Intelligent elevator fault prediction method based on Beidou positioning and multi-source data fusion

Abstract

The invention relates to the technical field of multi-modal data processing and predictive analysis, and particularly discloses an elevator fault intelligent prediction method based on Beidou positioning and multi-source data fusion. The method comprises the following steps: acquiring multi-source heterogeneous data containing dynamic time sequence information, physical sensing information and dynamic visual feature vectors; and generating a digital twinning early warning state matrix through a hoistway trajectory algorithm, inputting the digital twinning early warning state matrix into a model comprising a parallel processing path and an attention fusion layer, and calculating a fault probability. By constructing the specific attention fusion model, dynamic weighting and deep fusion are performed on heterogeneous data, and the technical problem that a traditional data processing method is difficult to effectively perform collaborative analysis on multi-source information is solved.

Description

Technical Field

[0001] This invention relates to the field of multimodal data processing and predictive analysis technology, specifically to an intelligent prediction method for elevator faults based on BeiDou positioning and multi-source data fusion. Background Technology

[0002] Predictive maintenance of the state of complex physical systems is crucial for ensuring their safe and stable operation. Current technologies rely on real-time monitoring and analysis of multi-source sensor data to detect anomalies in advance, but they face technical challenges in fusion of heterogeneous data.

[0003] Existing data transmission methods are mostly adept at analyzing structured time-series data from physical sensors, such as vibration, temperature, or location signals. While video surveillance is also frequently used for system monitoring, its data streams are typically processed within separate security or recording systems and are not effectively integrated into the data streams of predictive analytics. Current fusion strategies struggle to perform deep and meaningful correlation analysis between unstructured dynamic visual features extracted from video that characterize specific event behaviors and structured time-series data.

[0004] In summary, existing technologies suffer from bottlenecks when processing multimodal heterogeneous data, namely, the lack of a unified computational framework that can effectively integrate dynamic visual features and physical sensing time-series data. This fragmentation in data processing prevents predictive models from utilizing cross-modal information to identify complex anomaly patterns, thus limiting the accuracy and reliability of prediction results. Therefore, a new data processing method is urgently needed to address this issue.

[0005] To address this, a smart elevator fault prediction method based on BeiDou positioning and multi-source data fusion is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent prediction method for elevator faults based on BeiDou positioning and multi-source data fusion. By constructing an attention model that can deeply integrate dynamic visual features and heterogeneous time-series data, this invention solves the technical problem that existing technologies cannot effectively utilize multimodal information for collaborative analysis, thereby achieving a more comprehensive and accurate assessment of elevator operating status and earlier and more reliable fault prediction.

[0007] To achieve the above objectives, the present invention provides the following technical solution: The intelligent prediction method for elevator faults based on BeiDou positioning and multi-source data fusion includes the following steps: The system acquires holographic operation data of the elevator during operation. The holographic operation data includes dynamic time-series information collected by the Beidou positioning unit, operation information and environmental information collected by sensors, and dynamic visual feature vectors extracted by the video monitoring unit. The shaft trajectory algorithm is used to fuse the preprocessed holographic operation data. The process noise and observation noise covariance matrix parameters of the shaft trajectory algorithm are determined by adaptive optimization based on historical operation data using the particle swarm optimization algorithm, generating a fused digital twin early warning status matrix that can comprehensively characterize the current health status of the elevator. The digital twin early warning state matrix is ​​input into an attention fusion model. The attention fusion model extracts features from the dynamic temporal information, operational information, environmental information, and dynamic visual feature vectors through multiple parallel processing paths. Then, the attention fusion layer dynamically weights and fuses the feature vectors extracted from each path to calculate and output the probability of future failures. Preferably, the preprocessing of the holographic operational data includes the following steps: Based on the timestamp information carried by the holographic operation data, the dynamic time sequence information, operation information, environmental information and event visual features are time-aligned; For the vibration signal in the operation information, a bandpass filtering algorithm is used to process it and filter out noise signals outside the preset frequency band; all feature data after time alignment and denoising are normalized to map the data of each dimension to a unified numerical range.

[0008] Preferably, after calculating the fault probability, it is determined whether the fault probability is greater than a preset fault threshold, and when the fault probability is greater than the fault threshold, the system automatically generates and pushes fault warning information.

