Data anomaly detection method based on multi-scale feature extraction

Abstract

The present invention discloses a data anomaly detection method based on multi-scale feature extraction, which relates to the technical field of data anomaly detection. The method sequentially performs multi-source data acquisition and environmental calibration, noise robustness multi-scale feature extraction, scenario-aware anomaly detection and sensor redundancy verification, adaptive feedback and collaborative update decision-making, as well as intelligent threshold learning and scenario adaptation. First, multiple signals such as acceleration, current, and electromagnetic signals are acquired and labeled to construct environmental labels. Then, a TCN and LSTM fusion model is used to extract short-term impacts and long-term trends. Then, the threshold is dynamically adjusted according to the elevator operating conditions and noise index, and fault redundancy verification is performed. Subsequently, the model, sensor weights, and thresholds are corrected through false positive and false negative feedback. Finally, a monitoring-detection-feedback-update-adaptive closed loop is constructed based on reinforcement learning or meta-learning for cross-seasonal adaptation, achieving high precision, low false positives, and strong robustness, significantly improving elevator operation safety and maintenance efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of data anomaly detection, and in particular to a data anomaly detection method based on multi-scale feature extraction. Background Art

[0002] In modern industrial and commercial environments, the collection and analysis of operational data from equipment and systems is crucial to ensuring safety and efficiency. Faced with multi-source heterogeneous data, high noise interference, and dynamically changing operating conditions, traditional single data processing methods and fixed threshold detection technologies are unable to effectively cope, resulting in frequent false alarms or missed alarms. In recent years, with the development of artificial intelligence and Internet of Things technologies, data processing methods based on deep learning have gradually emerged, but they still face challenges such as high environmental noise, insufficient multi-source data fusion, and limited online update capabilities. How to use advanced data processing and intelligent algorithms to comprehensively analyze multi-source sensor signals and still achieve stable and accurate fault identification in high-interference environments has become one of the current problems that the industry needs to solve.

[0003] In recent years, with the development of artificial intelligence and industrial IoT technologies, fault diagnosis solutions based on deep learning models such as convolutional networks and recurrent neural networks have gradually emerged. Some studies have used one-dimensional convolution techniques to reduce the dimensionality of time-domain signals, combining them with temporal convolutional networks (TCNs) or neural ordinary differential equations (NeuralODEs) to extract local fault characteristics of equipment. Other teams have introduced self-attention mechanisms to capture relevant information globally. These methods have improved the accuracy of fault detection for short time series signals. However, practical deployment still faces the following challenges:

[0004] Large environmental noise and load fluctuations: When equipment is simultaneously affected by electromagnetic interference, mechanical vibration, temperature and humidity fluctuations, and other factors, sensor output signals often exhibit highly random and sudden fluctuations. This can easily cause algorithms to misinterpret high disturbances under normal operating conditions as faults, or to drown out true fault characteristics by drowning them out in the noise, leading to missed detections. Lack of multi-source data fusion and adaptability: Equipment monitoring typically involves multiple sensor channels, but traditional methods often process each channel's data independently or perform simple weighting. These methods lack deep modeling of inter-sensor correlations and the ability to automatically adjust fusion strategies based on operating conditions. Fixed fusion solutions often fail, especially when loads switch suddenly or operating modes change frequently. Insufficient online updates and continuous learning: Some intelligent fault detection methods are often trained on large offline datasets before being deployed in the field, making it difficult to track the long-term evolution of equipment and environmental characteristics. Without online learning or adaptive updates after deployment, detection accuracy can significantly decline when encountering new disturbances or load patterns outside the training set distribution.

[0005] In view of this, the purpose of the present invention is to provide a multi-source signal adaptive processing and intelligent fault detection method, which can perform real-time monitoring of key electromechanical equipment in actual places (such as shopping malls, airports, industrial production lines, etc.) where noise levels fluctuate greatly with time period and load conditions; through technical means such as dynamic threshold adjustment, sensor redundancy verification, short-term and long-term feature coupling and online adaptive learning, it solves the detection difficulties caused by environmental noise and load instability.

[0006] This method is not only applicable to certain vertical transportation equipment (such as elevators in large shopping malls), but can also be widely used in other electromechanical equipment fault detection scenarios with strong environmental interference, providing a feasible way to ensure equipment safety and reduce operation and maintenance costs. Summary of the Invention

[0007] (1) Technical problems solved

[0008] In response to the shortcomings of the existing technology, the present invention provides a data anomaly detection method based on multi-scale feature extraction. By first acquiring and labeling multiple signals such as acceleration, current, and electromagnetic, environmental labels are constructed; then, the TCN and LSTM fusion model is used to extract short-term impacts and long-term trends; the threshold is dynamically adjusted and faults are redundantly checked according to the elevator operating conditions and noise index; the model, sensor weights and thresholds are then corrected through false positive and missed negative feedback; finally, relying on reinforcement learning or meta-learning for cross-season adaptation, a monitoring-detection-feedback-update-adaptive closed loop is constructed, which significantly improves the safety of elevator operation and maintenance efficiency; thereby solving the technical problems recorded in the background technology.

[0009] (2) Technical solution

[0010] To achieve the above objectives, the present invention is implemented through the following technical solutions: a data anomaly detection method based on multi-scale feature extraction, including automatically calling multi-source sensor data when the elevator is in different operating conditions, using a unified order benchmark through filtering and time synchronization, combining passenger flow and situational factors to generate multi-dimensional environmental labels, and constructing a high-confidence dataset for subsequent multi-scale analysis;

[0011] After preliminary cleaning of multi-source sensor data, the LTCN structure is enabled in multiple noise enhancement scenarios. Short-term convolutional network (TCN) is used to capture transient impacts, and long-term dependencies are then extracted using LSTM. This outputs multi-scale features and enhances fault detection sensitivity.

[0012] After the multi-scale features are generated and the elevator operating conditions and environmental indexes are known, a dynamic threshold is constructed based on the difference. Through multi-sensor redundancy verification, alarms are triggered only for multi-source consistent anomalies, reducing false alarms caused by noise interference.

[0013] If a suspicious result is output and true feedback is obtained, the multi-source sensor data and multi-scale features are back-traced to perform false positive and negative sample re-labeling, incrementally revise the judgment threshold and sensor weight, and update the multi-scale model;

[0014] When the environment changes suddenly due to seasonal changes or cross-shopping mall deployment needs arise, reinforcement learning and meta-learning mechanisms are enabled to adaptively modify the judgment threshold and migrate some LTCN parameters based on the current state description and reward function.

