Multi-signal analysis and early warning system for vibration and noise of building electromechanical equipment

Abstract

This invention relates to the field of building equipment monitoring technology, specifically disclosing a multi-signal analysis and early warning system for vibration and noise of building electromechanical equipment. The system includes a multi-modal signal acquisition array, a signal preprocessing and synchronization module, a multi-source signal blind separation and reconstruction module, an equipment feature fingerprint database, a fault mode recognition and localization engine, and an early warning decision and visualization module. By extracting independent source signals from the equipment through blind source separation technology, and combining feature fingerprint comparison with vibration-acoustic fusion localization, it achieves accurate identification, mode classification, and three-dimensional spatial localization of equipment faults. Furthermore, it provides graded early warnings based on the degree of deviation, thereby improving the sensitivity and operational efficiency of building electromechanical equipment fault monitoring.

Description

Technical Field

[0001] This invention belongs to the field of building equipment monitoring technology, specifically relating to a multi-signal analysis and early warning system for vibration and noise of building electromechanical equipment. Background Technology

[0002] In the field of operation and maintenance management of large-scale building facilities, the stable operation of building electromechanical equipment is essential to ensure the building's functionality and comfort. Condition monitoring and fault early warning for this equipment are technical means to preventative maintenance, reduce operation and maintenance costs, and avoid unplanned downtime. Vibration and noise signals are the most direct and effective physical quantities reflecting the mechanical condition of the equipment.

[0003] The equipment condition monitoring and fault early warning technology based on vibration and noise analysis aims to collect vibration and noise signals generated during equipment operation, and use signal processing and data analysis methods to identify abnormal equipment conditions and provide early warnings of potential faults, thereby achieving intelligent management of equipment health status.

[0004] Existing technologies typically employ a single sensor to monitor the target device or monitor multiple devices independently. In complex building electromechanical environments, multiple devices are often densely arranged and operate simultaneously. The vibration and noise signals they generate superimpose and couple during propagation, forming a complex mixed signal field. Existing analysis methods struggle to effectively separate the characteristic components belonging to specific devices from this mixed signal, resulting in the submergence of fault characteristics and an inability to accurately identify the source of anomalies. Due to the complexity of signal propagation paths and mutual interference between devices, methods based on single signal sources or simple threshold judgments face difficulties in accurately locating fault sources, easily leading to false alarms or missed alarms. This makes it difficult for maintenance personnel to quickly and accurately locate faulty devices, delaying maintenance opportunities and increasing troubleshooting costs.

[0005] Therefore, how to effectively separate signals and accurately locate fault sources from mixed vibration noise from multiple devices has become a pressing technical challenge in the field of intelligent operation and maintenance of building electromechanical equipment. Summary of the Invention

[0006] This invention provides a multi-signal analysis and early warning system for vibration and noise of building electromechanical equipment, in order to solve the technical contradiction in the prior art that it is difficult to effectively separate the characteristics of specific equipment and accurately locate the fault source from the mixed vibration and noise signals of multiple equipment.

[0007] The Vibration and Noise Multi-Signal Analysis and Early Warning System for Building Mechanical and Electrical Equipment includes a multi-modal signal acquisition array, a signal preprocessing and synchronization module, a multi-source signal blind separation and reconstruction module, an equipment feature fingerprint database, a fault mode recognition and location engine, and an early warning decision and visualization module.

[0008] The multimodal signal acquisition array consists of a vibration sensor array and an acoustic sensor array deployed at monitoring points of the building's electromechanical equipment cluster. The vibration sensor array uses triaxial accelerometers, arranged in a specific spatial density grid along key vibration transmission paths on equipment foundations, pipe supports, and the building structure. The acoustic sensor array comprises multiple high-precision microphones, deployed in a specific three-dimensional spatial configuration in the equipment room and adjacent spaces to collect airborne noise field signals. All sensor nodes integrate a high-precision timing chip to ensure global time synchronization of the entire array's data acquisition, with a synchronization accuracy better than 1 microsecond.

[0009] The signal preprocessing and synchronization module receives and processes the raw signal stream from the multimodal signal acquisition array. First, it performs bandpass filtering on the raw vibration and acoustic signals to remove extremely low-frequency structural vibrations and high-frequency electromagnetic noise unrelated to equipment operation. Further, the module performs signal amplitude normalization and timestamp alignment, unifying the sampled data from all channels to the same time base and amplitude reference system, generating a strictly time-synchronized multi-channel hybrid observation signal matrix.