[0009] Preferably, the well trajectory algorithm includes a state prediction unit and an observation correction unit; The state prediction unit predicts the prior state estimate at the current moment based on the optimal state estimate of the previous moment and the kinematic model of the elevator. The observation correction unit integrates the real-time collected holographic operation data as the observation value at the current moment, calculates the Kalman gain according to the difference between the prior state estimate and the observation value, corrects the prior state estimate according to the Kalman gain, and outputs the digital twin early warning state matrix.

[0010] Preferably, the specific steps for adaptive optimization using the particle swarm optimization algorithm are as follows: the elements of the process noise and observation noise covariance matrix of the well trajectory algorithm are used as parameters to be optimized to form particles; the function that minimizes the prediction error of historical data is set as the fitness function; and the velocity and position of the particles are updated iteratively to find the parameter combination that makes the fitness function optimal.

[0011] Preferably, the kinematic model of the elevator defines the motion state of the elevator car in the shaft as a state vector containing position and velocity, and uses a constant velocity model to describe the evolution relationship of the state vector in adjacent time steps. Specifically, the predicted position at the current moment is obtained by adding the product of the optimal position estimate at the previous moment, the optimal velocity estimate at the previous moment, and the time step; the predicted velocity at the current moment is equal to the optimal velocity estimate at the previous moment.

[0012] Preferably, the attention fusion model includes: The multi-path feature extraction unit processes the dynamic temporal information, operational information, environmental information, and dynamic visual feature vectors in parallel, and outputs their respective feature vectors. The attention fusion unit receives the feature vectors output by the multi-path feature extraction unit and performs dynamic weighted calculation on the feature vectors to generate a fused global feature vector. The fault probability output unit calculates the fault probability based on the global feature vector through a fully connected layer and an activation function.

[0013] Preferably, the multi-path feature extraction unit includes: The dynamic temporal processing path uses a gated recurrent unit network to process the dynamic temporal information and operational information; the event visual processing path uses a temporal convolutional network to process the dynamic visual feature vectors; and the environmental context processing path uses a multilayer perceptron to process the environmental information.

[0014] Preferably, the attention fusion unit implements dynamic weighted calculation in the following manner: Calculate the corresponding attention score for the feature vector from each processing path; generate normalized attention weights for the feature vector using the attention scores; and perform a weighted summation of the feature vectors based on the attention weights to generate the global feature vector.

[0015] Preferably, after generating the fault warning information, the system automatically analyzes and determines the potential fault type and related components based on the feature data that contributes the most to the fault probability, and generates an intelligent maintenance work order that includes maintenance suggestions and maintenance personnel scheduling plan.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a high-fidelity digital twin early warning status matrix by collecting holographic operational data containing dynamic visual features and performing deep fusion using a shaft trajectory algorithm. This method overcomes the limitations of traditional technologies that rely on single or discrete sensor information, enabling a comprehensive and accurate mapping of the elevator's instantaneous health status from multiple dimensions. This panoramic status perception capability provides an unprecedented high-dimensional, high-quality data foundation for subsequent accurate predictions, fundamentally improving the accuracy of status assessment.

[0017] 2. The attention fusion model designed in this invention solves the technical challenge of effectively fusing heterogeneous data through multi-path parallel processing and dynamic attention weighting mechanisms. In particular, its independent event visual processing path enables a deep understanding and quantification of mechanical behavioral anomalies captured from video. This model can intelligently identify the key information that contributes most to fault prediction, achieving deep correlation mining across modal data and significantly improving the sensitivity and accuracy of fault prediction, especially for early visual signs that are difficult to detect using traditional methods.