[0015] Furthermore, various types of sensors are installed and activated to continuously collect multi-source data streams in different typical time periods to form a multi-source data stream collection, record the elevator operating conditions and related environmental conditions during the time period, and construct an environmental index function after preprocessing the collected data to measure the noise and disturbance levels under different market conditions.

[0016] Furthermore, the obtained multi-source data streams and environmental indices are time-synchronized to form a multi-dimensional record containing elevator operation data and environmental index functions, introducing the elevator working status and generating contextual elements from the mall operation information;

[0017] Combining elevator working status, environmental index and situational factors, multi-dimensional environmental labels are generated for multi-source data records.

[0018] Furthermore, a multi-source data stream set is used as the input of the short-term convolution channel, and transient fault signals under short time scales are extracted; the output of the short-term convolution channel is defined as a short-term feature matrix, and the short-term feature matrix is ​​fused with the environmental index and the elevator working status through element-level weighting or attention mechanism to output the corrected short-term feature vector.

[0019] Furthermore, the serialized short-term correction features are fed into the LSTM layer to obtain the long-term memory vector. At the same time, the short-term correction features are concatenated or fused with the long-term memory vector by attention weights to obtain the final multi-scale feature vector:

[0020] Combined with multi-dimensional environmental labels, noise enhancement is performed on normal operating condition data during the model training phase, and a multi-objective loss function is defined to enable the model to perform weighted learning in high-noise scenarios and reduce the risk of overfitting low-noise samples.

[0021] Furthermore, in order to dynamically adjust the alarm sensitivity under different working conditions and environmental interference, a scenario-dependent threshold function is introduced by combining the elevator operating status and environmental index. An anomaly score function is constructed to measure the difference between the multi-scale feature vector and the reference vector to measure the degree of potential fault deviation.

[0022] The comparison between the anomaly score function and the threshold function is used to determine whether a fault is triggered. Threshold function, it is determined to be a suspected fault, otherwise it is considered normal or noise fluctuation.

[0023] Furthermore, a single-point anomaly score is defined for each sensor to obtain a score vector. After the suspected fault is identified, a multi-source consistency check is performed on the score vector, and a matrix metric is introduced to characterize the overall coupling degree of all sensor anomaly scores.

[0024] If the matrix metric exceeds the redundant fusion threshold, it is determined to be a multi-source confirmation fault, triggering an alarm or entering adaptive feedback and disposal. Otherwise, the anomaly is regarded as a non-real fault situation such as noise interference or single-source sensor failure.

[0025] Furthermore, the labels of suspected faults and multi-source confirmed faults are matched with the actual status feedback to generate a judgment result table, including correct reports, false positives and missed reports, and normal or suspected fluctuation records;

[0026] Samples that are confirmed to be false positives or missed negatives are relabeled and supplemented with corresponding real fault labels or normal labels to form updated training or calibration data; samples that are correctly reported or judged to be normal are retained, and a feasible source review is carried out.

[0027] Furthermore, after collecting the latest correct error, false alarm and missed alarm information, the threshold function is updated;

[0028] If a single sensor channel frequently conflicts with the judgment of other channels, the corresponding fusion weight of the channel in the redundant check will be reduced accordingly; when the false alarm rate of a sensor is higher than expected, the corresponding weight will be lowered; if its contribution to fault judgment is highly reliable, the corresponding weight will be increased.

[0029] Furthermore, for the multi-scale model LTCN, incremental or periodic retraining is performed using re-labeled fault samples and normal samples to update the learnable parameters;

[0030] During the update process, a loss function with a penalty term is defined, and the multi-scale model is iteratively updated in small batches or online, so that it continuously approaches the optimal state under the guidance of new environment data and failure cases.

[0031] The updated dynamic threshold function, sensor weight vector, and retrained model parameters are packaged into the decision layer so that the latest configuration can be immediately adopted in the next round of detection.

[0032] Furthermore, the dynamically updated multi-dimensional environment label is combined with the current threshold configuration, sensor weight vector, and model parameter key index into a state description, which is used as the input of the reinforcement learning or meta-learning algorithm and a multi-dimensional reward function is established.

[0033] In the reinforcement learning framework, the threshold strategy or model migration method for the next decision moment is selected or updated based on the current state description and reward function.

[0034] Furthermore, we make migration decisions for multi-scale feature extraction models and define model migration strategies under the framework of meta-learning or reinforcement learning.

[0035] The threshold policy function and the model transfer strategy are encapsulated in a unified reinforcement learning or meta-learning update loop, and gradient optimization is performed through the reward function obtained through interaction.

[0036] Furthermore, the updated threshold strategy function and model migration strategy are implemented in real time in the elevator system. Whenever a new operating cycle arrives, the current state description is used to determine whether to revise the threshold and / or update the local model. The fault detection results after each decision are recorded.

[0037] When facing different shopping malls or seasonal switches, the existing threshold adjustment strategy and model migration strategy will be migrated as a meta-model. If there are persistent high false positives or high missed negatives, reminders will be issued to the outside world or feedback and collaborative updates will be triggered.

[0038] (3) Beneficial effects

[0039] The present invention provides a data anomaly detection method based on multi-scale feature extraction, which has the following beneficial effects:

[0040] Through the coordinated implementation of five steps—multi-source data collection and operating environment calibration, multi-scale feature extraction and noise robustness model construction, context-aware anomaly detection and sensor redundancy verification, adaptive feedback and collaborative update decisions, and intelligent threshold learning and scenario adaptation—the system significantly improves fault detection accuracy under the complex operating conditions of shopping mall elevators, demonstrating the following beneficial effects:

[0041] For multi-source data stream collection Perform environmental calibration and generate multi-dimensional environmental labels Finally, information such as different loads, speeds, and interference intensities is fully integrated into the data system, enabling subsequent recognition algorithms to maintain stability in various noise scenarios.

[0042] Rely on the LTCN model (combining short-term convolution TCN and long-term dependency LSTM) to obtain multi-scale feature vectors The noise enhancement strategy ensures that real fault signs can be captured even in situations such as electromagnetic interference and passenger flow impact;

[0043] Use context-aware dynamic thresholds , and integrate the elevator operation status and environmental index The system adjusts the fault judgment threshold in real time, significantly reducing the false alarm rate under full load or high noise conditions. Furthermore, with the sensor redundancy check mechanism, a false alarm is automatically generated when a single path is abnormal while other sensors are normal, further improving detection reliability.

