault characteristics from the enhanced component signals in multiple dimensions, and use feature fusion and dimensionality reduction techniques to compress the high-dimensional feature space into a low-dimensional sensitive feature vector that can accurately characterize the fault state. S3. Input the low-dimensional sensitive feature vector into the pre-constructed adaptive nuclear spectrum clustering model, calculate the similarity between samples and perform cluster analysis to realize the automatic identification and classification of early weak fault states of the bridge erecting machine. S4. Based on the structural failure mechanism model of the bridge erecting machine, the warning threshold is dynamically calculated and updated in real time according to the current real-time working conditions. When the identified fault indicators exceed the warning threshold, a multi-level warning mechanism is triggered and the failure evolution trend is predicted.

[0006] In the preferred embodiment, step S1, which involves using a mode decomposition algorithm to decompose and enhance the signal, includes: First, the variational mode decomposition algorithm is used to decompose the preprocessed vibration signal into several eigenmode function components with specific center frequencies; Calculate the correlation coefficient and kurtosis value of each intrinsic mode function component, and screen out the effective components that contain the main fault characteristic information; The selected effective components are reconstructed, and broadband random noise is removed, thereby enhancing the weak fault characteristic signal.

[0007] In the preferred embodiment, the steps of extracting physical quantities and performing dimensionality reduction in step S2 include: The statistical characteristics of the reconstructed signal are calculated from three dimensions: time domain, frequency domain, and time-frequency domain. The time domain characteristics include the effective value, peak factor, and impulse factor; the frequency domain characteristics include the centroid frequency and root mean square frequency; and the time-frequency domain characteristics are composed of the energy values ​​of the wavelet packet decomposition coefficients. The above features are combined to construct an initial high-dimensional fault feature matrix; A local preservation projection algorithm based on manifold learning is used to linearly map the high-dimensional fault feature matrix, which reduces feature redundancy while preserving local structural information of the data and obtains low-dimensional sensitive feature vectors.

[0008] In the preferred embodiment, the construction and identification process of the adaptive nuclear spectrum clustering model in step S3 relies on the feature processing unit, matrix calculation acceleration module, and clustering decision maker built into the early warning system. The steps include: The feature processing unit receives a set of low-dimensional sensitive fault feature vectors after dimensionality reduction, extracts the local spatial distribution density for each feature vector in the set, and derives a data-driven adaptive scaling factor based on the historical operation distribution of the bridge erecting machine. Specifically, for any given feature vector... eigenvectors Its adaptive scaling factor The derivation formula is as follows: ; in, This is the working condition compensation coefficient that varies with the lifting load of the bridge erecting machine. To prevent infinitely large smoothing constants, The total number of samples for the extracted feature vector. The trace of the feature covariance matrix of the historical health status data of the bridge erecting machine; The matrix computation acceleration module constructs a hybrid kernel similarity matrix containing a Markov structure penalty mechanism based on the adaptive scaling factor of each eigenvector. This matrix is ​​then used to calculate the similarity between any two eigenvectors. and Similarity elements between Establish a global similarity graph, with similarity elements. The derivation and calculation formula is as follows: ; in, It is the minimum constant. The dynamic weighting parameters are set based on the stress transfer gradient of the key structure of the bridge erecting machine. The second term on the right side of the equation is the structural correlation projection represented by the inverse matrix of the covariance matrix. The matrix computation acceleration module calculates the diagonal matrix containing node connectivity based on the global similarity graph. Furthermore, the optimized Laplace matrix incorporating a mechanical vibration noise regularization term is derived. The calculation formula is: ; in, It is the identity matrix. It is an angle matrix whose diagonal elements are similarity matrices. The sum of the elements in the corresponding row. This is the noise suppression coefficient. The base broadband noise projection matrix is ​​extracted in advance from the no-load operation condition of the bridge erecting machine; The matrix computation acceleration module employs the implicit restart Arnoldi iterative algorithm to optimize the Laplace matrix. Perform eigenvalue decomposition, solve and extract the frontier values. The eigenvectors corresponding to the non-zero minimum eigenvalues ​​are concatenated column by column and normalized by row to construct a low-dimensional orthogonal indicator matrix. The clustering decision unit receives the low-dimensional orthogonal indicator matrix and inputs its row vectors as smoothed data points after noise reduction into the density-based spatial clustering algorithm module. Based on the density connectivity and distribution boundary of the clusters, the current state features are automatically weighted and labeled. The stable main clusters are determined to be in a normal state, and the discrete micro-clusters that deviate from the main clusters and exhibit nonlinear decay in local density are determined to be in an early weak fault state. Finally, the weak fault identification result is output.

[0009] In the preferred scheme, the specific calculation steps and algorithm implementation process for dynamically setting the early warning threshold based on the fault mechanism and triggering multi-level early warnings in step S4 rely on the damage accumulation calculation engine, real-time operating condition compensation module, and decision-making executor deployed on the control terminal of the early warning system. The steps include: By extracting the transient stress time series of the key load-bearing structure of the bridge erecting machine using a damage accumulation calculation engine, and combining this with the nonlinear fatigue damage mechanism under high-intensity operation of lifting machinery, the basic health threshold is derived. The calculation formula is: ; in, This represents the initial fatigue attenuation coefficient of the bridge erecting machine's structural components. For transient dynamic stress, The yield strength of the material. A nonlinear fatigue index related to material density and microscopic defects. The activation energy for the propagation of microcracks in the structure. Let be the ideal gas constant. The absolute temperature of the transient environment; By collecting data on the current lifting weight of the bridge erecting machine, the operating frequency of the main beam, and wind load disturbance variables through a real-time working condition compensation module, a data-driven dynamic working condition compensation factor is constructed. The derivation formula is as follows: ; in, This represents the nonlinear coupling penalty coefficient between the load and frequency. This is to determine the real-time lifting weight of the bridge erecting machine. Main beam operating frequency, This represents the weighting of aerodynamic disturbances caused by wind loads. For real-time wind speed, The maximum wind speed for the safe operation of the bridge erecting machine; The basic health threshold is adjusted through the real-time working condition compensation module. Compensation factor for dynamic operating conditions Feature mapping is performed, and a historical decay memory term is introduced in the time dimension to calculate and update the dynamic warning threshold at the current time point. The derived calculation formula is as follows: ; in, Forgetting factors in historical state memory, The total length of the historical early warning threshold time window extracted. For the first Warning threshold data for each historical period; The decision executor receives the weak fault membership index output by the clustering model. The indicators are input into the false alarm prevention judgment function based on a sliding time window, and the trigger condition formula is as follows: ; in, This represents the total step size of the sliding decision time window. For time steps The corresponding weak fault membership index, For time steps The corresponding dynamic early warning threshold, For sign determination function, The confidence level triggers the bottom line; When the triggering condition of the false alarm prevention judgment function is met, the decision executor extracts the nonlinear gradient difference of the current weak fault membership index exceeding the dynamic early warning threshold, matches it with the system's pre-set multi-level alarm rule base, and calls the communication middleware to push the corresponding level of alarm signal, protection instruction or interlock instruction to the terminal.

[0010] In the preferred embodiment, the actual deployment steps of this method in a computer system include: Environment setup steps: Configure the Linux operating system on the industrial control computer or NAS hardware device, and deploy the Docker containerized environment; Software installation steps: Install the Python runtime environment and algorithm dependency libraries including NumPy, SciPy, Scikit-learn, and PyTorch in a Docker container; Module integration steps: Write and deploy the data acquisition driver module, signal processing script, clustering prediction engine module, and Web visualization dashboard module; API call steps: Receive raw JSON data uploaded by the sensor via the REST API interface, and push the processed warning results to the remote monitoring terminal.

[0011] In the preferred scheme, the multi-level response steps after the early warning is triggered include: Level 1 warning: When the fault indicator is in the first threshold range, the system sends an SMS reminder to the maintenance personnel and records the location of the fault. Level 2 warning: When the fault indicator is in the second threshold range, the system will issue an alarm through the sound and light alarm and flash the prompt on the visual display board, while also suggesting to reduce the operating speed of the bridge erecting machine; Level 3 warning: When the fault indicator exceeds the maximum safety limit, the system automatically sends a shutdown command to the PLC controller and generates a fault analysis report.