[0010] The multi-source signal blind separation and reconstruction module is the core processing unit of the system, used to separate and reconstruct the source signals corresponding to each independent electromechanical device from the time-synchronized multi-channel mixed observation signal matrix. This module is based on an improved joint diagonalization blind source separation algorithm. Specifically, the algorithm first calculates the set of cross-correlation matrices of the multi-channel mixed observation signals at multiple different time delays. Further, it searches for an optimal separation matrix that can simultaneously approximate the diagonalization of the aforementioned cross-correlation matrix set. Solving for this optimal separation matrix is ​​transformed into a constrained optimization problem of minimizing the weighted sum of the off-diagonal elements of all cross-correlation matrices, and is solved using an iterative Jacobian rotation method. Finally, applying the obtained optimal separation matrix to the original mixed observation signal matrix outputs a series of statistically independent estimated source signals, each corresponding to a physically independent vibration or acoustic source, i.e., a specific electromechanical device.

[0011] The equipment feature fingerprint database is a dynamically updated database used to store and match the baseline feature fingerprints of various electromechanical devices. The feature fingerprint of each device is established through the following process: Under the known healthy operating conditions of the device, the source signals of the device are separated by the multi-source signal blind separation and reconstruction module; features are extracted from the source signals, including but not limited to time-domain statistical features, frequency-domain spectral features, time-frequency domain wavelet packet energy features, and higher-order statistical features; the extracted feature set is associated with and stored with the device's unique identifier to form the device's baseline feature fingerprint. The equipment feature fingerprint database supports online incremental learning; when the device undergoes maintenance or component replacement, the system can update the original feature fingerprint based on the new health status data.

[0012] The fault mode recognition and localization engine is used to assess the state and locate faults in the isolated source signals of various devices. This engine first retrieves the baseline feature fingerprint of the corresponding device from the device feature fingerprint database. Further, it performs the same feature extraction operation on the currently isolated device source signals to obtain real-time feature vectors. The engine incorporates a multi-level fault diagnosis model. The first level is feature deviation calculation, which calculates the Mahalanobis distance between the real-time feature vector and the baseline feature fingerprint to obtain a comprehensive state deviation index. The second level is fault mode classification; when the comprehensive state deviation index exceeds a first preset threshold, a support vector machine-based multi-classifier is triggered. This classifier, trained based on historical fault cases, maps abnormal features to specific fault modes, such as bearing wear, rotor imbalance, and blade breakage. The third level is fault source spatial localization, which integrates data from vibration sensor arrays and acoustic sensor arrays. For vibration signals, a wave direction-of-arrival (WDA) estimation technique based on time difference of arrival (TDOA) is used. By calculating the time difference of arrival of the vibration wavefront at different vibration sensors, the propagation direction of the vibration source is inverted. For acoustic signals, a sound source localization algorithm based on controllable response power is employed to perform a sound energy focusing search in three-dimensional space to determine the spatial coordinates corresponding to the maximum sound energy. Finally, the vibration source direction estimate and the sound source spatial coordinates are fused, and through spatial geometric intersection calculations, the accurate coordinate estimate of the faulty equipment in three-dimensional space and its location confidence are output.

[0013] The early warning decision and visualization module generates tiered early warning commands and provides a human-machine interface based on the output of the fault mode recognition and localization engine. This module presets multiple early warning level thresholds, including attention level, early warning level, and alarm level. The early warning decision logic is as follows: When the overall state deviation index is greater than the attention level threshold but lower than the early warning level threshold, a device status deterioration trend indication is generated. When the overall state deviation index is greater than the early warning level threshold, and the fault mode classifier outputs a clear fault type, a corresponding early warning notification is generated, along with an estimated spatial coordinate of the fault source. When the overall state deviation index is greater than the alarm level threshold, and the fault source location confidence level is higher than 95%, the highest-level alarm command is generated, and the linkage interface with the building equipment management system is automatically triggered, providing a basis for decision-making regarding equipment shutdown or load adjustment. The visualization module integrates and displays the 3D model of the building's electromechanical equipment, sensor deployment diagrams, real-time signal flow graphs, health status dashboards for each device, 3D heatmaps of fault location, and historical early warning logs.

[0014] As one embodiment of the present invention, the improved joint diagonalization algorithm in the multi-source signal blind separation and reconstruction module adaptively allocates the weight coefficients of the off-diagonal elements in the objective function based on the eigenvalues ​​of each cross-correlation matrix. Cross-correlation matrices with larger eigenvalues ​​have correspondingly larger weight coefficients in the objective function. This enhances the algorithm's ability to separate strongly correlated signal components and improves the signal-to-noise ratio for separating weak fault characteristic signals under strong background noise.