[0044] Through unified backflow updates for false positives or missed negatives, thresholds and sensor weights can be adaptively revised to ensure long-term system evolution and gradually improve redundant fusion strategies.

[0045] The reinforcement learning and meta-learning introduced can not only iteratively train on existing data, but also quickly migrate or adjust in new scenarios such as cross-season and cross-shopping. On the one hand, it relies on the reward function to comprehensively measure accuracy, maintenance cost and omission risk. On the other hand, it uses the policy network to dynamically adjust the threshold. Online tuning of some LTCN parameters ensures high recognition rates even in environments with sudden increases in air conditioning load or high passenger flow.

[0046] Overall, this solution fully covers the entire chain of elevator fault detection from data-features-detection-feedback-adaptation, and can effectively solve the technical difficulties caused by noise interference, load changes and scene switching, providing strong technical support for real-time safety monitoring and subsequent large-scale deployment of shopping mall elevators. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 The figure is a flow chart of the data anomaly detection method based on multi-scale feature extraction of the present invention. DETAILED DESCRIPTION

[0048] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0049] See also Figure 1 The present invention provides a data anomaly detection method based on multi-scale feature extraction, including:

[0050] Step 1: When the elevator is in different working conditions such as empty, fully loaded or high speed, it automatically calls multi-source sensors, uses basic filtering and time synchronization to filter out spike interference and retain key waveform features, and generates multi-dimensional environmental labels based on environmental information. Then, by entering the temperature, humidity, floor station usage frequency and other situational factors Enrich the data context and form a high-confidence dataset that can be readily adapted to subsequent multi-scale feature extraction;

[0051] The step 1 includes the following:

[0052] Step 101: Environmental basic collection and environmental index extraction

[0053] Install and activate various types of sensors (including acceleration, vibration, current, voltage, temperature and humidity, and noise, etc.), and unify the data output by each sensor to form a preliminary multi-source data stream collection ,in, Indicates the current acquisition time, M indicates the number of sensors;

[0054] During different typical periods of mall operation (e.g., peak and off-peak periods), the aforementioned multi-source data streams are continuously collected and relevant environmental conditions (e.g., passenger flow, electromagnetic interference intensity, indoor temperature and humidity, etc.) are recorded. To facilitate subsequent calibration and analysis, the elevator's operating conditions during the period, such as empty, fully loaded, accelerating, and braking, are recorded.

[0055] Perform necessary processing on the collected data to remove extreme high-frequency noise (to avoid large sensor spikes affecting the overall analysis) and define the environmental index function , The transient and steady-state interference characteristics can be comprehensively captured to measure the noise and disturbance levels under different market conditions. The following definition is introduced:

[0056]

[0057] Where: = , represents the M-channel sensor signals collected at time, such as acceleration, current, and voltage; A subset of weight coefficients corresponding to each sensor (in this solution, they do not conflict with the adaptive coefficients of other steps), which are used to highlight the main noise sources or key fault signals in the environmental index;

[0058] and They represent the translation and scale in continuous wavelet transform respectively;

[0059] and are the lower and upper limits of the translation range, and The lower and upper limits of the scale range;

[0060] It is based on the matrix representation of multi-source signals in the time-frequency domain and can be defined as:

[0061]

[0062] in, Indicates the sensor signal In translation amount = ,scale The multi-component vector obtained after performing wavelet transform (for example, concatenating complex numbers or multi-channel coefficients into a column vector);

[0063] Express Perform matrix logarithm operation, here is the identity matrix (with Same dimension), the purpose is to avoid logarithmic singularity and improve the recognition of small amplitude perturbations;

[0064] is a weighting matrix of the same dimension (which can be a diagonal matrix or a bandpass coupling), used to further emphasize or weaken the influence of a specified frequency band or sensor channel at the matrix logarithmic level;

[0065] It is the matrix trace operator, which can compress the result of matrix logarithm into the scalar domain;

[0066] is the global scaling factor, which is used to amplify or compress the result after the integral calculation;

[0067] When used, it establishes an objective and quantifiable environmental benchmark for subsequent situational awareness and multi-scale fault detection, by setting different weights on different frequency bands and sensors. This allows for adaptively highlighting key interference sources within the complex environment of shopping mall elevators, such as vibration during peak and valley traffic flow, and electromagnetic interference from air conditioning and power equipment switching. By combining frequency-domain analysis with time-domain analysis, both transient noise and steady-state disturbance information can be captured, providing a solid data foundation for more effective separation of environmental noise from true anomalies in subsequent fault feature extraction.

[0068] Step 102: Data synchronization and multi-dimensional environment label generation

[0069] The obtained multi-source data streams and environmental indices Perform time synchronization operations to uniformly timestamp the data of each sensor at the same time or in the same operating cycle, forming a series of elevator operation data and environmental index function Multidimensional records of

[0070] Introducing elevator working status , for example with the symbol Indicates that the elevator is at time The running status (values ​​can be empty, full, accelerating, braking, etc.) of the mall is recorded simultaneously based on the mall operation information, including passenger flow, temperature and humidity, electromagnetic interference sources, and other situational factors. Unified representation of these situational elements ,For example:

[0071]

[0072] Combined with the elevator working status , Environmental Index and situational elements , generating multidimensional context labels for multi-source data records ,like: This label can not only reflect the elevator's own operating conditions, but also effectively distinguish different interference levels and situational information in the environmental dimension;

[0073] Multidimensional environment labels The synchronized data is stored in the environment tag library, providing a directly referenced integrated source of environment context and data input for subsequent multi-scale feature extraction and context-aware anomaly detection steps.

[0074] At this time, the environmental index extracted in step 101 And the multi-dimensional environment label generated in step 102 , are all uniformly saved in the data structure to ensure seamless query and use in subsequent steps;

[0075] When in use, through time synchronization and the establishment of multi-dimensional environment tags, it can ensure that data from different sensors, different frequency bands, and different working conditions can be managed uniformly at the same time. , can be achieved through multi-dimensional environment labels It directly retrieves complete information such as the elevator's operating status, the intensity of environmental disturbances, and passenger flow density, providing accurate contextual support for subsequent steps (such as distinguishing the acceleration difference between no-load and full-load conditions during fault determination). This reduces misjudgments or missed judgments caused by inconsistent data records or incomplete operating condition labeling, greatly improving the accuracy of the system's subsequent fault detection and diagnosis.