[0018] 3. This invention not only provides early warnings of faults but also achieves an intelligent closed loop from "early warning" to "maintenance." After generating early warning information, the system can automatically trace and analyze the source, determine the potential fault type and related components, and generate an intelligent maintenance work order containing specific repair suggestions. This function greatly shortens the time for manual fault diagnosis, reduces reliance on the experience of maintenance personnel, automates and refines predictive maintenance, significantly improves maintenance efficiency, and reduces the overall downtime and maintenance costs of elevators. Attached Figure Description

[0019] Figure 1 This is a flowchart of the intelligent prediction method for elevator faults based on BeiDou positioning and multi-source data fusion proposed in this invention. Figure 2 This is a schematic diagram illustrating the steps of the intelligent elevator fault prediction method based on BeiDou positioning and multi-source data fusion proposed in this invention. Figure 3 This is a schematic diagram of the attention fusion model proposed in this invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1 Please see Figures 1 to 2This invention provides an intelligent prediction method for elevator faults based on BeiDou positioning and multi-source data fusion. The technical solution is as follows: An intelligent prediction method for elevator faults based on BeiDou positioning and multi-source data fusion, such as Figures 1-2 This includes the following steps: The system acquires holographic operation data of the elevator during operation. The holographic operation data includes dynamic time-series information collected by the Beidou positioning unit, operation information and environmental information collected by sensors, and dynamic visual feature vectors extracted by the video monitoring unit. The shaft trajectory algorithm is used to fuse the preprocessed holographic operation data. The process noise and observation noise covariance matrix parameters of the shaft trajectory algorithm are determined by adaptive optimization based on historical operation data using the particle swarm optimization algorithm, generating a fused digital twin early warning status matrix that can comprehensively characterize the current health status of the elevator. The digital twin early warning state matrix is ​​input into the attention fusion model; the attention fusion model extracts features from the dynamic time-series information, operation information, environmental information and dynamic visual feature vectors through multiple parallel processing paths, and then dynamically weights and fuses the feature vectors extracted by each path through the attention fusion layer to calculate and output the probability of future failures.

[0022] Furthermore, the preprocessing of the holographic operational data includes the following steps: Based on the timestamp information carried by the holographic operation data, the dynamic time sequence information, operation information, environmental information and event visual features are time-aligned; For the vibration signal in the operation information, a bandpass filtering algorithm is used to process it and filter out noise signals outside the preset frequency band; all feature data after time alignment and denoising are normalized to map the data of each dimension to a unified numerical range.

[0023] The dynamic visual feature vector extracted by the video surveillance unit is specifically extracted by the following steps: the surveillance video of the elevator door area is divided into 2-second video segments at a frame rate of 25fps, the segments are input into a pre-trained I3D model, and the output of its last global average pooling layer is extracted to generate a 1024-dimensional feature vector, which is used to characterize the motion pattern of the elevator door during the time period.

[0024] Based on the timestamp information carried by the holographic operation data, and using the 10Hz sampling frequency of the dynamic timing information as a benchmark, the dynamic timing information, operation information, environmental information, and event visual features are time-aligned. Specifically, vibration signals with sampling frequencies higher than 10Hz are downsampled using a sliding window averaging method, and data with sampling frequencies lower than 10Hz (if any) are upsampled using a forward padding method to generate a multidimensional feature sequence that is strictly synchronized in time. In this embodiment, the bandpass filtering algorithm specifically adopts a fourth-order Butterworth bandpass filtering algorithm with a cutoff frequency of 10Hz to 500Hz. This preset frequency band can effectively cover the common fault characteristic frequencies of major rotating components such as elevator traction machines, bearings and gearboxes, while filtering out power frequency interference and other high-frequency random noise signals. For all feature data that have undergone time alignment and denoising, the max-min normalization method is used to calculate and linearly map each dimension of data to the numerical range of [0, 1] to eliminate the influence of different dimensions between features.

[0025] The preprocessing method of this invention achieves accurate extraction of effective fault features and targeted noise filtering by setting specific bandpass filters covering the core fault frequency, significantly improving the signal-to-noise ratio. Simultaneously, its explicit multi-rate data synchronization strategy ensures high consistency and effectiveness of the data sources in the time dimension before fusion. The final normalization process lays the foundation for stable training of the subsequent model.

[0026] Furthermore, after calculating the fault probability, it is determined whether the fault probability is greater than a preset fault threshold, and when the fault probability is greater than the fault threshold, the system automatically generates and pushes fault warning information.

[0027] The method for determining the preset fault threshold includes the following steps: testing the trained attention fusion model using a validation dataset containing normal samples and known fault samples; statistically analyzing the true positive rate and false positive rate of the model under different probability thresholds to plot the receiver operation feature curve; and selecting an optimal balance point on the receiver operation feature curve that balances high recall and low false positive rate based on actual operational needs, and using the corresponding probability value as the preset fault threshold. For example, in one embodiment, the probability value of 0.95, the point with the highest TPR when the FPR does not exceed 5%, is selected as the threshold.