[0012] A bridge erecting machine early fault warning system includes: The multi-source data acquisition and enhancement module has its signal output end connected to the feature fusion and dimensionality reduction module, which is used to transmit the reconstructed intrinsic mode function signal to the feature fusion and dimensionality reduction module. The feature fusion and dimensionality reduction module has its signal input end connected to the multi-source data acquisition and enhancement module, and its signal output end connected to the fault state identification module, which is used to transmit the mapped low-dimensional sensitive feature vector to the fault state identification module. The fault status identification module has a signal input end connected to the feature fusion and dimensionality reduction module, and a signal output end connected to the dynamic threshold early warning module. It is used to transmit the weak fault membership index generated by cluster analysis to the dynamic threshold early warning module. The dynamic threshold early warning module has its signal receiving end connected to the fault status identification module and the external operating condition sensor, respectively. It is used to receive weak fault membership indicators and real-time operating condition parameters, and output early warning control commands after logical judgment.

[0013] A computer device includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the above-mentioned early fault warning method for bridge erecting machines.

[0014] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described early fault warning method for bridge erecting machines.

[0015] This invention provides an early fault warning method for bridge erecting machines. By deeply integrating multi-source data acquisition with variational mode decomposition technology, this invention effectively removes broadband random noise during signal preprocessing. Subsequently, it fuses multi-dimensional features such as time-domain, frequency-domain, and wavelet packet energy, and introduces a local preserving projection algorithm for manifold dimensionality reduction. This series of processes greatly removes redundant information, successfully extracting highly sensitive low-dimensional fault features even under strong interference environments. It fundamentally overcomes the industry challenge of extracting weak fault signals under low signal-to-noise ratio conditions, significantly improving the ability to detect early-stage equipment problems.

[0016] The adaptive kernel clustering model constructed in this invention eliminates the dependence on massive amounts of labeled data. Through data-driven adaptive scaling factors and hybrid kernel similarity calculations incorporating a Markovian structure penalty mechanism, the model can automatically optimize based on the local spatial density of the bridge erecting machine's real-time features. This enables the early warning system to exhibit extremely excellent generalization ability and robustness when facing different models of bridge erecting machines or drastically different construction environments, significantly reducing the model tuning cost and deployment difficulty when applying across different machine models.

[0017] This invention creatively integrates the nonlinear fatigue damage mechanism of bridge erecting machine structures with real-time dynamic operating condition variables. The warning threshold is no longer a cold, fixed value, but a dynamic baseline that can adaptively compensate in real time for wind speed, lifting weight, operating frequency, and historical attenuation memory. Combined with a false alarm prevention judgment function based on a sliding time window, the system can extremely accurately filter out false anomalies caused by transient operating condition fluctuations. While ensuring extremely high warning accuracy, it provides on-site construction personnel and equipment maintenance teams with ample time for fault diagnosis and emergency response, completely changing the passive situation of traditional equipment maintenance of "post-event repair" or "blind scheduled maintenance."

[0018] The hardware and software deployment scheme of this invention adopts containerization technology and modular architecture design, realizing complete decoupling of data acquisition, feature extraction, model judgment, and multi-level response. This not only ensures lightweight and stable operation of the system on industrial control computers or edge computing devices, but also provides standardized and efficient interfaces for subsequent functional iterations and large-scale clustered device monitoring, possessing extremely high industrial promotion value and engineering practicality. Attached Figure Description

[0019] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the data acquisition and signal enhancement process of this invention; Figure 2 This is a flowchart of the feature fusion and dimensionality reduction process of this invention; Figure 3 This is a flowchart of the adaptive nuclear spectrum clustering identification process of the present invention; Figure 4 This is a flowchart of the dynamic threshold and multi-level early warning process of the present invention. Detailed Implementation

[0020] Example 1 like Figure 1-4 As shown, a method for early fault warning of a bridge erecting machine includes the following steps: S1. Real-time acquisition of multi-source operating data of key parts of the bridge erecting machine through sensor array, and decomposition and enhancement processing of the acquired signals using mode decomposition algorithm to extract weak fault components hidden in strong noise background. S2. Extract physical quantities reflecting fault characteristics from the enhanced component signals in multiple dimensions, and use feature fusion and dimensionality reduction techniques to compress the high-dimensional feature space into a low-dimensional sensitive feature vector that can accurately characterize the fault state. S3. Input the low-dimensional sensitive feature vector into the pre-constructed adaptive nuclear spectrum clustering model, calculate the similarity between samples and perform cluster analysis to realize the automatic identification and classification of early weak fault states of the bridge erecting machine. S4. Based on the structural failure mechanism model of the bridge erecting machine, the warning threshold is dynamically calculated and updated in real time according to the current real-time working conditions. When the identified fault indicators exceed the warning threshold, a multi-level warning mechanism is triggered and the failure evolution trend is predicted.

[0021] In the preferred embodiment, step S1, which involves using a mode decomposition algorithm to decompose and enhance the signal, includes: First, the variational mode decomposition algorithm is used to decompose the preprocessed vibration signal into several eigenmode function components with specific center frequencies; Calculate the correlation coefficient and kurtosis value of each intrinsic mode function component, and screen out the effective components that contain the main fault characteristic information; The selected effective components are reconstructed, and broadband random noise is removed, thereby enhancing the weak fault characteristic signal.

[0022] The detailed operation process and mathematical basis of the specific implementation method for acquiring signals and performing mode decomposition and signal enhancement in step S1 are described below.

[0023] The system first acquires the preprocessed vibration signal and then processes it using a variational mode decomposition (VMD) algorithm. VMD is a completely non-recursive signal decomposition and estimation method. Its core idea is to construct and solve a constrained variational problem. This algorithm adaptively decomposes the original broadband complex signal into a given number of eigenmode function components, each closely revolving around its respective center frequency, thus effectively avoiding the mode aliasing phenomenon commonly found in traditional empirical mode decomposition. The total number of modes to be decomposed is set to... The set of intrinsic mode function components obtained after decomposition can be expressed as: The corresponding set of center frequencies is .in The total number of modes preset by the system based on the structure of the bridge erecting machine's transmission system. The decomposition yields the first... Each intrinsic mode function component Indicates the relationship with the first The center frequencies corresponding to the components of each intrinsic mode function.

[0024] After obtaining several intrinsic mode function (IMF) components, the system further calculates the correlation coefficient and kurtosis value for each component, using this as a quantitative basis for feature selection. The correlation coefficient measures the degree of linear correlation between the decomposed IMF components and the original vibration signal, reflecting the proportion of the original true signal characteristics retained by that component. For the [specific component name]... Each intrinsic mode function component Compared with the original preprocessed signal The formula for calculating its correlation coefficient is defined as follows: ; In the formula, Indicates the first The correlation coefficient of each component This represents the total length of the signal's data sampling points. This represents the time step index number of the current discrete sampling point. Indicates the first The component in the first The magnitude of each sampling point Indicates the first The arithmetic mean of the amplitudes of all sampling points of each component This indicates that the original preprocessed signal is at the 1st... The magnitude of each sampling point This represents the arithmetic mean of the amplitudes of all sampling points of the original preprocessed signal.

[0025] Simultaneously, the system introduces kurtosis as a second screening dimension. Kurtosis, a dimensionless statistical parameter highly sensitive to the transient impact characteristics of a signal, is used to assess the presence of high-frequency impact components in the signal caused by early, minor damage to mechanical parts. Each intrinsic mode function component The formula for calculating kurtosis is defined as follows: ; In the formula, Indicates the first The system calculates the kurtosis values ​​of each component, while maintaining the same physical meaning as the variables in the correlation coefficient formula. Based on the calculated results, the system extracts the intrinsic mode function components whose correlation coefficients are greater than the preset baseline correlation coefficients and whose kurtosis values ​​are greater than the preset baseline kurtosis values. These components that satisfy the dual joint condition are marked as valid components containing the main fault characteristic information.

[0026] For all selected valid components, the system performs a final signal reconstruction operation. The system directly linearly adds the eigenmode function components marked as valid components in the time domain sequence to generate a denoised reconstructed signal. The mathematical expression of the reconstruction formula is: ; In the formula, This represents the reconstructed time-domain enhanced signal. Represents a continuous time independent variable. This represents the set of valid component indices that have passed the joint test of correlation coefficient and kurtosis value. Components not included in the set of valid component indices are determined to be interference terms mainly composed of external broadband random noise or useless high-order harmonics, and are directly discarded by the system during the summation reconstruction process.