[0076] In this step (step 1), a systematic environmental calibration system is established through the progressive processing of steps 101 and 102: on the one hand, the environmental index is used The system characterizes the interference level of multi-source data in different frequency bands. Furthermore, it integrates the actual operating conditions of elevators with contextual factors in shopping malls to manage the collected data using multi-dimensional environmental labels. This environmental calibration system not only provides highly accurate input data and contextual support for subsequent fault detection, but also lays a foundation for applications that can adapt to noise, passenger flow fluctuations, and electromagnetic interference, enabling accurate and reliable elevator fault monitoring and identification in noisy and volatile shopping mall environments.

[0077] Step 2: After completing the environmental calibration and obtaining the multi-source data stream collection With multidimensional environment labels Finally, under different noise levels and with the help of data enhancement, the short-term convolutional neural network (TCN) and long-term memory (LSTM) are used to build an LTCN model. The convolution kernel is used to capture transient impact signals and then stabilize the trend. The multi-scale feature output can improve fault sensitivity.

[0078] The second step includes the following:

[0079] Step 201: Short-term convolution channel and environment weighted fusion

[0080] Aggregate multiple source data streams As the input of the short-term convolution channel (hereinafter referred to as TCN channel), a one-dimensional or two-dimensional convolution kernel (depending on the channel structure of the sensor data) is used to extract features at a short time scale to capture transient fault signals such as impact and collision;

[0081] Define the output of the short-term convolution channel as the short-term feature matrix , where S represents Short-TermFeatures. In each receptive field, the convolution kernel can be combined with hyperparameters such as stride and dilation rate to achieve sensitive capture of abnormal impacts;

[0082] In order to improve the noise suppression ability of short-term convolution, the environmental index and operating status It is regarded as a dynamic weighting factor, which enables the short-term convolution channel output to adaptively adjust the sensitivity to different noise environments;

[0083] Specifically, a fusion function can be defined , the short-term feature matrix Element-level weighting or attention mechanism and environmental index , elevator working status Perform fusion and output the corrected short-term feature vector , its operational form can be written as:

[0084]

[0085] Where: Represents element-wise multiplication For A tensor of all ones of matching shape, used for controllable magnification / scaling on the base value; Represents a smooth nonlinear activation function, such as the Sigmoid function;

[0086] The environmental index amplification factor can be set according to the environmental noise sensitivity and can be between 0.5 and 2.0;

[0087] This is a correction item for the current operating state of the elevator. Different bonuses or reductions can be given to certain states (such as high-speed full load) through table lookup or simple mapping function;

[0088] This fusion process can improve the robustness to environmental disturbances: when the environmental index When the noise level is high (strong), the model (short-term convolution channel (TCN channel) and its environment weighted fusion submodule) automatically reduces the excessive response to transient spikes. When the load is high, the fault trigger sensitivity should be appropriately increased;

[0089] When using, make full use of the environmental index obtained in the first step and elevator working status Dynamically weighting short-term convolutional features can distinguish between noise impacts and real fault impacts: Combined with convolution operations within a short time range, the model captures sudden fault events more precisely, achieving a balance between short-term high sensitivity and noise suppression. The environmental index and working condition labels are explicitly introduced into the weighting mechanism of the convolutional output layer, making the model more adaptable in noisy environments.

[0090] Step 202: Long-term memory integration and noise robustness training

[0091] After completing step 201, the short-term correction characteristics are obtained. ,Next, LSTM (Long Short-Term Memory) is introduced to capture the ,stable operation mode and slowly changing trend of the elevator under ,different loads and speeds;

[0092] The short-term correction characteristics after serialization Input into the LSTM layer to obtain the long-term memory vector , where L represents Long-Term Features. In this process, LSTM can learn the periodic or trend characteristics of elevator operation across multiple time steps and distinguish the difference between normal slow wear and potential fault signs;

[0093] At the same time, the short-term correction characteristics With long-term memory vector Perform splicing or attention weight fusion to obtain the final multi-scale feature vector :

[0094]

[0095] in represents a mapping function (e.g. linear projection or attention fusion) that integrates short-term shock information with long-term trend information;

[0096] Combined with the multi-dimensional environment labels in the first step During the model training phase, noise enhancement is performed on normal operating data, that is, random superposition of interference fragments from actual measurements or simulations, so that the model can still maintain the stability of feature extraction in high-noise scenarios;

[0097] In order to further improve the model's sensitivity to real faults, a multi-objective loss function is defined , which enables the model to perform weighted learning in high-noise scenarios and reduces the risk of overfitting low-noise samples. Represents all learnable parameters in this step (including convolution kernels, LSTM weights, fusion parameters, etc.), as follows:

[0098]

[0099] Where: is the reference feature or label vector (which can come from the statistical characteristics of normal samples, or from the difference labels of known fault samples); is a weight term related to the environmental index or working condition label, which is used to distinguish between high noise and low noise situations during training, such as Or based on the elevator working status Set different values ​​for the status categories;

[0100] Represents the square of the Euclidean distance, which is combined with the logarithmic function to produce an optimization gradient that is more sensitive to extreme errors;

[0101] When used, LSTM's characterization of long-term trends is combined with the short-term convolution channel TCN's capture of short-term shocks, effectively distinguishing between transient random fluctuations and trend-based fault evolution; the noise enhancement strategy and environmental weighted training enable the model to maintain its ability to extract real fault features even in the face of extreme interference (such as strong electromagnetic noise generated by short-term switching of electrical equipment in shopping malls), combining multi-scale convolution with LSTM in a unified training framework and using environmental label weights Change the attention of the loss function to different noise scenes.

[0102] By building a multi-scale feature extraction model that can simultaneously capture short-term instantaneous shocks and long-term stable trends, and improving overall robustness through noise enhancement and environment weighted training strategies, the short-term convolution channel in step 201 first combines the environmental index and the operating status Dynamic weighting is performed to enable transient fault signals to be captured even in noisy backgrounds. The LSTM module in step 202 models long-term changes under different load and speed conditions, and uses a multi-objective loss function to highlight the ability to distinguish between high-noise and real fault samples. The multi-scale features output after the fusion of the two are This not only offers significant advantages in noise suppression but also provides high feature confidence and discrimination for subsequent context-aware anomaly detection. In the noisy and volatile environment of shopping mall elevators, this method truly achieves short-term sensitivity, long-term robustness, and environmental adaptability in fault feature extraction.