[0015] In one embodiment of the present invention, the fault source spatial localization level in the fault mode recognition and localization engine employs a Kalman filter algorithm for its data fusion process. This Kalman filter algorithm establishes a state-space model of the system using vibration wave direction-of-arrival estimation as the observation vector and sound source localization coordinates as the state vector. Through the prediction and update loop of the Kalman filter, optimal fusion and smoothing filtering are performed on the spatial coordinate estimation of the fault source, effectively suppressing random errors in the localization results of a single sensor array, thereby reducing the final localization error.

[0016] As one embodiment of the present invention, the incremental learning process of the device feature fingerprint database introduces a feature stability assessment based on Spearman's rank correlation coefficient. When the Spearman correlation coefficient between the extracted feature vector and the historical baseline feature fingerprint of new health status data collected after equipment maintenance is less than 0.8, the system determines that the equipment operating status baseline has changed, automatically overwrites the old feature fingerprint with the feature vector generated from the new data, and records the version update log.

[0017] In one embodiment of the present invention, the system adopts a cloud-edge collaborative computing architecture. A multimodal signal acquisition array and a signal preprocessing and synchronization module are deployed on edge computing nodes for real-time acquisition and preliminary processing of high-frequency raw data. The computational tasks of the multi-source signal blind separation and reconstruction module, the device feature fingerprint database, and the fault mode recognition and localization engine are deployed on a cloud server, receiving preprocessed data uploaded from edge nodes and executing complex blind separation, pattern recognition, and localization algorithms. The early warning decision and visualization module acts as a client, subscribing to processing results and raw alarm streams from both the cloud and edge nodes. This architecture ensures both the real-time performance of data processing and meets the computational resource requirements of large-scale algorithm operations.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention, by deploying a spatiotemporally synchronized multimodal sensor array and employing a blind source separation algorithm based on joint diagonalization, achieves for the first time in the field of building electromechanical operation and maintenance the effective separation and reconstruction of source signals from multiple simultaneously operating devices within a complex mixed vibration and noise field. This technology solves the problem of fault characteristic signals being submerged by environmental noise and other device signals, providing a clean, device-specific signal input for subsequent accurate diagnosis, and improving the sensitivity and specificity of fault detection.

[0019] 2. This invention innovatively constructs a fault source spatial localization engine that fuses vibration and acoustic signals. It combines the direction-of-arrival estimation of vibration signals with the two-dimensional spatial coordinate localization of acoustic signals, and performs optimal data fusion through Kalman filtering. This technology overcomes the limitations of traditional single-signal-source localization methods, which suffer from insufficient accuracy and susceptibility to interference. It can output high-confidence coordinates of faulty equipment in three-dimensional space, enabling maintenance personnel to quickly and accurately locate the fault point, greatly shortening the fault diagnosis time and cost.

[0020] 3. This invention establishes a dynamically updatable equipment feature fingerprint database and designs a multi-level intelligent diagnostic model, forming a complete analysis chain from state deviation detection to fault mode classification and spatial positioning. The system can not only identify "whether there is an anomaly," but also determine "what kind of anomaly" and "where the anomaly is," achieving intelligent operation throughout the entire process from monitoring to diagnosis to positioning. Combined with a cloud-edge collaborative architecture, the system possesses both real-time response capabilities and powerful data analysis capabilities, providing a complete and reliable technical solution for predictive maintenance and intelligent operation and maintenance of large-scale building electromechanical equipment. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the principle framework of the multi-source signal blind separation and reconstruction module in this invention; Figure 3 This is a multi-level logical flow diagram of the fault mode recognition and localization engine in this invention; Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow of data processing and task flow under the cloud-edge collaborative architecture in this invention; Figure 5 This is a schematic diagram illustrating the technical effect of the vibration and acoustic signal fusion positioning principle compared with a single method in this invention. Detailed Implementation

[0022] Example 1: Please refer to the appendix Figure 1 To be continued Figure 5 The overall technical architecture of the multi-signal analysis and early warning system for vibration and noise of building electromechanical equipment described in this invention is as follows: Figure 1 As shown, it consists of six functional units: a multimodal signal acquisition array, a signal preprocessing and synchronization module, a multi-source signal blind separation and reconstruction module, a device feature fingerprint database, a fault mode recognition and location engine, and an early warning decision and visualization module. These units interact in an orderly manner through strictly defined data interfaces and logical control flows, jointly completing the entire process from raw mixed signal acquisition to accurate fault location and hierarchical early warning.