[0028] This invention scientifically determines fault warning thresholds by employing receiver operating characteristic curve analysis, transforming threshold selection from traditional experience-based setting to a data-driven optimization process. This method enables the system to achieve a quantifiable optimal balance between false alarm and missed alarm rates, and allows for customized adjustments based on actual operational needs. This significantly improves the reliability, accuracy, and practical value of warnings, effectively preventing invalid alarms or missed critical faults due to improper threshold settings.

[0029] Furthermore, the well trajectory algorithm includes a state prediction unit and an observation correction unit; The state prediction unit predicts the prior state estimate at the current moment based on the optimal state estimate of the previous moment and the kinematic model of the elevator. The observation correction unit integrates the real-time collected holographic operation data as the observation value at the current moment, calculates the Kalman gain according to the difference between the prior state estimate and the observation value, corrects the prior state estimate according to the Kalman gain, and outputs the digital twin early warning state matrix.

[0030] The observation correction unit maps the prior state estimate to the observation space through an observation model, and integrates the real-time acquired holographic operation data as the observation value at the current moment; the observation model includes: The observation matrix H is used to define the linear relationship between the state vector and the observation vector; The observation noise covariance matrix R is used to characterize the measurement uncertainty of each data source in the holographic operation data; In one specific embodiment, when the state vector is [position, velocity], the observation vector mainly uses the position information of the Beidou positioning unit, and the observation matrix H can be set to [1, 0]; the value of the observation noise covariance matrix R is preset based on the long-term test of the Beidou positioning unit in a static environment and the statistical variance of its position drift error.

[0031] The shaft trajectory algorithm is used to generate a smooth elevator state estimate. Its state prediction and observation correction unit employs a Kalman filter, which uses only the position information from the BeiDou positioning unit as observations to optimally estimate the elevator's position and speed. Then, in addition to the filter's output, this optimal state estimate (position and speed) is temporally aligned with other preprocessed holographic operational data (such as vibration signals, environmental information, and dynamic visual feature vectors) to form an augmented feature vector, which is the digital twin early warning state matrix.

[0032] In this embodiment, the digital twin early warning state matrix is ​​a two-dimensional matrix of shape (N, T), where N is the number of features and T is the number of time steps. T is set to 100, corresponding to a 10-second time window at a 10Hz sampling rate; the N feature dimensions include: position and velocity output by the well trajectory algorithm, vibration signal after bandpass filtering and normalization, normalized temperature and humidity, and a 16-dimensional dynamic visual feature vector.

[0033] This invention transforms the fuzzy data fusion process into rigorous and implementable mathematical steps by explicitly defining the observation model, particularly the observation matrix H and the observation noise covariance matrix R. This method not only establishes an explicit correlation between the model's internal state and external sensor data but also enables the algorithm to perform optimal weighted fusion based on the statistically calibrated reliability of each data source. This significantly improves the accuracy and robustness of state estimation, providing a core technological guarantee for generating high-fidelity digital twin state matrices.

[0034] Furthermore, the specific steps for adaptive optimization using the particle swarm optimization algorithm are as follows: the elements of the process noise and observation noise covariance matrix of the well trajectory algorithm are used as parameters to be optimized to form particles; the function that minimizes the prediction error of historical data is set as the fitness function; and the optimal combination of parameters for the fitness function is found by iteratively updating the velocity and position of the particles.

[0035] In this embodiment, the core of the wellbore trajectory algorithm is the Kalman filter, whose performance is highly dependent on the settings of the process noise covariance matrix Q and the observation noise covariance matrix R. To overcome the blindness of manual parameter tuning, this invention uses the particle swarm optimization algorithm (PSO) for adaptive optimization.

[0036] The key diagonal elements in matrices Q and R are selected as parameters to be optimized.

[0037] In the PSO algorithm, the position vector of each "particle" represents a set of candidate Q and R parameters.

[0038] The fitness function is defined as minimizing the prediction error of the filter on a historical "holographic running data". Specifically, a set of Q and R parameters are substituted into the well trajectory algorithm to filter the historical data, and the root mean square error between the output state estimate and the true value (or high-precision reference value) is calculated. This error value is the fitness of the particle; the smaller the error, the higher the fitness.