[0027] Implementing the above data processing and enhancement steps has significant technical benefits. Because bridge erecting machines not only bear complex alternating mechanical loads during construction but are also exposed to strong aerodynamic disturbances and background noise from motor operation, transient impact signals from early structural component micro-cracks or initial bearing wear are easily masked, resulting in extremely low signal-to-noise ratios. This application overcomes the shortcomings of traditional filtering methods in extracting weak signals by introducing adaptive frequency band division capabilities through variational mode decomposition. By combining correlation coefficients to ensure undistorted reconstructed signals and utilizing kurtosis values ​​to keenly capture fault impact waveforms, it successfully and accurately removes weak fault features mixed in with strong background interference. This series of operations physically cuts off the transmission path of broadband random noise, achieving high-fidelity and high signal-to-noise ratio enhancement of early weak fault signals. The above scheme not only effectively overcomes the technical pain point of difficulty in capturing weak fault features under complex working conditions but also lays a high-quality data foundation for the accurate extraction of subsequent spatial features and the stable convergence of clustering recognition models, significantly improving the sensitivity of the entire early warning system to early hazard detection.

[0028] In the preferred embodiment, the steps of extracting physical quantities and performing dimensionality reduction in step S2 include: The statistical characteristics of the reconstructed signal are calculated from three dimensions: time domain, frequency domain, and time-frequency domain. The time domain characteristics include the effective value, peak factor, and impulse factor; the frequency domain characteristics include the centroid frequency and root mean square frequency; and the time-frequency domain characteristics are composed of the energy values ​​of the wavelet packet decomposition coefficients. The above features are combined to construct an initial high-dimensional fault feature matrix; A local preservation projection algorithm based on manifold learning is used to linearly map the high-dimensional fault feature matrix, which reduces feature redundancy while preserving local structural information of the data and obtains low-dimensional sensitive feature vectors.

[0029] The detailed operational process and mathematical basis for extracting physical quantities and performing dimensionality reduction in step S2, after comprehensive correction of letter conflicts, are explained below: The system first extracts time-domain features from the reconstructed enhanced signal, mainly including RMS value, peak factor, and impulse factor. To maintain consistency with global variables, the total length of the discrete data sampling points of the reconstructed signal is uniformly denoted as [missing information]. , No. The signal amplitude corresponding to each sampling point is denoted as . Valid values ​​are denoted as Its calculation formula is This indicator reflects the overall energy distribution and average fluctuation level of a signal over a period of time. The peak factor is denoted as... Its calculation formula is The numerator of this formula is the maximum abrupt change value in the absolute value sequence of the signal. This index is not limited by the magnitude of the absolute amplitude of the signal and can sensitively reflect whether a sudden mechanical impact occurs in the signal. The impulse factor is denoted as... Its calculation formula is The denominator of the formula is the arithmetic mean of the absolute values ​​of the entire sampling sequence. This index further enhances the system's ability to capture and amplify the early weak impact characteristics of the bridge erecting machine's gears or bearings.

[0030] Next, the system extracts frequency domain features, mainly including the centroid frequency and root mean square frequency. After converting the time-domain signal to the frequency domain using a Fourier transform, let the total number of frequency spectral lines be denoted as... , No. The frequency value corresponding to each spectral line is denoted as The corresponding amplitude spectrum amplitude is denoted as The frequency of the center of gravity is denoted as... Its calculation formula is This parameter characterizes the physical location of the centroid of the frequency distribution of the signal's spectral energy across the entire frequency band. The root mean square frequency is denoted as... Its calculation formula is This parameter precisely describes the degree of energy dispersion within the main frequency band and is highly sensitive to the bandwidth broadening phenomenon caused by the cumulative wear of mechanical structural components.

[0031] The system then extracts time-frequency domain features, specifically using the energy values ​​of wavelet packet decomposition coefficients as the characterization quantifier. The signal is then subjected to wavelet packet decomposition at a specific number of levels; let the total number of decomposition levels be denoted as . Then, multiple independent child nodes will be generated at the bottom layer. Let any number of child nodes be... The total number of discrete coefficients contained in each node is denoted as . The first node under this node The decomposition coefficients are denoted as The energy value of the wavelet packet decomposition coefficients corresponding to this node is denoted as... Its calculation formula is By calculating the energy value of each independent node, the system can accurately obtain the dynamic distribution law of the reconstructed signal energy changing over time in different sub-bands.

[0032] After completing the feature calculations across the three dimensions mentioned above, the system sequentially concatenates the feature indicators extracted within each sampling time period to form a high-dimensional column vector describing the current operating state. As the equipment operates continuously, the high-dimensional column vectors from multiple sampling periods are combined and concatenated to construct an initial high-dimensional fault feature matrix containing rich information on fault state evolution. Let this high-dimensional matrix be denoted as... The system employs a manifold-based local preserving projection algorithm for linear dimensionality reduction mapping. The total number of sample features extracted from the high-dimensional space is denoted as . This is absolutely consistent with the physical meaning of the total number of samples in the subsequent adaptive kernel clustering model. Let the th... A high-dimensional sample is denoted as , No. A high-dimensional sample is denoted as The algorithm first constructs a nearest neighbor weight matrix. When two samples are neighbors in the high-dimensional space, a corresponding similarity weight is assigned; otherwise, the weight is set to zero. Then, the diagonal matrix is ​​calculated. Its diagonal elements are strictly equal to the sum of the elements of the corresponding rows of the weight matrix. Further, the high-dimensional Laplacian matrix is ​​derived. .

[0033] The core of the local preservation projection algorithm lies in finding an optimal linear transformation projection matrix. This ensures that after a high-dimensional sample is mapped to a low-dimensional subspace, it still retains its original local nearest neighbor structure. Let the mapped low-dimensional sensitive feature vector be denoted as... This low-dimensional vector will be directly used as the feature vector of the input sample set of the adaptive kernel spectral clustering model in claim 4, and its algebraic mapping relationship with the high-dimensional samples is as follows: The objective function is to minimize the physical distance between locally similar samples in the low-dimensional space, which can be mathematically expressed as minimizing... At the same time, geometric scaling constraints must be met. ,in The identity matrix is ​​used. The projection transformation matrix is ​​obtained by solving the generalized eigenvalue problem. Then, a linear mapping is performed on the high-dimensional fault feature matrix to obtain a low-dimensional sensitive feature vector with significantly reduced redundant dimensions.

[0034] Implementing the aforementioned feature fusion and dimensionality reduction mapping steps yields significant technical benefits. The early, subtle fault mechanisms of bridge erecting machines under complex operating conditions are extremely complex. Single-dimensional features often only reflect partial information about fault evolution, easily leading to the omission of crucial early warning information. This application constructs a comprehensive high-dimensional feature matrix by combining the impact sensitivity in the time domain, the intuitiveness of energy distribution in the frequency domain, and the dynamic local detailing capabilities of the time and frequency domains. This ensures comprehensive coverage of early, subtle fault feature information from the underlying logic. Simultaneously, a local preserving projection algorithm is introduced for dimensionality reduction, which not only efficiently eliminates collinearity interference and useless redundant information between multidimensional statistical features but also keenly captures and perfectly preserves the inherent local manifold structure of the high-dimensional nonlinear dataset. This mechanism ensures that feature samples under the same degradation trend are more tightly clustered in the lower-dimensional space after dimensionality reduction, making the boundaries between different state types exceptionally clear. This provides a high-quality, sensitive data stream with extremely high purity and strong separability for the subsequent construction of an adaptive nuclear spectral clustering model, fundamentally eliminating the fuzzy judgment of the model under complex cross data, and greatly improving the accuracy of early identification of minor faults in bridge erecting machines and the efficiency of model execution.