[0023] The multimodal signal acquisition array is the sensing front end of the system, deployed in the monitoring area of ​​the building's electromechanical equipment cluster. This multimodal signal acquisition array consists of a vibration sensor array and an acoustic sensor array.

[0024] The vibration sensor array employs triaxial accelerometers, deployed in a grid pattern along key vibration transmission paths of equipment foundations, pipe supports, floor support beams, and building structures, with a spatial density of at least one node per 20 square meters. Each triaxial accelerometer has a range of ±50g, a frequency response range of 0.5 Hz to 10 kHz, and a sampling rate set at 25600 Hz to ensure complete capture of broadband vibration information generated during equipment operation.

[0025] The acoustic sensor array consists of at least eight high-precision electret microphones, deployed in a three-dimensional configuration within the equipment room and adjacent corridors, shafts, and other spaces. The microphone spacing is no less than 1.5 meters to meet the sound field resolution requirements under spatial reverberation conditions. All sensor nodes integrate a timing chip based on the IEEE 1588 precision time protocol, achieving global time synchronization of the entire array's data acquisition through a master-slave clock synchronization mechanism, with a synchronization accuracy better than 1 microsecond. Each sensor node has a built-in 16-bit analog-to-digital converter that converts analog signals into digital signal streams in real time and uploads them to edge computing nodes via Industrial Ethernet or Time-Sensitive Networking Protocol (TSN).

[0026] The signal preprocessing and synchronization module is deployed on edge computing nodes to receive and process the raw data stream from the multimodal signal acquisition array. This module first performs bandpass filtering on the raw vibration and acoustic signals from each channel. For vibration signals, the filter passband is set to 5 Hz to 8000 Hz to effectively filter out extremely low-frequency disturbances below 5 Hz caused by wind loads or foot traffic in the building structure, while suppressing high-frequency electromagnetic interference above 8000 Hz. For acoustic signals, the passband is set to 100 Hz to 6000 Hz to eliminate low-frequency ambient hum and ultrasonic interference.

[0027] The filtered signal enters the amplitude normalization submodule. This submodule dynamically adjusts the gain coefficient of the current signal based on the historical maximum amplitude of each sensor, ensuring that the output amplitude of all channels is uniformly within the [-1, 1] range. Subsequently, the timestamp alignment submodule interpolates and resamples the data from all channels based on the precise timestamps reported by each sensor, ensuring that all signals are strictly aligned on the time axis, generating a dimension of... Multi-channel hybrid observation signal matrix ,in This represents the total number of sensor channels. This represents the number of sampling points in a single processing window, typically 8192. This multi-channel mixed observation signal matrix X serves as the input reference for subsequent blind separation processing.

[0028] The multi-source signal blind separation and reconstruction module is the system's processing engine, and its working principle is shown in the attached figure. Figure 2 As shown. The goal of this multi-source signal blind separation and reconstruction module is to extract data from a mixed observation signal matrix. Separate from Statistically independent source signals , Given m original signals to be separated, where each original signal corresponds to a vibration or noise source emitted by a physically independent electromechanical device, this multi-source signal blind separation and reconstruction module employs an improved joint diagonalization blind source separation algorithm. The algorithm first calculates the mixed signal in... Different delays The set of cross-correlation matrices below ,in , This represents the mathematical expectation. In practical engineering implementation, the mathematical expectation is approximated by time averaging, i.e. Algorithm search Separation matrix of dimension This makes the transformed signal cross-correlation matrix In all The matrix should be as close to a diagonal matrix as possible. This is the index for the time sampling points. The problem is modeled as a constrained optimization problem: ; in, This represents the operation of extracting the off-diagonal elements of a matrix and setting the diagonal elements to zero. For the Frobenius norm, These are the weighting coefficients of the k-th cross-correlation matrix. According to the improved strategy of this invention, the weighting coefficients... in accordance with Maximum eigenvalue Adaptive allocation, i.e.: ; For the first One delay, For the first One delay, It is the first The maximum eigenvalue corresponding to each delay.

[0029] This design allows cross-correlation matrices with larger eigenvalues ​​(typically corresponding to strongly correlated signal components) to have higher weights in the objective function, thereby enhancing the algorithm's ability to separate dominant signals and improving the identifiability of weak fault features under strong background noise. This optimization problem is solved using an iterative Jacobian rotation method: in each iteration, a pair of row indices is selected. Calculate the optimal rotation angle Construct the Givens rotation matrix And update W←WG. Iteration continues until the rate of change of the objective function value is lower than... Or it can reach the maximum number of iterations of 500. The final output is the estimated source signal matrix. Each line represents the reconfiguration source signal for an independent device.