[0039] The algorithm initializes a set of random particles and iterates multiple times according to the velocity and position update formulas of the standard PSO. In each iteration, the particles follow the "individual historical best solution" and the "population historical best solution," gradually searching the entire parameter space until they find a set of Q and R parameters that minimize the prediction error.

[0040] This optimization process can be performed offline periodically, updating the optimal values ​​of Q and R with the latest historical data and configuring these optimal values ​​into the online well trajectory algorithm.

[0041] This invention introduces a particle swarm optimization algorithm to achieve adaptive optimization of key parameters in wellbore trajectory algorithms, overcoming the shortcomings of traditional Kalman filters that rely on manual experience and have fixed parameters. This method can automatically tune the optimal noise parameters based on actual historical data, resulting in higher accuracy and robustness of the state estimation algorithm, thus providing a more reliable data foundation for subsequent fault prediction.

[0042] Furthermore, the kinematic model of the elevator defines the motion state of the elevator car in the shaft as a state vector containing position and velocity, and uses a constant velocity model to describe the evolution relationship of the state vector in adjacent time steps. Specifically, the predicted position at the current moment is obtained by adding the product of the optimal position estimate at the previous moment, the optimal velocity estimate at the previous moment, and the time step; the predicted velocity at the current moment is equal to the optimal velocity estimate at the previous moment.

[0043] Preferably, to further improve the model's tracking accuracy of the elevator's actual motion, the kinematic model also incorporates acceleration information as a control input to compensate for the predicted position and predicted velocity. Simultaneously, to fully implement the Kalman filtering prediction step, the kinematic model also includes a preset process noise covariance matrix Q. This matrix is ​​used to quantify the uncertainty between the model and the elevator's actual motion, and is used to calculate the prior covariance matrix at the current moment while predicting the state vector.

[0044] By introducing measured acceleration as a control input, the idealized constant velocity model is upgraded to a dynamic model that is closer to physical reality, significantly improving the instantaneous accuracy of state prediction. Simultaneously, by introducing the process noise covariance matrix Q to quantify the model's uncertainty, the robustness and long-term stability of the filtering algorithm are greatly enhanced, avoiding tracking failures caused by the accumulation of model biases.

[0045] The beneficial effects of this kinematic model lie in its dual optimization: by using measured acceleration compensation for prediction, the model closely reflects the actual dynamics of the elevator, improving instantaneous prediction accuracy; and by quantifying the uncertainty of the model through a process noise matrix, the robustness and stability of the algorithm in long-term tracking are guaranteed. This prediction mechanism, combining accuracy and robustness, provides a core guarantee for generating high-fidelity state estimates.

[0046] Furthermore, such as Figure 3 As shown, the attention fusion model includes: The multi-path feature extraction unit processes the dynamic temporal information, operational information, environmental information, and dynamic visual feature vectors in parallel, and outputs their respective feature vectors. The attention fusion unit receives the feature vectors output by the multi-path feature extraction unit and performs dynamic weighted calculation on the feature vectors to generate a fused global feature vector. The fault probability output unit calculates the fault probability based on the global feature vector through a fully connected layer and an activation function.

[0047] The activation function used by the fault probability output unit is the Sigmoid function, which maps the output of the fully connected layer to the (0, 1) interval to obtain the fault probability characterizing the possibility of binary classification faults.

[0048] This invention explicitly employs the Sigmoid activation function in the fault probability output unit, giving the model's final output a clear and standardized probabilistic meaning. This design transforms the abstract numerical values ​​within the neural network into a fault probability score within the (0, 1) interval that can be intuitively understood externally. This not only greatly enhances the interpretability of the prediction results but also provides a standardized mathematical foundation for subsequent scientific and reliable threshold judgments and early warning decisions, ensuring the rigor and practical value of the entire system's output.

[0049] Furthermore, the multi-path feature extraction unit includes: The dynamic temporal processing path uses a gated recurrent unit network to process the dynamic temporal information and operational information; the event visual processing path uses a temporal convolutional network to process the dynamic visual feature vectors; and the environmental context processing path uses a multilayer perceptron to process the environmental information.