[0035] Example 2 The construction and recognition process of the adaptive nuclear spectrum clustering model in step S3 relies on the feature processing unit, matrix calculation acceleration module, and clustering decision maker built into the early warning system. The steps include: The feature processing unit receives a set of low-dimensional sensitive fault feature vectors after dimensionality reduction, extracts the local spatial distribution density for each feature vector in the set, and derives a data-driven adaptive scaling factor based on the historical operation distribution of the bridge erecting machine. Specifically, for any given feature vector... eigenvectors Its adaptive scaling factor The derivation formula is as follows: ; in, This is the working condition compensation coefficient that varies with the lifting load of the bridge erecting machine. To prevent infinitely large smoothing constants, The total number of samples for the extracted feature vector. The trace of the feature covariance matrix of the historical health status data of the bridge erecting machine; The matrix computation acceleration module constructs a hybrid kernel similarity matrix containing a Markov structure penalty mechanism based on the adaptive scaling factor of each eigenvector. This matrix is ​​then used to calculate the similarity between any two eigenvectors. and Similarity elements between Establish a global similarity graph, with similarity elements. The derivation and calculation formula is as follows: ; in, It is the minimum constant. The dynamic weighting parameters are set based on the stress transfer gradient of the key structure of the bridge erecting machine. The second term on the right side of the equation is the structural correlation projection represented by the inverse matrix of the covariance matrix. The matrix computation acceleration module calculates the diagonal matrix containing node connectivity based on the global similarity graph. Furthermore, the optimized Laplace matrix incorporating a mechanical vibration noise regularization term is derived. The calculation formula is: ; in, It is the identity matrix. It is an angle matrix whose diagonal elements are similarity matrices. The sum of the elements in the corresponding row. This is the noise suppression coefficient. The base broadband noise projection matrix is ​​extracted in advance from the no-load operation condition of the bridge erecting machine; The matrix computation acceleration module employs the implicit restart Arnoldi iterative algorithm to optimize the Laplace matrix. Perform eigenvalue decomposition, solve and extract the frontier values. The eigenvectors corresponding to the non-zero minimum eigenvalues ​​are concatenated column by column and normalized by row to construct a low-dimensional orthogonal indicator matrix. The clustering decision unit receives the low-dimensional orthogonal indicator matrix and inputs its row vectors as smoothed data points after noise reduction into the density-based spatial clustering algorithm module. Based on the density connectivity and distribution boundary of the clusters, the current state features are automatically weighted and labeled. The stable main clusters are determined to be in a normal state, and the discrete micro-clusters that deviate from the main clusters and exhibit nonlinear decay in local density are determined to be in an early weak fault state. Finally, the weak fault identification result is output.

[0036] The detailed calculation process and mathematical logic of the specific implementation method for constructing the adaptive nuclear spectrum clustering model and using fault features for weak fault identification in step S3 are explained below.

[0037] The system's built-in feature processing unit first receives the set of low-dimensional sensitive fault feature vectors after dimensionality reduction. In the complex operating environment of the bridge erecting machine, the feature distribution under different working conditions exhibits significant non-uniformity. To overcome the shortcomings of traditional clustering algorithms where fixed scale parameters lead to the neglect of local sparse fault features, the system extracts the local spatial distribution density for each feature vector in the set and derives a data-driven adaptive scaling factor based on the historical operating distribution of the bridge erecting machine. For any given... eigenvectors Its adaptive scaling factor The derivation formula is as follows: ; In the derivation formula of the adaptive scaling factor, This is a working condition compensation coefficient that varies with the lifting load of the bridge erecting machine. This coefficient can dynamically adjust the scaling ratio according to the real-time lifting weight. To prevent the appearance of a smoothing constant with an infinite or zero value inside the logarithmic function; The total number of feature vector samples extracted; For the set of the first 1 eigenvector; Indicates the first The eigenvector and the eigenvector The squared Euclidean distance between the eigenvectors; denominator This constitutes the kernel density estimate, which measures local density. The smaller the local spatial density, the larger the derived scale factor. The covariance matrix of historical health status data of bridge erecting machines; The trace of the covariance matrix of the historical health characteristics represents the overall energy distribution range of the bridge erecting machine in the characteristic space under healthy conditions.

[0038] Subsequently, the system's built-in matrix calculation acceleration module constructs a hybrid kernel similarity matrix containing a Markovial structure penalty mechanism based on the adaptive scaling factor calculated independently for each eigenvector. Traditional Gaussian kernel functions only consider Euclidean distance, making it difficult to reflect the directional correlation of mechanical components inside the bridge erecting machine during stress transmission. and Similarity elements between To build a global similarity graph, similarity elements The derivation and calculation formula is as follows: ; In this similarity derivation formula, the first term on the right-hand side of the equation is the improved adaptive Gaussian kernel function, where... To prevent the minimum constant of the denominator being zero; the second term on the right side of the equation is the structural correlation projection part represented by the inverse of the covariance matrix, i.e., the Markov structure penalty mechanism; The dynamic weighting parameter is set based on the stress transfer gradient of the key structure of the bridge erecting machine, and is used to balance the proportion of local density similarity and global structural correlation in the total similarity. This is the inverse matrix of the historical health feature covariance matrix. By introducing this inverse matrix for feature space transformation, the dimensional differences and correlation interference between the various dimensions of the features can be effectively eliminated, making the similarity calculation more consistent with the actual physical damage evolution and transmission law of the bridge erecting machine.

[0039] After establishing the global similarity map, the matrix calculation acceleration module extracts and calculates the diagonal matrix containing node connectivity based on the map. Based on this, the system derives the optimized Laplace matrix that incorporates a mechanical vibration noise regularization term. The specific calculation formula is as follows: ; In this optimized formula for calculating the Laplacian matrix, It is the standard identity matrix; Represents all similarity elements The complete global similarity matrix formed by these elements; For an angle matrix, the elements on its diagonal are strictly equal to the similarity matrix. The sum of all elements in the corresponding row accurately reflects the spatial connectivity of each feature node in the global similarity graph; This is the noise suppression coefficient set by the system based on the on-site working conditions; The base broadband noise projection matrix is ​​extracted in advance from the no-load operation condition of the bridge erecting machine; This is the transpose of the broadband noise projection matrix of the base.

[0040] The matrix computation acceleration module further employs the implicit restart Arnoldi iterative algorithm to optimize the Laplace matrix derived above. Eigenvalue decomposition is performed. This iterative algorithm significantly reduces device memory consumption and greatly improves solution speed in large-scale sparse matrix operations. The system uses this algorithm to solve and extract the eigenvalues. Extract these non-zero minimum eigenvalues ​​simultaneously. The system extracts a series of eigenvectors corresponding to each eigenvalue. These extracted eigenvectors are concatenated column-wise, and then the concatenated eigenma matrix is ​​normalized by row-wise to construct a low-dimensional orthogonal indicator matrix that can smoothly represent the boundary of the cluster space distribution.

[0041] Finally, the system's built-in clustering decision-maker receives the low-dimensional orthogonal indicator matrix. The clustering decision-maker treats each row vector in the indicator matrix as a smoothed, denoised data point and inputs it into the density-based spatial clustering algorithm module. This module automatically assigns state weights and labels the current operating state characteristics based on the density connectivity and distribution boundaries of the clusters formed by the input data points in the low-dimensional manifold space. In the specific partitioning decision logic, the system determines the stable, highly dense main cluster as the normal healthy state of the bridge erecting machine; for those discrete, small clusters that deviate from the main cluster and whose local spatial density exhibits non-linear decay, the system accurately identifies them as early, minor fault states. After completing all label partitioning, the clustering decision-maker finally outputs the minor fault identification results of the bridge erecting machine.

[0042] Implementing the aforementioned adaptive kernel clustering model construction and identification steps yields significant technical benefits. First, the feature data distribution of bridge erecting machines under different lifting and hoisting tasks and varying wind loads often exhibits highly unevenness. Traditional clustering algorithms use a globally fixed kernel function scale, which easily forces early, weak fault features at the margins to be classified as normal operating noise or directly ignored. The data-driven adaptive scale factor proposed in this application can automatically adjust the perception sensitivity according to the local spatial data density. A small scale is used in data-dense areas to accurately distinguish fine state structures, while a large scale is used in data-sparse areas to maintain the connectivity of the manifold space, thus effectively avoiding the loss of weak, hidden fault feature information from an algorithmic perspective. Second, a Markovian structure penalty mechanism is creatively introduced into the hybrid kernel similarity matrix. This not only deeply considers the closeness of the Euclidean distance between feature vectors in high-dimensional mathematical space but also substantially incorporates prior knowledge of stress transmission in the physical mechanical structure of the bridge erecting machine. This dual-driven fusion mechanism of physical mechanism and pure data ensures that the division of cluster state boundaries highly matches the actual fatigue degradation evolution trajectory of large engineering machinery components. Furthermore, the creatively injected basis broadband noise projection matrix into the optimized Laplace matrix acts as a high-pass digital filter barrier for the subsequent spectral clustering analysis process, blocking the malicious interference of no-load operation noise on the core feature value decomposition process from the underlying logic of matrix algebra operations. This interweaving and fusion of a series of deeply customized algorithms enables the early warning system provided by this invention to achieve highly accurate unsupervised adaptive stripping and state identification of early, subtle faults in complex working conditions of bridge erecting machines without any manual intervention in threshold setting or the preparation of massive amounts of labeled prior fault samples. This completely overcomes the common technical barrier in the industry where subtle faults in engineering machinery are easily missed or frequently falsely reported under noisy and complex working conditions.