[0030] The equipment feature fingerprint database is a dynamically evolving database used to store multi-dimensional feature representations of various electromechanical equipment in a healthy state. The feature fingerprint establishment process for each piece of equipment is as follows: After the equipment completes installation, commissioning, or major overhaul, and confirms that it is in a stable and healthy operating state, the system initiates a baseline fingerprint acquisition process. During this period, the multi-source signal blind separation and reconstruction module continuously outputs the source signals corresponding to the equipment. Four types of features are extracted from the source signals: The first type is time-domain statistical features, including mean, variance, skewness, kurtosis, root mean square value, peak factor, and margin factor, totaling 8 indicators; the second type is frequency-domain spectral features, which extract the frequency, amplitude, and phase of the first 50 spectral lines by performing a fast Fourier transform on the source signals corresponding to the equipment, totaling 150 indicators; the third type is time-frequency domain wavelet packet energy features, which use the db4 wavelet basis to perform a 5-level decomposition of the signal and calculate the percentage of energy of each sub-band to the total energy, totaling 32 indicators; the fourth type is higher-order statistical features, including bispectral slice amplitude, trispectral diagonal slice, etc., totaling 20 indicators.

[0031] The aforementioned 210-dimensional feature vectors are bound to the device's unique identifier (such as device serial number or asset code) and stored in the device feature fingerprint database. This database supports online incremental learning: when the device undergoes maintenance or component replacement, the system automatically triggers a new health status data collection. The newly collected feature vectors are compared with the historical baseline fingerprint to calculate the Spearman-level correlation coefficient ρ. If ρ < 0.8, it is determined that the device's operating baseline has drifted, and the system automatically uses... cover The system records the version number, update time, and operator information in the log to ensure that the fingerprint database always reflects the device's current true health benchmark.

[0032] The logical flow of the fault mode identification and localization engine is shown in the attached figure. Figure 3 As shown, its functionality is divided into three levels. The first level is feature deviation calculation. The engine receives the source signals from each device from the multi-source signal blind separation and reconstruction module in real time, and performs the same feature extraction process as when building the fingerprint database on each signal to obtain real-time feature vectors. Retrieve the reference feature fingerprint of the corresponding device. Calculate the Mahalanobis distance between the two. : ; in This is the covariance matrix of the baseline feature vectors on the historical health dataset. Distance This is the comprehensive state deviation index, which considers the correlation and variance between various feature dimensions and reflects the true degree of anomaly better than Euclidean distance. The second level is fault mode classification. When Greater than the first preset threshold When the system determines that the equipment is in an abnormal state, it triggers a multi-classifier based on support vector machines. Before system deployment, the classifier is trained using a historical fault case dataset containing at least 500 labeled samples, covering 10 common fault modes such as bearing wear, rotor imbalance, blade breakage, coupling loosening, and gear tooth breakage. The classifier uses a radial basis function kernel, optimizing the penalty parameter C and kernel parameter γ through grid search and cross-validation. The classifier outputs the fault mode category and its confidence score. The third level is the spatial localization of the fault source. This level integrates vibration and acoustic localization information.

[0033] For vibration signals, the time difference between the arrival of the vibration wavefront at each node is calculated using at least four non-coplanar nodes in the vibration sensor array. The direction of arrival (azimuth and elevation angles) of the vibration source is estimated by solving the hyperboloid equations. For acoustic signals, a controllable response power-phase transformation algorithm is used to perform a grid search in the 3D building space model with a step size of 0.1 meters. The acoustic energy focusing response value of each grid point is calculated, and the spatial coordinates (x, y, y) corresponding to the maximum response value are taken. ,y ,z This serves as the initial result for locating the sound source. The direction of arrival of the vibration wave is fused with the coordinates of the sound source.

[0034] This invention employs a Kalman filter algorithm to achieve fusion: the state vector is defined as the three-dimensional coordinates of the sound source. The observation vector is the vibration direction. The system state equation is set as follows: ,in For process noise, for The three-dimensional coordinates of the sound source at that time; the observation equation is ,in Let be the geometric mapping function from spatial coordinates to direction vectors. To observe the noise, a Kalman filter prediction and update loop is used to optimally estimate and smooth the sound source coordinates, ultimately outputting the fused fault source coordinates. The location reliability P is determined by the Kalman gain and the residual covariance, with a typical value ranging from 80% to 99%.