[0103] Step 3: When outputting multi-scale feature vectors And the elevator is already in working condition and environmental index Afterwards, the multi-scale feature vector With reference vector Use difference Evaluate anomaly score and calculate threshold function ,If the anomaly score exceeds the limit and the multi-sensor channels are redundant and consistent, it is determined to be a fault and an alarm is output;

[0104] The step three includes the following:

[0105] Step 301: Dynamic threshold setting and situational awareness anomaly identification

[0106] Take the multi-scale feature vector output in the second step , to measure the moment The potential fault deviation degree, compared with the reference vector (e.g. statistical benchmarks of historical normal operating condition characteristics, or projections of normal state subspaces marked by maintenance personnel) to perform difference calculations;

[0107] Define the anomaly score function To measure the multi-scale feature vector With reference vector To ensure sufficient resolution in a high noise environment, the formula can be written as:

[0108]

[0109] in: express -norm, optional (like is the Euclidean distance) to cope with multi-dimensional features;

[0110] It is the amplification exponent, which can take values ​​between 1.5 and 3 to enhance the sensitivity to extreme deviations;

[0111] It is learned in the initial stage or from historical normal data and can be updated during deployment or periodic maintenance.

[0112] In order to dynamically adjust the alarm sensitivity under different working conditions and environmental interference, a threshold function that changes with the scenario is introduced. , combined with the elevator operation status defined in the first step and environmental index , which can take the following forms:

[0113]

[0114] Where: is the benchmark threshold mapping function, Used to select or interpolate a reference threshold value based on the current elevator operating status (such as no load, full load, high speed, etc.);

[0115] is a smooth nonlinear activation function, such as the Sigmoid function, which is used to convert the environmental index Mapping to a domain suitable for amplification or attenuation; To adjust the coefficient of the environmental index influence, the value can be between 0.5 and 1.5;

[0116] Indicates the overall scaling factor, which can be between 0.1 and 0.3 and is used to control the threshold value level;

[0117] When the environmental index When it is higher, The elevator status will also rise, thereby increasing the threshold to avoid a large number of false alarms; Indicates low speed or no load, It may be too small, making the model more sensitive to capture minor fault signs;

[0118] Finally at the moment , according to the anomaly score function With threshold function Comparison is used to determine whether a fault is suspected;

[0119] like , it is determined to be a suspected fault, otherwise it is considered normal or noise fluctuation;

[0120] When in use, the threshold value is dynamically raised under high noise or high load conditions, effectively reducing frequent false alarms caused by normal fluctuations. In low noise or special sensitive conditions, the threshold value is automatically tightened to avoid missing potential hidden dangers by comparing with the reference vector The difference measurement can realize the intuitive quantification of fault tendency and can adapt to the environment. By forming an explicit anomaly score-dynamic threshold mechanism and using logarithmic and exponential construction, it can achieve good discrimination in different amplitude ranges.

[0121] Step 302: Multi-sensor redundancy check and comprehensive judgment

[0122] In addition to using the multi-scale feature vector output by the second step fusion In addition to detection, the local anomaly scores of each sensor at different scales can be retained, for example, the short-term convolutional channel (TCN) unit or LSTM output is split into sensor components, which are recorded as , ; For each sensor , defining the single point anomaly score Get a set of score vectors :

[0123]

[0124] In this way, when a suspicious fluctuation is detected at a certain point, the consistency can be judged by comparing the scores of other sensors; Calculate the absolute deviation between the observed value and the predicted (or reference) value, amplify and logarithmically smooth it, and get the single point anomaly score; it is the logarithmic score vector Perform multi-source consistency detection and introduce matrix measurement To characterize the overall coupling degree of all sensor anomaly scores:

[0125]

[0126] Where: It is the outer product matrix, which can reflect the correlation and amplitude between different sensor scores;

[0127] is the global scaling factor, which can be between 0.1 and 2;

[0128] It is the matrix trace operator, which can compress the result of matrix logarithm into the scalar domain for comparison with the threshold;

[0129] If the matrix metric If it is significantly greater than a set trigger threshold, it means that multiple sensors have generated high anomaly scores and the credibility increases accordingly; if only a single sensor has a high score while the others are low, it means that the matrix metric The trigger threshold is difficult to reach, thus reducing the probability of false alarms.

[0130] After determining the suspected fault, the score vector Perform multi-source consistency check, if the matrix metric If the redundant fusion threshold is exceeded, it is determined to be a multi-source confirmed fault, triggering an alarm or entering adaptive feedback and disposal: otherwise, the anomaly is regarded as a non-real fault situation such as noise interference or single-source sensor failure; the above judgment results can be recorded in the subsequent adaptive feedback stage to update the reliability weight or failure mode library of each sensor.

[0131] When used, by performing outer product matrix logarithm operation on the single sensor anomaly score and taking the trace, the coupling characteristics of multiple sources with high anomalies at the same time can be captured, effectively reducing the false alarm caused by single source channel noise or fault illusion; the threshold function of step 301 is Combining this process with multi-source redundancy verification ensures both sensitivity and reliability, adapting to varying operating conditions while significantly reducing the risk of missed and false alarms. Utilizing context-aware threshold adjustment and multi-sensor redundancy verification, a flexible and highly reliable fault anomaly determination process has been established. This process adapts to varying passenger loads and environmental interference conditions, significantly reducing the false alarm and missed alarm rates through multi-source mutual verification, providing greater reliability for the safe diagnosis of elevator faults in shopping malls.

[0132] Step 4: When the suspected or confirmed fault result is output and the real operation and maintenance feedback is obtained, the multi-source data stream collection is recovered , multi-scale feature vector and missed and false positive samples to the training library, and revised the dynamic threshold The multi-scale model parameters Ω are incrementally updated with the sensor weight w, and the fault detection accuracy is improved iteratively to suppress frequent false alarms caused by noise fluctuations.

[0133] The step 4 includes the following contents:

[0134] Step 401: Fault determination result backflow and data re-marking

[0135] Match the suspected faults generated in the third step with the labels of the multi-source confirmed faults, as well as the actual status feedback from subsequent manual inspections or maintenance systems, to generate a judgment result table.

[0136] If the system determines that it is a fault and manual or maintenance personnel confirm that it is a fault, it will be recorded as a correct error report, which is the elevator fault detection and diagnosis system;

[0137] If the system determines that there is a fault, but there is no abnormality, it will be recorded as a false alarm;

[0138] If the system is judged to be normal, but there is actually a fault, it is recorded as a missed alarm;

[0139] Other cases are recorded as normal or suspected fluctuations;

[0140] For samples that are confirmed to be false positives or missed negatives, the multi-source environment perception training dataset (including the multi-source data stream set) used in the first and second steps should be , multi-scale feature vector and multi-dimensional environmental labels ) and supplemented with corresponding real fault labels or normal labels to form updated training or calibration data;

[0141] Keep samples of correct error reports or normal judgments to continue expanding the boundary information between normal and faulty conditions; and conduct a review of feasible sources. If false alarms or missed alarms are too concentrated in certain specific sensor data segments, the hardware or software sampling process of the sensor can be reviewed at this step to eliminate judgment errors caused by sensor failure or serious deviations. If it is found that false alarms are mostly concentrated in a certain type of operating condition (such as full load and high-speed operation), it indicates that the model or threshold setting under this operating condition scenario is insufficient, and the subsequent steps need to focus on optimization of this scenario.