[0050] Preferably, the gated recurrent unit network of the dynamic temporal processing path contains two hidden layers, each with 128 hidden units, and finally outputs a 64-dimensional feature vector. The temporal convolutional network of the event visual processing path contains 4 residual blocks with a kernel size of 3. The dilation coefficient increases exponentially with a base of 2 and finally outputs a 64-dimensional feature vector. The multilayer perceptron in the environmental context processing path contains two fully connected hidden layers, each with 32 neurons and ReLU as the activation function, and finally outputs a 16-dimensional feature vector.

[0051] This invention designs a heterogeneous and dedicated network architecture for the multi-path feature extraction unit. For three data modalities—dynamic temporal, event-driven visual, and environmental context—GRU, TCN, and MLP networks, respectively, are employed to best match their characteristics. This tailored design ensures in-depth mining and optimal feature extraction of unique information from each data source, providing a high-quality feature foundation for efficient subsequent attention fusion and high prediction accuracy.

[0052] Furthermore, the attention fusion unit implements dynamic weighted calculation in the following manner: Calculate the corresponding attention score for the feature vector from each processing path; generate normalized attention weights for the feature vector using the attention scores; and perform a weighted summation of the feature vectors based on the attention weights to generate the global feature vector.

[0053] The attention score is calculated using an additive attention model, which is implemented through a fully connected neural network with a single hidden layer and shared parameters, and a context vector that can be learned jointly. Specifically, the feature vectors of each path are nonlinearly transformed through the fully connected neural network, and the transformation result is multiplied by the context vector to obtain the attention score corresponding to each feature vector. The normalized attention weights are generated by applying the Softmax function to all attention scores.

[0054] This invention upgrades feature fusion from static concatenation to a dynamic, intelligent weighting process by employing a specific additive attention model. This model automatically assesses the importance of each data source based on real-time data content using a learnable neural network, enabling the system to focus on the most critical fault symptoms, regardless of whether they originate from temporal, visual, or environmental information. This adaptive fusion capability significantly enhances the accuracy of capturing complex cross-modal faults and is a core guarantee for improving the final prediction accuracy.

[0055] Furthermore, after generating the fault warning information, based on the feature data that contributes the most to the fault probability, the system automatically analyzes and determines the potential fault type and associated components, and generates an intelligent maintenance work order that includes maintenance suggestions and maintenance personnel scheduling plans.

[0056] The calculation of the failure probability contribution is achieved by analyzing the attention weights generated by the attention fusion unit for each path feature vector. Feature data with higher attention weights are identified as feature data with the highest contribution. The automatic analysis and determination of potential fault types and associated components is achieved by querying a preset fault knowledge base. The fault knowledge base stores the correlation between different high-contribution feature data combinations and specific fault types in the form of rules and / or mapping tables. The intelligent maintenance work order is generated based on the fault knowledge base, which stores standard maintenance suggestions, required spare parts models, and recommended maintenance personnel skill levels for each fault type. The system retrieves and populates this information from the knowledge base according to the determined fault type, and automatically generates the work order.

[0057] This invention cleverly utilizes the weights inherent in the attention model to provide intrinsic interpretability for deep learning predictions, transforming a "black box" model into a transparent decision-making system. This method can clearly trace and identify the key feature data leading to the warning, turning a vague alarm signal into a clearly directional diagnostic clue. This provides a core basis for accurately determining the fault type, greatly improving the efficiency and accuracy of fault diagnosis.

[0058] This invention constructs holographic elevator operation data by combining dynamic visual feature vectors characterizing elevator door opening and closing behavior with dynamic time-series information from BeiDou positioning and traditional sensor information. This method not only generates a high-fidelity digital twin state matrix through initial fusion, but the subsequent attention fusion model can intelligently and adaptively dynamically weight these heterogeneous multimodal data. This allows the model to automatically focus on key signs that contribute most to fault prediction, thereby deeply mining and identifying complex cross-modal fault modes that cannot be detected by traditional single data sources, ultimately significantly improving the accuracy of fault prediction and the lead time for early warning.