[0035] The early warning decision and visualization module executes tiered early warning logic based on the output of the fault mode recognition and localization engine. The system presets three early warning level thresholds: attention level threshold... =3.0, warning level threshold =4.5, alarm level threshold =6.0. When ∈[ , When this occurs, the module generates an "Equipment Status Deterioration Trend Alert," which includes the equipment name, deviation index, main abnormal characteristic dimensions, and suggested re-inspection cycle. ≥ Furthermore, when the fault classifier outputs a clear fault type (confidence > 85%), an "early warning notification" is generated, which includes the fault type, possible causes, suggested handling measures, and the three-dimensional coordinates of the fault source. .when ≥ When the location reliability P ≥ 95%, an "emergency alarm command" is generated and automatically pushed to the building equipment management system via a standard API interface, triggering coordinated operations such as equipment load reduction, switching to standby units, or requesting shutdown for maintenance. The visualization module uses WebGL technology to build a 3D interactive interface, integrating the building BIM model, sensor deployment topology diagram, real-time signal waterfall diagram, health dashboards for each device, 3D heatmaps for fault location, and historical warning log tables. All interface elements support multi-dimensional filtering and drill-down by time, device, fault type, and other dimensions.

[0036] The entire system adopts a cloud-edge collaborative architecture, as shown in the attached diagram. Figure 4 As shown, edge computing nodes are deployed locally in each equipment room to drive the multimodal signal acquisition array, cache raw data, and perform real-time computations for signal preprocessing and synchronization modules, ensuring data processing latency is less than 100 milliseconds. The edge nodes compress and upload the preprocessed hybrid observation signal matrix X to the cloud server at 10-second intervals. The cloud server cluster handles all computational tasks for the multi-source signal blind separation and reconstruction module, the device feature fingerprint database, and the fault mode recognition and location engine, utilizing GPU acceleration for large-scale matrix operations and machine learning inference. The early warning decision and visualization module acts as a client, subscribing to advanced analysis results from the cloud and directly subscribing to raw alarm streams from the edge nodes, achieving dual guarantees of millisecond-level local alarms and second-level intelligent diagnosis. In a real-world deployment of this cloud-edge collaborative architecture in a large commercial complex, it successfully reduced the single complete analysis cycle from 8 seconds in the traditional centralized architecture to 2.3 seconds, while simultaneously reducing server resource consumption.

[0037] Appendix Figure 5 This invention demonstrates the effectiveness of vibration and acoustic signal fusion localization. Under a single vibration localization method, due to multipath reflections from the building structure, the localization result exhibits a large sector-shaped error zone; single acoustic localization is susceptible to reverberation interference in low signal-to-noise ratio environments, resulting in false hotspots. The fusion localization method of this invention effectively suppresses the random errors of both methods through Kalman filtering, reducing localization errors and improving the engineering practicality of fault location.

[0038] Example 2: In another embodiment, the algorithm implementation of the multi-source signal blind separation and reconstruction module can be further optimized to adapt to the operating characteristics of a specific equipment cluster. When a building electromechanical equipment cluster contains a large number of devices of the same model and speed, the vibration and noise signals are highly similar in the frequency domain, leading to a decrease in the performance of traditional blind separation algorithms. This embodiment introduces a semi-blind separation strategy assisted by prior equipment knowledge. During the system initialization phase, maintenance personnel input structural parameters such as the rated speed, number of motor pole pairs, and number of blades for each device through the human-machine interface. Based on this, the system calculates the theoretical characteristic frequency set of each device, including rotational frequency, harmonics, blade passing frequency, and its sidebands.

[0039] In the objective function of the blind separation algorithm, a regularization term is added. This regularization term penalizes estimated source signals with excessively low energy at theoretical characteristic frequencies. That is, for the One estimated source signal Calculate its theoretical characteristic frequency set at device j. Energy percentage , This represents the frequency of the signal. If... Preliminary identification as equipment The signal (e.g., through initial clustering), but If the value is less than 0.3, an additional penalty is applied to the objective function. This guides the solution process of the separation matrix W, making it more likely to output source signals with energy at theoretical characteristic frequencies, thus effectively distinguishing similar devices. In a field test at a data center cooling station, this semi-blind separation strategy improved the signal separation accuracy of water pumps of the same model.