[0142] When in use, by comparing the actual fault information with the system judgment results, the limitations of the current model can be discovered in a timely manner, and the new information can be fed back into the data set for re-labeling, laying the foundation for subsequent threshold adjustment and model training. A dynamic update mechanism for the fault case library is established, which can continuously enrich the environment-fault control samples in a real shopping mall environment, and also has the ability to continuously learn for rare working conditions that are difficult to simulate.

[0143] Step 402: Dynamic threshold and sensor weight update

[0144] After collecting the latest correct error, false alarm and missed alarm information, the threshold function used in the third step is Make adjustments, for example, introducing differentiable learning rates based on historical accuracy feedback , define a new threshold update relationship:

[0145]

[0146] in: is the dynamic threshold; To express the error metric of this judgment result (for example, the correct error can be expressed as If it is less than a certain threshold, false alarm or missed alarm will be larger); is the threshold correction rate at this moment, based on the multi-dimensional environment label (such as passenger flow, electromagnetic interference intensity) and the accuracy of recent judgments are dynamically determined;

[0147] is the offset constant, which can be between 0,1 and 0.5 and is used to stabilize the correction process;

[0148] When the judgment results have large deviations multiple times ( exceeds a certain threshold), will deviate significantly from 1, forcing Make adjustments in the corresponding direction (raise or lower);

[0149] In the process of multi-source sensor fusion, if a single sensor channel frequently conflicts with the judgment of other channels, the corresponding fusion weight of the channel in the redundancy check is reduced accordingly; in the specific implementation, a weight vector can be defined , and calculate the outer product matrix The process of combining, such as:

[0150]

[0151] Where: Represents the abnormal score vector for each sensor Perform weighted adjustments;

[0152] is the identity matrix; For the moment Sensor weight After adjustment, the measurement results of the coupling strength of multi-sensor anomaly scores are is the global scaling factor, which can be between 0.1 and 2; For the Road sensor at all times Single point anomaly score; is the matrix trace operator;

[0153] When the false alarm rate of a sensor is high, that is, higher than the expected level, the corresponding weight is lowered. If its contribution to fault judgment is reliable, the corresponding weight is increased. ; This enables differentiated adjustment of multi-source redundant verification, and maintains a high overall accuracy even in scenarios with significant noise interference.

[0154] When used, the dynamic threshold correction mechanism can quickly compensate for false positives or missed positives, allowing the system to find a balance between fewer false positives and fewer missed positives in continuous iterations; sensor weight redistribution can prevent overall misjudgments caused by software and hardware defects or frequent abnormal interference in a certain channel, and can also ensure that the most reliable data source is highlighted during multi-source fusion, thereby improving diagnostic accuracy. A two-pronged update of the weight vector w in the logarithm of the redundant check matrix is ​​performed to build an adaptive and multi-dimensional tuning mechanism for the elevator fault detection system.

[0155] Step 403: Multi-scale model parameter retraining and collaborative decision output

[0156] The multi-scale model LTCN constructed in the second step refers to a composite network of short-term (TCN) and long-term (LSTM) in parallel or in series. The fault samples and normal samples re-labeled in step 401 are used for incremental or periodic retraining to update the convolution kernel, LSTM weights and other learnable parameters. ;

[0157] During the update, a loss function with a penalty term can be defined to highlight the extreme concern for false positive and false negative scenarios, for example:

[0158]

[0159] in: Indicates that the parameter Next, the model predicts the characteristic value of the input signal at the moment;

[0160] According to the real features or expected output after re-labeling;

[0161] express -norm, is the amplification index, which can be between 2.0 and 4.0 and is used to strengthen the penalty for significant deviations;

[0162] It is a weighting coefficient, which can be between 1.0 and 10. It gives higher weight to false positive / missing samples, so that they can generate larger gradient updates during training. Represents a small batch of training samples;

[0163] Through small batches ( ) or update the model iteratively online so that it continuously approaches the optimal state under the guidance of new environmental data and fault cases.

[0164] The dynamic threshold function updated in step 402 and the sensor weight vector And the retrained model parameters here The three are packaged together in the decision-making layer so that the latest configuration can be immediately adopted in the next round of testing;

[0165] If there are structural changes in the shopping mall environment (such as the addition of a large number of large equipment, seasonal changes in customer flow patterns, etc.), the model's adaptability to the new operating conditions can be ensured through regular batch retraining or online fine-tuning. The collaborative update results will continue to be recycled in the next round of monitoring-detection-feedback-update closed loop, forming a continuously evolving fault detection system.

[0166] When in use, the recognition capability and threshold strategy are coordinated and improved by training the parameters of the multi-scale model (TCN, etc.). When there are new changes in the external environment and elevator operation mode, this mechanism can enable the model to learn the new data distribution and fault symptom characteristics in a timely manner, avoiding continuous misjudgment caused by outdated models.

[0167] Through the triple linkage of threshold + sensor fusion weight + deep model parameters, the system can evolve collaboratively in different dimensions to maximize the detection stability and safety under the changing working conditions of shopping mall elevators, enabling the system to continuously optimize the sensitivity and robustness of fault identification while constantly receiving real operation and maintenance information and environmental changes.

[0168] Combining the multi-source data collection, noise-robust feature extraction, and context-aware anomaly detection built in the previous three steps, this fourth-step closed-loop update mechanism can minimize the probability of false alarms and missed alarms, ensuring that shopping mall elevators maintain a high level of safe operation and fault prevention in noisy and changing environments.