[0059] Example 2 This embodiment describes the application of the method of the present invention in diagnosing a specific early-stage fault related to a local location in the hoistway: "car micro-vibration caused by guide shoe wear and momentary jamming of the door operator." This fault is characterized by its intermittency, concealment, and correlation of multiple phenomena, making it extremely difficult to detect using traditional single-sensor threshold methods.

[0060] In this scenario, premature wear on one of the elevator's guide shoes causes a short-duration, low-amplitude high-frequency vibration when the car travels to a specific area between the 15th and 16th floors (where there may be minor unevenness in the guide rails). When the elevator decelerates and stops at the 16th floor in this area, this micro-vibration is briefly transmitted to the door operator system, causing a barely perceptible, extremely slight "lag" or "shaking" in the first 100 milliseconds of the door opening action.

[0061] In this process, the holographic operation data collected by the method of this invention is as follows: Dynamic timing information: The Beidou positioning unit accurately records the millimeter-level position trajectory of the elevator car as it passes through the 15-16 floor interval each time it runs.

[0062] Operational information: The accelerometer installed on the car collects a similar, low-amplitude, non-dominant high-frequency vibration signal each time it passes through this specific area.

[0063] Dynamic visual feature vector: The video surveillance unit continuously monitors the door operator system. When the elevator stops at the 16th floor, the initial motion component of the dynamic visual feature vector extracted by the unit has a slight, quantifiable difference compared with the feature vector when the elevator stops at other floors.

[0064] After the holographic operation data is preprocessed and initially fused by the well trajectory algorithm, the resulting digital twin early warning state matrix is ​​input into the attention fusion model.

[0065] The GRU network with dynamic temporal processing path learns and identifies the weak vibration pattern that recurs at a specific location from the mixed vibration signals and outputs it as an abnormal temporal feature.

[0066] The TCN network for the event visual processing path identifies the visual feature vector of the initial opening of the door on the 16th floor as a low-probability abnormal event by comparing a large number of normal door opening and closing events.

[0067] The MLP network of the environmental context processing path confirms that the current environmental temperature, humidity and other information are normal.

[0068] When the two anomalous features (weak vibration and delayed door opening) occur simultaneously, the attention fusion unit, through training, has learned that there is a strong correlation between this "micro-vibration at a specific location" and "minor anomalies in door opening and closing at that location." Therefore, the attention mechanism simultaneously assigns extremely high attention weights to feature vectors from both the dynamic temporal processing path and the event visual processing path. After these two highly weighted anomalous features are fused, the fault probability calculated by the "fault probability output unit" instantly spikes to 0.98, far exceeding the preset threshold of 0.95, thus triggering an early warning.

[0069] After sending out the warning information, the system immediately enters the diagnostic phase: System analysis revealed that the most significant contributing data to this early warning came from "BeiDou positioning data (15-16 floor range)," "high-frequency components of vibration signals," and "visual features of the door operator opening event in its initial stage."

[0070] The system uses this "high-contribution feature combination" to query the preset fault knowledge base. The knowledge base contains a mapping rule: "IF abnormal vibration at a specific location AND abnormal visual opening and closing of the door at that location THEN potential fault is 'guide shoe / rail problem'".

[0071] The system automatically generates an intelligent maintenance work order with the following content: "Warning type: Highly suspected guide shoe wear or local unevenness of the guide rail. Fault location: Area between floors 15 and 16 of the hoistway. Maintenance suggestion: Maintenance personnel should focus on checking the wear of the guide rail joints and the guide shoes on the corresponding side of the car in this area." Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent elevator fault prediction method based on Beidou positioning and multi-source data fusion, characterized in that, The method comprises the following steps: acquiring holographic operation data of the elevator in operation, the holographic operation data comprising dynamic time sequence information collected by a Beidou positioning unit, operation information and environmental information collected by a sensor, and a dynamic visual feature vector extracted by a video monitoring unit; fusing the preprocessed holographic operation data by using a shaft trajectory algorithm, process noise and observation noise covariance matrix parameters of the shaft trajectory algorithm being determined by using a particle swarm optimization algorithm to perform adaptive optimization according to historical operation data, to generate a fused digital twin early warning state matrix capable of comprehensively representing the current health status of the elevator; inputting the digital twin early warning state matrix into an attention fusion model; the attention fusion model performs feature extraction on the dynamic time sequence information, the operation information and environmental information, and the dynamic visual feature vector through a plurality of parallel processing paths, and performs dynamic weighted fusion on the feature vectors extracted by the paths through an attention fusion layer to calculate a fault probability of a possible future fault.