[0040] In this embodiment, a time decay factor is introduced into the incremental learning mechanism of the device feature fingerprint database. This is because device performance slowly degrades over time, and its health state is not absolutely static. Therefore, the baseline feature fingerprint... Instead of a fixed vector, it is dynamically maintained using an exponentially weighted moving average model: ,in The learning rate is typically 0.01. This multi-source signal blind separation and reconstruction model assigns higher weight to recent health data, enabling the fingerprint database to adaptively track the gradual aging process of devices and avoid frequent false alarms caused by minor performance drift. Simultaneously, the system sets a maximum drift tolerance threshold. =0.5 (measured by Mahalanobis distance), if within 30 consecutive days The cumulative drift is greater than It will automatically remind maintenance personnel to conduct in-depth health assessments to prevent model overfitting from masking real faults.

[0041] In the fault mode recognition and localization engine, this embodiment enhances the detection capability for intermittent faults. Traditional methods rely on continuous signal windows, which are prone to missing short-term sudden faults. This embodiment introduces an event triggering mechanism in the feature extraction stage: when the instantaneous amplitude of any sensor channel exceeds its historical 99.9th percentile, a signal segment of 1 second before and after is immediately extracted for independent blind separation and feature extraction. The analysis result of this signal segment does not participate in routine periodic diagnosis but is stored as an independent event in the event database. If the same device triggers such an event more than 3 times within 24 hours, even if its routine... Even if the threshold is not exceeded, the system will still generate a "suspected intermittent fault" warning, prompting users to check for potential hazards such as poor electrical contact or loose connections. This event triggering mechanism successfully provided a warning 72 hours in advance in a hospital's HVAC system for an intermittent sparking fault caused by a loose motor terminal, preventing equipment burnout.

[0042] In summary, this embodiment further enhances the robustness and adaptability of the system in complex and dynamic operating environments by introducing prior knowledge of the equipment, dynamic fingerprint maintenance, and intermittent fault detection mechanisms, fully demonstrating the scalability and engineering value of the technical solution of this invention.

Claims

1. A multi-signal analysis and early warning system for vibration and noise of building electromechanical equipment, characterized in that, include: A multimodal signal acquisition array is used to acquire mixed vibration and acoustic signals of building electromechanical equipment clusters; The signal preprocessing and synchronization module is used to receive and process the raw signal stream from the multimodal signal acquisition array; The multi-source signal blind separation and reconstruction module is used to separate and reconstruct the source signals corresponding to each independent electromechanical device from the time-strictly synchronized multi-channel hybrid observation signal matrix. Equipment feature fingerprint database, used to store and match the baseline feature fingerprints of various electromechanical equipment; The fault mode recognition and location engine is used to perform status assessment and fault location on the isolated source signals of each device. The early warning decision and visualization module is used to generate graded early warning instructions and provide a human-computer interaction interface based on the output results of the fault mode recognition and localization engine. The early warning decision and visualization module presets multiple early warning level thresholds and generates early warning notifications or alarm instructions of different levels based on the comprehensive state deviation index, fault mode classification results and fault source location reliability.

2. The multi-signal analysis and early warning system for vibration and noise of building electromechanical equipment according to claim 1, characterized in that, The multimodal signal acquisition array includes a vibration sensor array and an acoustic sensor array. All sensor nodes are integrated with a high-precision timing chip to ensure global time synchronization of data acquisition across the entire array. The signal preprocessing and synchronization module performs bandpass filtering on the original vibration and acoustic signals, and performs signal amplitude normalization and timestamp alignment operations to generate a multi-channel hybrid observation signal matrix with strict time synchronization. The multi-source signal blind separation and reconstruction module is based on an improved joint diagonalization blind source separation algorithm. The algorithm calculates the set of cross-correlation matrices of the multi-channel mixed observation signals under multiple different time delays and finds the optimal separation matrix such that the multi-source signal blind separation matrix can simultaneously approximate the set of cross-correlation matrices as diagonal. The process of solving the optimal separation matrix is ​​transformed into a constrained optimization problem of minimizing the weighted sum of the off-diagonal elements of all cross-correlation matrices, and is solved by the iterative Jacobian rotation method. Finally, the optimal separation matrix obtained by the solution is applied to the original mixed observation signal matrix to output a series of statistically independent estimated source signals.