[0043] In the preferred scheme, the specific calculation steps and algorithm implementation process for dynamically setting the early warning threshold based on the fault mechanism and triggering multi-level early warnings in step S4 rely on the damage accumulation calculation engine, real-time operating condition compensation module, and decision-making executor deployed on the control terminal of the early warning system. The steps include: By extracting the transient stress time series of the key load-bearing structure of the bridge erecting machine using a damage accumulation calculation engine, and combining this with the nonlinear fatigue damage mechanism under high-intensity operation of lifting machinery, the basic health threshold is derived. The calculation formula is: ; in, This represents the initial fatigue attenuation coefficient of the bridge erecting machine's structural components. For transient dynamic stress, The yield strength of the material. A nonlinear fatigue index related to material density and microscopic defects. The activation energy for the propagation of microcracks in the structure. Let be the ideal gas constant. The absolute temperature of the transient environment; By collecting data on the current lifting weight of the bridge erecting machine, the operating frequency of the main beam, and wind load disturbance variables through a real-time working condition compensation module, a data-driven dynamic working condition compensation factor is constructed. The derivation formula is as follows: ; in, This represents the nonlinear coupling penalty coefficient between the load and frequency. This is to determine the real-time lifting weight of the bridge erecting machine. Main beam operating frequency, This represents the weighting of aerodynamic disturbances caused by wind loads. For real-time wind speed, The maximum wind speed for the safe operation of the bridge erecting machine; The basic health threshold is adjusted through the real-time working condition compensation module. Compensation factor for dynamic operating conditions Feature mapping is performed, and a historical decay memory term is introduced in the time dimension to calculate and update the dynamic warning threshold at the current time point. The derived calculation formula is as follows: ; in, Forgetting factors in historical state memory, The total length of the historical early warning threshold time window extracted. For the first Warning threshold data for each historical period; The decision executor receives the weak fault membership index output by the clustering model. The indicators are input into the false alarm prevention judgment function based on a sliding time window, and the trigger condition formula is as follows: ; in, This represents the total step size of the sliding decision time window. For time steps The corresponding weak fault membership index, For time steps The corresponding dynamic early warning threshold, For sign determination function, The confidence level triggers the bottom line; When the triggering condition of the false alarm prevention judgment function is met, the decision executor extracts the nonlinear gradient difference of the current weak fault membership index exceeding the dynamic early warning threshold, matches it with the system's pre-set multi-level alarm rule base, and calls the communication middleware to push the corresponding level of alarm signal, protection instruction or interlock instruction to the terminal.

[0044] The detailed calculation process and mathematical logic of the specific implementation method for dynamically setting the early warning threshold and triggering multi-level early warnings in step S4, based on the fault mechanism, are explained below.

[0045] The damage accumulation calculation engine deployed at the control terminal of the early warning system first extracts the transient stress time series of the key load-bearing structure of the bridge erecting machine. Considering the complex stress alternation that occurs under high-intensity construction operations, the system derives the foundation health threshold by combining nonlinear fatigue damage mechanisms. The foundation health threshold is denoted as... The derived calculation formula is as follows: ; In this formula, The initial fatigue attenuation coefficient of the bridge erecting machine's structural components represents the initial health baseline level of the equipment after it leaves the factory or undergoes major repairs. t' represents the current total running time; t' is a continuous integral time variable. The transient dynamic stress borne by the bridge erecting machine at time t'; Yield strength of materials used in bridge erecting machine manufacturing; The nonlinear fatigue index is related to the material's density and micro-defects, and is used to characterize the nonlinear acceleration characteristics of damage accumulation in materials under alternating stress. The activation energy for the propagation of microcracks in the structure; It is the ideal gas constant; Let t' be the transient ambient absolute temperature. The product of the integral and exponential terms in the formula performs a physical-level deep coupling calculation of the cumulative law of mechanical fatigue and the thermodynamic material degradation effect caused by ambient temperature, thereby establishing a safety baseline that gradually deteriorates over time.

[0046] Subsequently, the real-time load compensation module collects real-time data on the current lifting weight of the bridge erecting machine, the operating frequency of the main beam, and wind load disturbance variables, thereby constructing a fully data-driven dynamic load compensation factor. The dynamic load compensation factor is denoted as... The derived calculation formula is as follows: ; In this formula, This is the nonlinear coupling penalty coefficient between the load and the frequency; The real-time lifting weight of the bridge erecting machine; The operating frequency of the main beam; The weighting of wind load aerodynamic disturbance; Real-time wind speed; The formula defines the maximum wind speed at which the bridge erecting machine can operate safely. It uses the natural logarithm function to nonlinearly smooth and suppress high-frequency heavy-load coupled conditions, while employing the square root function to evaluate the aerodynamic disturbance limit of the wind load. This accurately quantifies the degree to which external transient adverse conditions weaken the safety margin of the bridge erecting machine.

[0047] The real-time operating condition compensation module further maps the basic health threshold to the dynamic operating condition compensation factor, and creatively introduces a historical decay memory term in the time dimension to calculate and update the dynamic warning threshold at the current time point in real time. The dynamic warning threshold is denoted as... The derived calculation formula is as follows: ; In this formula, This is a historical state memory forgetting factor, used to adjust the system's dependence weight on historical data; This refers to the total length of the historical early warning threshold time window extracted by the system. For the sequence step index within the historical time window; For the first The warning threshold data is stored for each historical period. The second term on the right side of the formula introduces an inverse proportional time weighting mechanism, assigning a higher reference weight to historical thresholds closer to the current time. This ensures that the dynamic warning threshold has both the ability to respond quickly to sudden operating conditions and the memory inertia of long-term hidden degradation trends of the equipment.

[0048] The decision executor receives the weak fault membership index output by the clustering model and inputs it into the false alarm prevention decision function based on a sliding time window for logical verification. The weak fault membership index is defined as follows: The trigger condition formula for the false alarm prevention function is: ; In this formula, This represents the total step size of the sliding decision time window; This is the index of the current time step within the sliding time window; For time steps The corresponding weak fault membership index; For time steps Synchronize the corresponding dynamic early warning threshold; This is a sign determination function. Its operating logic is to output the value one when the internal value is greater than or equal to zero, and otherwise output the value zero. This sets a confidence threshold for the system. This threshold strictly requires that within a continuous time window, the proportion of fault indicators exceeding the dynamic threshold must reach the confidence threshold before it can be ultimately determined as a valid potential hazard.

[0049] When the triggering conditions of the false alarm prevention function are fully met, the decision executor will further extract the nonlinear gradient difference between the current weak fault membership index and the dynamic early warning threshold. The system will then accurately match this difference with a pre-set multi-level alarm rule base. Based on the severity of the matching result, the system will call the communication middleware to push the corresponding level of audible and visual alarm signal, bridge erecting machine speed reduction protection command, or emergency stop interlock command to the operation terminal or remote control center.