[0169] Step 5: After completing multiple rounds of feedback updates, when facing sudden changes in the new season or new market environment, the reinforcement learning or meta-learning module describes the state according to the state. With the reward function , using the threshold strategy function Coordination determination threshold Scheduling and combining model migration strategies , achieving adaptive fault detection across loads and noises and shortening the retraining cycle;

[0170] The step five includes the following:

[0171] Step 501: Scene state representation and reward mechanism definition

[0172] Multidimensional environment labels that will be dynamically updated (Including elevator status , Environmental Index , as well as external passenger flow, electromagnetic interference intensity, etc.) and current threshold configuration , sensor weight vector , model parameter key index, etc. are combined into a state description ;

[0173] The status description It is the input of reinforcement learning or meta-learning algorithms, used to characterize the current environmental context of the system and its own detection configuration; in order to guide the system to spontaneously learn good thresholds and model migration strategies, a multi-dimensional reward function needs to be established. , combined with multiple factors such as accuracy indicators obtained after real-time operation, false positive / missing negative distribution, and maintenance costs:

[0174]

[0175] Where: represents the comprehensive accuracy measure of fault detection in the current period; and represent the missed alarm rate and the false alarm rate respectively;

[0176] Indicates the cost or expense of this scene adaptation process (such as additional resource consumption caused by excessively frequent adjustment of thresholds and sensor weightings); is a positive real number weight used to balance accuracy, false positives / missing negatives, and system overhead;

[0177] In the reinforcement learning framework, each cycle (such as half an hour, one hour, or other suitable windows) can be regarded as a decision moment. With the reward function Select or update the threshold strategy or model migration method for the next decision moment; under the meta-learning framework, the learning experience of similar operating scenarios in historical operating data can be used to achieve rapid adaptation to new scenarios or new seasons, as follows:

[0178] First, the current scene feature vector Input policy network (including environment labels, current thresholds, sensor weights, and recent performance indicators) , which outputs two actions: one is the continuous threshold correction , the second is discrete model update instructions ("no update", "fine-tune" or "retrain"); after executing these actions, the system calculates the immediate reward in the next cycle based on the detection accuracy, false positive / missing rate and update cost , and then use the policy gradient algorithm to Iterative optimization is performed to achieve adaptive joint decision-making of threshold and model migration.

[0179] When used, the environment state and adaptive configuration are combined into a unified state description , which can more precisely identify the current noise level and working condition characteristics in a changing shopping mall situation, and introduce the reward function Finally, while seeking high accuracy, we can also take into account the false alarm rate, missed alarm rate and cost constraints to maximize the overall benefit.

[0180] Step 502: Joint learning of threshold strategy and model migration strategy

[0181] The dynamic threshold value obtained Based on the introduction of threshold strategy function (in represents a learnable parameter), according to the current state description Automatically output a set of control variables (such as threshold correction terms, sensor fusion fine-tuning coefficients, etc.);

[0182] The updated threshold is , which can be expressed as:

[0183]

[0184] in: is a smooth mapping to avoid excessive changes in the threshold (saturation function, nonlinear activation, etc. can be used); when the threshold strategy function Output positive adjustment amount when the threshold It will be raised or lowered accordingly to adapt to the current or upcoming scene noise level;

[0185] Make migration decisions for the multi-scale feature extraction model (including convolution + LSTM or LTCN structure) formed in the second and fourth steps, and define the model migration strategy under the meta-learning or reinforcement learning framework ,in It is its internal parameter, which is used to determine what kind of parameter partial freezing or local retraining is done in what situation;

[0186] For example: If the current state describes Reflecting the new strong seasonal characteristics (air conditioning overload), the model migration strategy It may output a high-level retraining request to make targeted fine-tuning on certain channels of the convolutional layer; if there is only a small fluctuation, it will output a decision to retain the existing parameters or do slight online learning; the parameters refer to: the weight parameters of the multi-scale feature extraction model (i.e. all learned parameters of deep networks such as convolution kernels and LSTM units), parameters of the policy network (Threshold Adjustment Strategy) and (Model migration strategy);

[0187] Threshold strategy function Model migration strategy The two strategies are encapsulated in a unified reinforcement learning or meta-learning update loop, and the rewards obtained through interaction Perform gradient optimization:

[0188]

[0189] Where: To review historical periods or mini-batches; is the time discount factor or meta-learning step size; is the learning rate; is a multi-dimensional reward function;

[0190] Through this update rule, continuous iteration can be achieved in time series or across scenarios, allowing the threshold strategy and model migration strategy to cooperate with each other and continuously approach the global optimal solution.

[0191] Integrating threshold adjustment and model retraining into the same decision-making framework significantly reduces the limitations of relying solely on fixed threshold functions or offline training. The system can quickly detect performance degradation and autonomously respond to seasonal or scenario changes. By combining time discounting with meta-learning rules, it achieves a flexible balance between short-term benefits and long-term adaptability, providing greater flexibility for elevator fault detection in complex shopping mall environments.

[0192] Step 503: Online execution and cross-scenario adaptive output

[0193] The threshold strategy function obtained in step 502 Model migration strategy In the elevator system, it is online in real time. Whenever a new operation cycle arrives, the current state description is used. To decide whether to revise the threshold and / or update the local model; at the same time, the fault detection effect (accuracy, missed and false positives) after each decision is recorded and a new reward function is calculated Continuously enrich learning samples; represents the dynamic fault judgment threshold after being modified by the reinforcement learning (RL) strategy;

[0194] When the system faces different shopping malls or seasonal switching, the existing threshold adjustment strategy can be and model migration strategies Migrate as a meta-model: In new scenarios, only a small amount of online tuning or small-scale training is needed to achieve usable threshold configuration and feature extraction capabilities. For completely unfamiliar extreme scenarios, certain security policies can be retained (such as increasing alarm redundancy) to minimize the risk of missed alerts, and gradually converge to more reasonable thresholds through learning. During execution, if a high number of false positives or missed alerts persists, maintenance personnel will be alerted or the detailed feedback and collaborative update process in step 4 will be triggered again.

[0195] When in use, the online execution mechanism ensures that the system can make timely adjustments to thresholds and model strategies based on the latest situation during the actual operation of the elevator, shortening the feedback-to-improvement delay. Through the meta-model migration method, it can be applied to different shopping malls or seasons with one click without training from scratch, further improving the versatility of the detection solution; combining real-time online execution with cross-scenario migration to achieve maximum adaptive effect with minimal human intervention, the elevator fault detection system is transformed from passive update to active learning in the IoT application of shopping mall elevators, thereby effectively improving overall safety and operation and maintenance efficiency, and has the potential to be extended to other complex electromechanical equipment monitoring scenarios.