3. The multi-signal analysis and early warning system for vibration and noise of building electromechanical equipment according to claim 2, characterized in that, The device feature fingerprint database establishes the feature fingerprint of each device through the following process: Under the known healthy operating state of the device, the source signal of the device is separated by the multi-source signal blind separation and reconstruction module, and the source signal is feature extracted. The extracted feature set includes time-domain statistical features, frequency-domain spectral features, time-frequency domain wavelet packet energy features, and higher-order statistical features. The extracted feature set is then associated with and stored with the unique identifier of the device. The fault mode recognition and localization engine has a built-in multi-level fault diagnosis model. The first level is feature deviation calculation, which calculates the Mahalanobis distance between the real-time feature vector and the reference feature fingerprint retrieved from the device feature fingerprint database to obtain the comprehensive state deviation index. The second level is fault mode classification. When the overall state deviation index is greater than the first preset threshold, a multi-classifier based on support vector machine is triggered. The third level is the spatial localization of the fault source. The data of the vibration sensor array and the acoustic sensor array are fused. For the vibration signal, the direction of arrival estimation technology based on the time difference of arrival method is used. For the acoustic signal, the sound source localization algorithm based on the controllable response power is used. The vibration source direction estimation and the sound source spatial coordinates are fused to output the coordinate estimation of the faulty equipment in three-dimensional space and the location confidence.

4. The multi-signal analysis and early warning system for vibration and noise of building electromechanical equipment according to claim 3, characterized in that, In the improved joint diagonalization blind source separation algorithm, the weight coefficients of the off-diagonal elements in the objective function of the constrained optimization problem are adaptively allocated according to the eigenvalues ​​of each cross-correlation matrix. The weight coefficients of cross-correlation matrices with larger eigenvalues ​​are correspondingly increased in the objective function. The vibration sensor array uses triaxial accelerometers and is arranged in a grid with a specific spatial density on the equipment foundation, pipe support and key vibration transmission paths of the building structure. The acoustic sensor array consists of multiple high-precision microphones and is deployed in a specific three-dimensional spatial configuration in the equipment room and adjacent spaces.

5. The multi-signal analysis and early warning system for vibration and noise of building electromechanical equipment according to claim 4, characterized in that, In the third level of the fault mode recognition and localization engine, the data fusion process adopts the Kalman filter algorithm. The Kalman filter algorithm uses the vibration wave direction of arrival estimation as the observation vector and the sound source location coordinates as the state vector to establish the system's state space model. Through the prediction and update loop of the Kalman filter, the spatial coordinate estimation of the fault source is optimally fused and smoothed. The device feature fingerprint database supports online incremental learning. When the device is repaired or parts are replaced, the system updates the original feature fingerprint based on the new health status data. The update process introduces feature stability evaluation based on Spearman rank correlation coefficient. When the Spearman correlation coefficient between the feature vector extracted from the newly collected health status data and the historical baseline feature fingerprint is less than 0.8, the system determines that the device operating status baseline has changed and automatically overwrites the old feature fingerprint with the feature vector generated by the new data.

6. The multi-signal analysis and early warning system for vibration and noise of building electromechanical equipment according to claim 5, characterized in that, The system adopts a cloud-edge collaborative computing architecture. The multimodal signal acquisition array and the signal preprocessing and synchronization module are deployed on edge computing nodes. The computing tasks of the multi-source signal blind separation and reconstruction module, the device feature fingerprint library, and the fault mode recognition and location engine are deployed on cloud servers. The early warning decision and visualization module acts as a client, subscribing to processing results and raw alarm streams from both the cloud and edge nodes.

7. The multi-signal analysis and early warning system for vibration and noise of building electromechanical equipment according to claim 6, characterized in that, The early warning decision-making logic of the early warning decision and visualization module is as follows: when the comprehensive status deviation index is greater than the attention level threshold but lower than the early warning level threshold, a device status deterioration trend prompt is generated; When the overall status deviation index is greater than the warning level threshold and the fault mode classifier outputs a clear fault type, a warning notification for the corresponding fault type is generated, along with an estimate of the spatial coordinates of the fault source. When the overall status deviation index is greater than the alarm level threshold and the location confidence of the fault source is higher than 95%, the highest level alarm command is generated, and the linkage interface with the building equipment management system is automatically triggered.

8. The multi-signal analysis and early warning system for vibration and noise of building electromechanical equipment according to claim 7, characterized in that, In the second layer of the fault mode recognition and localization engine, the multi-classifier based on support vector machines, trained according to historical fault cases, can map abnormal features to specific fault modes.

9. The multi-signal analysis and early warning system for vibration and noise of building electromechanical equipment according to claim 8, characterized in that, During the incremental learning process, the device feature fingerprint database records the version update log of the feature fingerprints.

10. The multi-signal analysis and early warning system for vibration and noise of building electromechanical equipment according to claim 9, characterized in that, The edge computing node compresses and uploads the preprocessed hybrid observation signal matrix to the cloud server at periodic intervals.

Images