[0050] Implementing the aforementioned dynamic early warning threshold setting and multi-level early warning triggering steps has extremely significant technical benefits. Traditional bridge erecting machine monitoring systems generally use fixed values ​​as alarm thresholds. When the equipment faces full-load lifting or sudden gusts of wind, the surge in transient stress can easily break through the fixed threshold, resulting in a large number of invalid false alarms. Furthermore, when the equipment is in a state of long-term fatigue deterioration, the fixed threshold becomes too insensitive and easily leads to missed alarms of fatal faults. This application completely breaks through the technical bottleneck of rigid thresholds. By integrating nonlinear fatigue damage mechanisms with external transient operating condition variables, a dynamic baseline is established that can adaptively and elastically adjust with the performance degradation of the bridge erecting machine itself and harsh external environments. Combined with a sliding time window sign verification mechanism, it not only perfectly filters out transient false alarms caused by high-frequency random impacts at the underlying mathematical logic level, but also ensures accurate interception of early, subtle hidden dangers in real equipment. This solution greatly improves the anti-interference robustness and alarm confidence of the early warning system, providing irreplaceable and solid technical support for the intelligent operation and maintenance and proactive safety protection of large-scale engineering machinery.

[0051] Example 3 The actual deployment steps of this method in a computer system include: Environment setup steps: Configure the Linux operating system on the industrial control computer or NAS hardware device, and deploy the Docker containerized environment; Software installation steps: Install the Python runtime environment and algorithm dependency libraries including NumPy, SciPy, Scikit-learn, and PyTorch in a Docker container; Module integration steps: Write and deploy the data acquisition driver module, signal processing script, clustering prediction engine module, and Web visualization dashboard module; API call steps: Receive raw JSON data uploaded by the sensor via the REST API interface, and push the processed warning results to the remote monitoring terminal.

[0052] In the preferred scheme, the multi-level response steps after the early warning is triggered include: Level 1 warning: When the fault indicator is in the first threshold range, the system sends an SMS reminder to the maintenance personnel and records the location of the fault. Level 2 warning: When the fault indicator is in the second threshold range, the system will issue an alarm through the sound and light alarm and flash the prompt on the visual display board, while also suggesting to reduce the operating speed of the bridge erecting machine; Level 3 warning: When the fault indicator exceeds the maximum safety limit, the system automatically sends a shutdown command to the PLC controller and generates a fault analysis report.

[0053] The actual deployment process of the system within the computer system is the foundational engineering for ensuring the stable operation of the early warning algorithm. During the environment setup phase, an industrial control computer or network-attached storage hardware device is selected as the physical host, and a highly stable and real-time Linux operating system is configured on it. Based on this operating system, a Docker containerized runtime environment is deployed. Containerization technology enables the standardized packaging of complex early warning algorithms and their underlying environments, achieving complete decoupling between the application and the underlying infrastructure. During the software installation phase, the system centrally installs the Python runtime environment and related core algorithm dependency libraries within the Docker container. Specifically, the system installs the NumPy library to support fast parallel computation of underlying high-dimensional matrices, the SciPy library to perform efficient low-level processing tasks such as digital signal filtering and signal decomposition, the Scikit-learn library to call optimized machine learning manifold dimensionality reduction and clustering algorithm interfaces, and the PyTorch library to leverage its powerful tensor computation engine to accelerate the iterative convergence process of complex similarity matrices.

[0054] After completing the underlying software environment configuration, the system enters the module integration and interface call phase. The system independently develops and deploys a data acquisition driver module, a signal processing script, a clustering prediction engine module, and a Web visualization dashboard module. These modules interact with each other with low latency via an internal high-speed bus and shared memory mechanism. To achieve seamless integration with external hardware devices, the system develops and opens a standard expressive state transition application programming interface (API). Various sensors deployed on-site by the bridge erecting machine convert the analog signals they collect into standardized JSON format raw data, which is then frequently uploaded to the server via this interface. The JSON format offers advantages such as lightweight design, cross-platform compatibility, and ease of parsing, significantly reducing network bandwidth consumption. After completing data processing and state identification, the clustering prediction engine module also pushes the final warning results, fault membership indicators, and dynamic threshold curves to the remote monitoring terminal in real time via a standard network protocol.

[0055] To address the multi-level response mechanism triggered by the early warning, the system employs a tiered safety defense strategy. When the fault indicator output by the decision-making executor falls within the first threshold range, the system determines that the bridge erecting machine exhibits very early-stage minor wear or non-fatal fatigue. At this point, the system triggers a Level 1 warning, automatically pushing a notification containing an anomaly code to the designated maintenance personnel's mobile terminal via an external communication gateway. The system also permanently records the physical coordinates of the fault in its database, guiding maintenance personnel to focus on this area during the next routine maintenance shutdown. When the fault indicator continues to climb and enters the second threshold range, the system determines that the deterioration trend of critical components of the bridge erecting machine has become apparent and may affect the safety of lifting operations in the short term. The system then triggers a Level 2 warning, directly driving the audible and visual alarm installed in the bridge erecting machine's control room via the fieldbus to issue a high-frequency alarm. The system also prominently displays the abnormal component information in a flashing red highlight on the web visualization dashboard. Simultaneously, the system will forcefully advise on-site operators in a pop-up window to immediately reduce the main beam running speed and the winch lifting speed of the bridge erecting machine to delay further deterioration of the fault.

[0056] When fault indicators surge and exceed the system's maximum safety limits, the system determines that the equipment faces an imminent catastrophic risk of structural fracture or mechanical jamming. The system immediately triggers the highest-level (Level 3) warning. The decision-making actuator bypasses all manual verification steps and directly sends a high-priority emergency shutdown interlock command via the industrial Ethernet protocol to the programmable logic controller (PLC) at the bottom layer of the bridge erecting machine, forcibly cutting off the power supply to the core drive motor. After the equipment safely shuts down, the system automatically summarizes all multi-source characteristic data before and after the incident, generating a structured fault analysis report for incident review and subsequent damage assessment.

[0057] Implementing the aforementioned hardware and software deployment and multi-level response mechanism yields significant technical benefits. On one hand, the containerized microservice deployment architecture completely resolves the engineering pain points of traditional industrial monitoring software, such as difficulties in porting it across different industrial control computers and conflicts in environmental dependencies, endowing the system with strong cross-platform rapid delivery capabilities and agile version iteration. On the other hand, the tiered multi-level early warning mechanism precisely matches the physical damage evolution law of large lifting equipment from quantitative to qualitative changes, effectively avoiding the huge economic losses caused by frequent false shutdowns due to one-size-fits-all alarms. Through early flexible information intervention, mid-term audible and visual speed reduction warnings, and later rigid power-off control, the system, while maximizing the absolute safety of the bridge erecting machine's structure and the lives of on-site personnel, greatly improves the continuous and efficient operation capability and intelligent operation and maintenance management level of large engineering machinery, completely filling the last gap in the practical application of traditional monitoring solutions.

[0058] Example 4 A bridge erecting machine early fault warning system includes: The multi-source data acquisition and enhancement module has its signal output end connected to the feature fusion and dimensionality reduction module, which is used to transmit the reconstructed intrinsic mode function signal to the feature fusion and dimensionality reduction module. The feature fusion and dimensionality reduction module has its signal input end connected to the multi-source data acquisition and enhancement module, and its signal output end connected to the fault state identification module, which is used to transmit the mapped low-dimensional sensitive feature vector to the fault state identification module. The fault status identification module has a signal input end connected to the feature fusion and dimensionality reduction module, and a signal output end connected to the dynamic threshold early warning module. It is used to transmit the weak fault membership index generated by cluster analysis to the dynamic threshold early warning module. The dynamic threshold early warning module has its signal receiving end connected to the fault status identification module and the external operating condition sensor, respectively. It is used to receive weak fault membership indicators and real-time operating condition parameters, and output early warning control commands after logical judgment.

[0059] This paper details the overall architecture and collaborative operation mechanism of the early fault warning system for bridge erecting machines. The system's first end is configured with a multi-source data acquisition and enhancement module. This module acts as the data input barrier and purification hub for the entire warning system. Its signal output is strictly unidirectionally connected to the feature fusion and dimensionality reduction module via a physical link or software communication bus. This module receives the raw, complex signals collected by the underlying sensor array. After internal modal decomposition and noise reduction filtering, it continuously transmits the highly purified reconstructed intrinsic mode function signals to the downstream feature fusion and dimensionality reduction module without distortion through a high-speed data interface or shared memory area. This front-end deep purification and direct transmission architecture design has extremely significant technical benefits. It completely cuts off the backward propagation path of broadband mechanical noise and environmental electromagnetic interference from the bridge erecting machine at the source of the system's physical data flow, avoiding invalid noise occupying the system's internal bus bandwidth, thus providing a high-quality data source with an extremely high signal-to-noise ratio for the precise feature extraction of subsequent modules.