2. The intelligent elevator fault prediction method based on Beidou positioning and multi-source data fusion according to claim 1, characterized in that, The preprocessing of the holographic operation data comprises the following steps: time aligning the dynamic time sequence information, the operation information, the environmental information, and the event visual features based on timestamp information carried by the holographic operation data; for a vibration signal in the operation information, performing band-pass filtering to filter out noise signals outside a preset frequency band; and performing normalization calculation on all feature data after time alignment and denoising to map the data in each dimension to a unified numerical interval.

3. The elevator fault intelligent prediction method based on Beidou positioning and multi-source data fusion according to claim 1, characterized in that: after the fault probability is calculated, it is determined whether the fault probability is greater than a preset fault threshold, and when the fault probability is greater than the fault threshold, the system automatically generates and pushes fault early warning information.

4. The intelligent elevator fault prediction method based on Beidou positioning and multi-source data fusion according to claim 1, characterized in that: The shaft trajectory algorithm comprises a state prediction unit and an observation correction unit. The state prediction unit predicts a prior state estimation value at the current time based on the optimal state estimation at the previous time and in combination with a kinematic model of the elevator; the observation correction unit fuses the holographic operation data collected in real time as observation values at the current time, calculates a Kalman gain according to the difference between the prior state estimation value and the observation values, modifies the prior state estimation value according to the Kalman gain, and outputs the digital twin early warning state matrix.

5. The intelligent elevator fault prediction method based on Beidou positioning and multi-source data fusion according to claim 1, characterized in that: The specific steps of adaptive optimization by using the particle swarm optimization algorithm are as follows: elements of the process noise and observation noise covariance matrix of the shaft trajectory algorithm are taken as particles to be optimized; a function of minimizing historical data prediction error is set as a fitness function; and the speed and position of the particles are updated through iteration to find a parameter combination that optimizes the fitness function.

6. The elevator fault intelligent prediction method based on Beidou positioning and multi-source data fusion according to claim 4, characterized in that: The kinematic model of the elevator defines the motion state of the elevator car in the shaft as a state vector containing position and velocity, and uses a constant velocity model to describe the evolution relationship of the state vector within adjacent time steps, specifically: the predicted position at the current time is obtained by adding the optimal position estimate at the last time, the optimal velocity estimate at the last time and the product of the time step; the predicted velocity at the current time is equal to the optimal velocity estimate at the last time.

7. The intelligent elevator fault prediction method based on Beidou positioning and multi-source data fusion according to claim 1, characterized in that, The attention fusion model comprises: A multi-path feature extraction unit which processes the dynamic timing information, operation information, environment information and dynamic visual feature vector in parallel and outputs respective feature vectors; An attention fusion unit which receives the feature vectors output by the multi-path feature extraction unit and performs dynamic weighting calculation on the feature vectors to generate a fused global feature vector; A fault probability output unit which calculates the fault probability based on the global feature vector through a fully connected layer and an activation function.

8. The intelligent elevator fault prediction method based on Beidou positioning and multi-source data fusion according to claim 7, characterized in that, The multi-path feature extraction unit comprises: A dynamic timing processing path which processes the dynamic timing information and operation information using a gated recurrent unit network; an event visual processing path which processes the dynamic visual feature vector using a timing convolution network; An environment context processing path which processes the environment information using a multi-layer perceptron. 9.The intelligent elevator fault prediction method based on Beidou positioning and multi-source data fusion according to claim 7, characterized in that, The attention fusion unit performs dynamic weighting calculation in the following manner: Calculate the corresponding attention scores for the feature vectors from each processing path; use the attention scores to generate normalized attention weights for the feature vectors; and perform weighted summation on the feature vectors according to the attention weights to generate the global feature vector.

10. The Beidou positioning and multi-source data fusion-based elevator fault intelligent prediction method according to claim 3, characterized in that: After generating the fault warning information, the feature data with the highest contribution to the fault probability is used to automatically analyze and determine the potential fault type and associated components, and an intelligent operation work order containing maintenance suggestions and operation personnel scheduling schemes is generated.

Images