[0196] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0197] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0198] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is only for some logical functions. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0199] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0200] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A data anomaly detection method based on multi-scale feature extraction, characterized by: include, When the elevator is in different operating conditions, it automatically calls multi-source sensor data, adopts a unified order benchmark for filtering and time synchronization, combines passenger flow and situational factors to generate multi-dimensional environmental labels, and constructs a high-confidence dataset for subsequent multi-scale analysis; After preliminary cleaning of multi-source sensor data, the LTCN structure is enabled in multiple noise enhancement scenarios. Short-term convolutional network (TCN) is used to capture transient impacts, and long-term dependencies are then extracted using LSTM. This outputs multi-scale features and enhances fault detection sensitivity. After the multi-scale features are generated and the elevator operating conditions and environmental indexes are known, a dynamic threshold is constructed based on the difference. Through multi-sensor redundancy verification, alarms are triggered only for multi-source consistent anomalies, reducing false alarms caused by noise interference. If a suspicious result is output and true feedback is obtained, the multi-source sensor data and multi-scale features are back-traced to perform false positive and negative sample re-labeling, incrementally revise the judgment threshold and sensor weight, and update the multi-scale model; When the environment changes seasonally or cross-shopping deployment needs arise, reinforcement learning and meta-learning mechanisms are enabled to adaptively modify the judgment threshold and migrate some LTCN parameters based on the current state description and reward function; After collecting the latest correct error, false alarm and missed alarm information, the threshold function is updated; If a single sensor channel frequently conflicts with the judgments of other channels, the corresponding fusion weight of the channel in the redundancy check will be reduced accordingly; when the false alarm rate of a sensor is higher than expected, the corresponding weight will be lowered; if its contribution to fault judgment is reliable, the corresponding weight will be increased; For the multi-scale LTCN model, incremental or periodic retraining is performed using relabeled fault samples and normal samples to update the learnable parameters; During the update process, a loss function with a penalty term is defined, and the multi-scale model is iteratively updated in small batches or online, so that it continuously approaches the optimal state under the guidance of new environment data and failure cases. The updated dynamic threshold function, sensor weight vector, and retrained model parameters are packaged into the decision layer so that the latest configuration can be immediately adopted in the next round of detection.

2. The data anomaly detection method based on multi-scale feature extraction according to claim 1, characterized in that: Install and activate various types of sensors to continuously collect multi-source data streams at different typical time periods to form a multi-source data stream collection, recording the elevator operating conditions and related environmental conditions during the period; pre-process the collected data and construct an environmental index function to measure the noise and disturbance levels under different market conditions; Time synchronization is performed on the acquired multi-source data streams and environmental indices to form a multi-dimensional record containing elevator operation data and environmental index functions. The elevator operating status is introduced and contextual elements are generated from the mall operation information. Combining elevator working status, environmental index and situational factors, multi-dimensional environmental labels are generated for multi-source data records.

3. The data anomaly detection method based on multi-scale feature extraction according to claim 2, characterized in that: The multi-source data stream collection is used as the input of the short-term convolution channel to extract the transient fault signal at a short time scale; The output of the short-term convolution channel is defined as a short-term feature matrix. The short-term feature matrix is ​​fused with the environmental index and the elevator working status through element-level weighting or attention mechanism to output the corrected short-term feature vector.

4. The data anomaly detection method based on multi-scale feature extraction according to claim 3 is characterized in that: The serialized short-term correction features are fed into the LSTM layer to obtain the long-term memory vector. At the same time, the short-term correction features are concatenated or fused with the long-term memory vector by attention weights to obtain the final multi-scale feature vector. Combined with multi-dimensional environmental labels, noise enhancement is performed on normal operating condition data during the model training phase, and a multi-objective loss function is defined to enable the model to perform weighted learning in high-noise scenarios and reduce the risk of overfitting low-noise samples.

5. The data anomaly detection method based on multi-scale feature extraction according to claim 4 is characterized in that: In order to dynamically adjust the alarm sensitivity under different working conditions and environmental interference, a scenario-dependent threshold function is introduced in combination with the elevator operating status and environmental index. An anomaly score function is constructed to measure the difference between the multi-scale feature vector and the reference vector to measure the degree of potential fault deviation. The triggering of a suspected fault is determined by comparing the anomaly score function with the threshold function. If the anomaly score function is greater than the threshold function, it is determined as a suspected fault; otherwise, it is considered normal or noise fluctuation.

6. The data anomaly detection method based on multi-scale feature extraction according to claim 5, characterized in that: A single-point anomaly score is defined for each sensor to obtain a score vector. After identifying suspected faults, a multi-source consistency check is performed on the score vector, and a matrix metric is introduced to characterize the overall coupling degree of all sensor anomaly scores. If the matrix metric exceeds the redundant fusion threshold, it is determined to be a multi-source confirmation fault, triggering an alarm or entering adaptive feedback and disposal. Otherwise, the anomaly is regarded as a non-real fault situation such as noise interference or single-source sensor failure.

7. The data anomaly detection method based on multi-scale feature extraction according to claim 6, characterized in that: Match the labels of suspected faults and multi-source confirmed faults with the actual status feedback to generate a judgment result table, including correct errors, false alarms and missed alarms, and normal or suspected fluctuation records; samples with confirmed false alarms or missed alarms are re-labeled and supplemented with corresponding real fault labels or normal labels to form updated training or calibration data; samples with correct errors or normal judgments are retained, and a feasible source review is carried out.

8. The data anomaly detection method based on multi-scale feature extraction according to claim 7, characterized in that: The dynamically updated multi-dimensional environmental label is combined with the current threshold configuration, sensor weight vector, and key index of model parameters to form a state description, which serves as the input to the reinforcement learning or meta-learning algorithm and establishes a multi-dimensional reward function. In the reinforcement learning framework, the threshold strategy or model migration method for the next decision moment is selected or updated based on the current state description and reward function. The migration decision is made for the multi-scale feature extraction model, and the model migration strategy is defined within the meta-learning or reinforcement learning framework. The threshold policy function and the model transfer strategy are encapsulated in a unified reinforcement learning or meta-learning update loop, and gradient optimization is performed through the reward function obtained through interaction.

9. The data anomaly detection method based on multi-scale feature extraction according to claim 8, characterized in that: The updated threshold policy function and model migration strategy are put into real-time operation in the elevator system. Whenever a new operating cycle arrives, the current state description is used to decide whether to revise the threshold and / or update the local model, and the fault detection effect after each decision is recorded. When facing different shopping malls or seasonal switching, the existing threshold adjustment strategy and model migration strategy are migrated as a meta-model. If there is a persistent high number of false positives or missed negatives, an external reminder will be issued or feedback and collaborative updates will be triggered.

Images