[0060] The feature fusion and dimensionality reduction module closely follows the front-end data flow. Its signal input is connected at a high frequency synchronously with the output of the multi-source data acquisition and enhancement module, while its signal output is directly connected to the input bus of the fault state identification module. Upon receiving the intrinsic mode function reconstruction signal, this module performs parallel multi-dimensional matrix splicing operations on time-domain features, frequency-domain features, and time-frequency-domain features in its independent processing unit. It then immediately invokes the local preserving projection dimensionality reduction engine to perform spatial compression. After linear mapping, the module accurately pushes a highly concise, low-dimensional sensitive feature vector that accurately characterizes the equipment degradation trend to the fault state identification module. The beneficial effect of implementing this module connection and data flow architecture is that by forcibly performing physical dimensionality reduction on the complex high-dimensional feature matrix at the system's intermediate layer, the data throughput computational pressure of the subsequent core clustering identification module is completely offloaded. This not only significantly improves the data flow efficiency and real-time response rate of the entire early warning system but also eliminates the communication congestion risks caused by high-dimensional spatial feature redundancy at the system architecture level.

[0061] The fault status identification module, as the core intelligent decision-making hub of the system, has its signal input closely connected to the feature fusion and dimensionality reduction module, while its signal output establishes a highly reliable data link with the backend dynamic threshold early warning module. Upon receiving a low-dimensional sensitive feature vector, this module utilizes its internally configured matrix acceleration computing hardware unit to quickly construct an adaptive similarity map and solve for the optimized Laplace matrix. Through unsupervised density space clustering analysis, it accurately quantifies the degree of degradation of the current bridge erecting machine's operating status from the normal baseline. Subsequently, the module transmits the quantified weak fault membership index as a standardized digital signal to the dynamic threshold early warning module at the end of the system. The advantage of this node design lies in completely encapsulating the extremely complex nonlinear mathematical space clustering process within an independent hardware or software sandbox module, exposing only standardized one-dimensional index output to the subsequent layers. This highly decoupled modular data output mechanism greatly enhances the system's fault tolerance and maintainability. Even if the field equipment model changes, only the scaling factor needs to be adaptively adjusted within the identification module, without requiring any modification to the entire system's data link architecture.

[0062] The dynamic threshold early warning module is located at the very end of the control data flow of the early warning system, employing a unique dual-source signal receiving architecture. On one hand, the module's signal receiver connects to the fault status identification module via an internal system bus to receive weak fault membership indicators. On the other hand, it establishes a real-time connection with external condition sensors deployed on the bridge erecting machine via an external industrial communication interface to synchronously receive real-time condition parameters such as lifting weight, main beam operating frequency, and wind speed. The module's built-in logic actuator performs a high-frequency comparison between the received weak fault membership indicators and the dynamically calculated dynamic early warning threshold based on real-time condition parameters. Upon meeting the false alarm triggering conditions, it outputs multi-level early warning control commands to the external hardware actuator after logical judgment. This closed-loop design at the end of the system offers significant advantages, completely breaking away from the rigid mode of traditional monitoring systems that rely solely on internal vibration data for blind judgment. By forcibly introducing external real working environment parameters as judgment compensation factors at the physical connection level, the final output early warning control command not only includes a keen insight into micro mechanical fatigue, but also takes into account the macroscopic resistance to harsh external working environments. This fundamentally achieves zero missed alarms and extremely low false alarms for the multi-level early warning signals of the bridge erecting machine, giving the entire early warning system extremely strong adaptability to on-site working conditions and industrial-grade continuous operation reliability.

[0063] A computer device includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the above-mentioned early fault warning method for bridge erecting machines.

[0064] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described early fault warning method for bridge erecting machines.

[0065] The internal logic structure and specific execution mechanism of the computer equipment and computer-readable storage medium used to implement the early fault warning method for bridge erecting machines are described in detail.

[0066] The computer device, in its physical architecture, includes a memory and a processor. The memory, serving as the physical carrier of system instructions and massive amounts of state data, specifically employs high-speed random access memory to ensure high-frequency caching requirements during clustering algorithm operation, while also being equipped with non-volatile memory for persistently storing historical health covariance matrices and dynamic early warning threshold time series. The memory fully maps and encapsulates all the aforementioned logic control code instruction sets, from multi-source signal acquisition, modality enhancement, manifold dimensionality reduction to nuclear spectrum clustering identification and dynamic early warning; these instruction sets constitute the computer program that implements the early warning function. The processor, as the device's computation and scheduling core, interacts with the memory continuously via an internal high-speed system bus. When the early warning system is activated, the processor actively extracts and executes the computer program instructions from the memory according to a set periodic frequency. During execution, the processor uses its built-in floating-point arithmetic unit to perform eigenvalue decomposition of the complex Laplace matrix and calls the logic control unit to perform symbol comparison for sliding time window false alarm prevention, ultimately outputting precise multi-level early warning control signals to external actuators. The beneficial effect of implementing this computer equipment is that by deeply embedding the abstract early warning algorithm logic into a specific hardware computing platform, the entire early warning method is endowed with industrial-grade computing capabilities for independent operation, offline computation, and real-time response. This ensures that the method can still stably and efficiently complete the continuous monitoring and judgment of early and minor faults in the complex electromagnetic interference environment at the bridge erecting machine site.

[0067] The computer-readable storage medium, as an independent data carrier, can take physical forms including, but is not limited to, portable solid-state drives, optical discs, and distributed storage nodes deployed in cloud server clusters. This storage medium contains complete computer program code that executes all core steps of the early fault warning method for bridge erecting machines. When this storage medium is legally connected to any compatible computing terminal via a physical interface or network protocol and read and executed by its built-in processor, the complete early warning monitoring process can be instantly replicated and initiated within the target terminal. The beneficial effects of implementing this computer-readable storage medium are that it achieves complete software-based and highly portable implementation of complex industrial early warning algorithms. This design not only greatly reduces the distribution cost and cross-platform portability of core early warning algorithms but also enables the system to be deployed on a large scale in vehicle-mounted control units or remote monitoring centers of different models of bridge erecting machines at extremely low marginal costs. Simultaneously, it provides a standardized and regulated software distribution carrier for subsequent online hot updates of algorithm logic, remote cloud upgrades, and periodic synchronization of the fault sample library, significantly improving the commercial promotion value and subsequent maintenance efficiency of the engineering machinery safety monitoring system.

[0068] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A method for early fault warning of a bridge erecting machine, characterized by: Includes the following steps: S1. Real-time acquisition of multi-source operating data of key parts of the bridge erecting machine through sensor array, and decomposition and enhancement processing of the acquired signals using mode decomposition algorithm to extract weak fault components hidden in strong noise background. S2. Extract physical quantities reflecting fault characteristics from the enhanced component signals in multiple dimensions, and use feature fusion and dimensionality reduction techniques to compress the high-dimensional feature space into a low-dimensional sensitive feature vector that can accurately characterize the fault state. S3. Input the low-dimensional sensitive feature vector into the pre-constructed adaptive nuclear spectrum clustering model, calculate the similarity between samples and perform cluster analysis to realize the automatic identification and classification of early weak fault states of the bridge erecting machine. S4. Based on the structural failure mechanism model of the bridge erecting machine, the warning threshold is dynamically calculated and updated in real time according to the current real-time working conditions. When the identified fault indicators exceed the warning threshold, a multi-level warning mechanism is triggered and the failure evolution trend is predicted.

2. The method for early fault warning of a bridge erecting machine according to claim 1, characterized in that: Step S1, which involves using a mode decomposition algorithm to decompose and enhance the signal, includes the following steps: First, the variational mode decomposition algorithm is used to decompose the preprocessed vibration signal into several eigenmode function components with specific center frequencies; Calculate the correlation coefficient and kurtosis value of each intrinsic mode function component, and screen out the effective components that contain the main fault characteristic information; The selected effective components are reconstructed, and broadband random noise is removed, thereby enhancing the weak fault characteristic signal.

3. The method for early fault warning of a bridge erecting machine according to claim 1, characterized in that: Step S2 involves extracting physical quantities and performing dimensionality reduction, including: The statistical characteristics of the reconstructed signal are calculated from three dimensions: time domain, frequency domain, and time-frequency domain. The time domain characteristics include the effective value, peak factor, and impulse factor; the frequency domain characteristics include the centroid frequency and root mean square frequency; and the time-frequency domain characteristics are composed of the energy values ​​of the wavelet packet decomposition coefficients. The above features are combined to construct an initial high-dimensional fault feature matrix; A local preservation projection algorithm based on manifold learning is used to linearly map the high-dimensional fault feature matrix, which reduces feature redundancy while preserving local structural information of the data and obtains low-dimensional sensitive feature vectors.

4. The method for early fault warning of a bridge erecting machine according to claim 1, characterized in that: The construction and recognition process of the adaptive nuclear spectrum clustering model in step S3 relies on the feature processing unit, matrix calculation acceleration module, and clustering decision maker built into the early warning system. The steps include: The feature processing unit receives a set of low-dimensional sensitive fault feature vectors after dimensionality reduction, extracts the local spatial distribution density for each feature vector in the set, and derives a data-driven adaptive scaling factor based on the historical operation distribution of the bridge erecting machine. Specifically, for any given feature vector... eigenvectors Its adaptive scaling factor The derivation formula is as follows: ； in, This is the working condition compensation coefficient that varies with the lifting load of the bridge erecting machine. To prevent infinitely large smoothing constants, The total number of samples for the extracted feature vector. The trace of the feature covariance matrix of the historical health status data of the bridge erecting machine; The matrix computation acceleration module constructs a hybrid kernel similarity matrix containing a Markov structure penalty mechanism based on the adaptive scaling factor of each eigenvector. This matrix is ​​then used to calculate the similarity between any two eigenvectors. and Similarity elements between Establish a global similarity graph, with similarity elements. The derivation and calculation formula is as follows: ； in, It is the minimum constant. The dynamic weighting parameters are set based on the stress transfer gradient of the key structure of the bridge erecting machine. The second term on the right side of the equation is the structural correlation projection represented by the inverse matrix of the covariance matrix. The matrix computation acceleration module calculates the diagonal matrix containing node connectivity based on the global similarity graph. Furthermore, the optimized Laplace matrix incorporating a mechanical vibration noise regularization term is derived. The calculation formula is: ； in, It is the identity matrix. It is an angle matrix whose diagonal elements are similarity matrices. The sum of the elements in the corresponding row. This is the noise suppression coefficient. The base broadband noise projection matrix is ​​extracted in advance from the no-load operation condition of the bridge erecting machine; The matrix computation acceleration module employs the implicit restart Arnoldi iterative algorithm to optimize the Laplace matrix. Perform eigenvalue decomposition, solve and extract the frontier values. The eigenvectors corresponding to the non-zero minimum eigenvalues ​​are concatenated column by column and normalized by row to construct a low-dimensional orthogonal indicator matrix. The clustering decision unit receives the low-dimensional orthogonal indicator matrix and inputs its row vectors as smoothed data points after noise reduction into the density-based spatial clustering algorithm module. Based on the density connectivity and distribution boundary of the clusters, the current state features are automatically weighted and labeled. The stable main clusters are determined to be in a normal state, and the discrete micro-clusters that deviate from the main clusters and exhibit nonlinear decay in local density are determined to be in an early weak fault state. Finally, the weak fault identification result is output.

5. The method for early fault warning of a bridge erecting machine according to claim 1, characterized in that: Step S4, which dynamically sets the early warning threshold based on the fault mechanism and triggers multi-level early warnings, involves specific calculation steps and algorithm implementation processes that rely on the damage accumulation calculation engine, real-time operating condition compensation module, and decision-making executor deployed on the early warning system control terminal. The steps include: The transient stress time series of the key load-bearing structure of the bridge erecting machine is extracted by a damage accumulation calculation engine. Combined with the nonlinear fatigue damage mechanism under high-intensity operation of lifting machinery, the basic health threshold is derived. The calculation formula is: ； in, The initial fatigue attenuation coefficient of the bridge erecting machine structural components is given. For transient dynamic stress, The yield strength of the material. The nonlinear fatigue index is related to the material's density and microscopic defects. The activation energy for the propagation of microcracks in the structure. Let be the ideal gas constant. The absolute temperature of the transient environment; By collecting data on the current lifting weight of the bridge erecting machine, the operating frequency of the main beam, and wind load disturbance variables through a real-time working condition compensation module, a data-driven dynamic working condition compensation factor is constructed. The derivation formula is as follows: ； in, This represents the nonlinear coupling penalty coefficient between the load and frequency. This is to determine the real-time lifting weight of the bridge erecting machine. Main beam operating frequency, This represents the weighting of aerodynamic disturbances caused by wind loads. For real-time wind speed, The maximum wind speed for the safe operation of the bridge erecting machine; The basic health threshold is adjusted through the real-time working condition compensation module. Compensation factor for dynamic operating conditions Feature mapping is performed, and a historical decay memory term is introduced in the time dimension to calculate and update the dynamic warning threshold at the current time point. The derived calculation formula is as follows: ； in, Forgetting factors in historical state memory, The total length of the historical early warning threshold time window extracted. For the first Warning threshold data for each historical period; The decision executor receives the weak fault membership index output by the clustering model. The indicators are input into the false alarm prevention judgment function based on a sliding time window, and the trigger condition formula is as follows: ； in, This represents the total step size of the sliding decision time window. For time steps The corresponding weak fault membership index, For time steps The corresponding dynamic early warning threshold, For sign determination function, The confidence level triggers the bottom line; When the triggering condition of the false alarm prevention judgment function is met, the decision executor extracts the nonlinear gradient difference of the current weak fault membership index exceeding the dynamic early warning threshold, matches it with the system's pre-set multi-level alarm rule base, and calls the communication middleware to push the corresponding level of alarm signal, protection instruction or interlock instruction to the terminal.

6. A method for early fault warning of a bridge erecting machine according to any one of claims 1 to 5, characterized in that: The actual deployment steps of this method in a computer system include: Environment setup steps: Configure the Linux operating system on the industrial control computer or NAS hardware device, and deploy the Docker containerized environment; Software installation steps: Install the Python runtime environment and algorithm dependency libraries including NumPy, SciPy, Scikit-learn, and PyTorch in a Docker container; Module integration steps: Write and deploy the data acquisition driver module, signal processing script, clustering prediction engine module, and Web visualization dashboard module; API call steps: Receive raw JSON data uploaded by the sensor via the REST API interface, and push the processed warning results to the remote monitoring terminal.

7. The method for early fault warning of a bridge erecting machine according to claim 6, characterized in that: The multi-level response steps after an early warning is triggered include: Level 1 warning: When the fault indicator is in the first threshold range, the system sends an SMS reminder to the maintenance personnel and records the location of the fault. Level 2 warning: When the fault indicator is in the second threshold range, the system will issue an alarm through the sound and light alarm and flash the prompt on the visual display board, while also suggesting to reduce the operating speed of the bridge erecting machine; Level 3 warning: When the fault indicator exceeds the maximum safety limit, the system automatically sends a shutdown command to the PLC controller and generates a fault analysis report.

8. An early fault warning system for bridge erecting machines, characterized in that: include: The multi-source data acquisition and enhancement module has its signal output end connected to the feature fusion and dimensionality reduction module, which is used to transmit the reconstructed intrinsic mode function signal to the feature fusion and dimensionality reduction module. The feature fusion and dimensionality reduction module has its signal input end connected to the multi-source data acquisition and enhancement module, and its signal output end connected to the fault state identification module, which is used to transmit the mapped low-dimensional sensitive feature vector to the fault state identification module. The fault status identification module has a signal input end connected to the feature fusion and dimensionality reduction module, and a signal output end connected to the dynamic threshold early warning module. It is used to transmit the weak fault membership index generated by cluster analysis to the dynamic threshold early warning module. The dynamic threshold early warning module has its signal receiving end connected to the fault status identification module and the external operating condition sensor, respectively. It is used to receive weak fault membership indicators and real-time operating condition parameters, and output early warning control commands after logical judgment.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: The processor executes the steps of any one of claims 1 to 7 when executing a computer program.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When a computer program is executed by a processor, it implements the steps of any one of claims 1 to 7.

Images