Data anomaly detection method based on multi-scale feature extraction

Abstract

The invention discloses a data anomaly detection method based on multi-scale feature extraction, relates to the technical field of data anomaly detection, and sequentially performs multi-source data acquisition and environment calibration, noise robustness multi-scale feature extraction, context awareness anomaly detection and sensor redundancy check, and adaptive feedback and collaborative update decision. Intelligent threshold learning and scene self-adaption are realized; the method comprises the following steps: firstly, acquiring and marking acceleration, current, electromagnetic and other multi-path signals, and constructing an environment label; then, utilizing a TCN and LSTM fusion model to extract short-term impact and long-term trend; according to the elevator working condition and the noise index, the threshold value is dynamically adjusted, and the fault is subjected to redundancy check; the model, the sensor weight and the threshold value are corrected through false alarm and missing alarm backflow; and finally, a monitoring-detection-feedback-updating-self-adaptive closed loop is constructed by means of reinforcement learning or meta learning cross-seasonal adaptation, high precision, low false alarm and high robustness are achieved, and elevator operation safety and maintenance efficiency are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data anomaly detection, specifically 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 operating data of devices and systems are crucial for ensuring safety and efficiency. Facing multi-source heterogeneous data, high-noise interference, and dynamically changing operating conditions, traditional single data processing methods and fixed-threshold detection technologies are difficult to effectively cope with, resulting in frequent false alarms or missed detections. 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 still face challenges such as high environmental noise, insufficient multi-source data fusion, and limited online update capabilities. How to comprehensively analyze multi-source sensor signals using advanced data processing and intelligent algorithms and still achieve stable and accurate fault identification in a high-interference environment has become one of the urgent problems in the current industry.

[0003] In recent years, with the development of artificial intelligence and industrial Internet of Things technologies, fault diagnosis solutions based on deep learning models such as convolutional networks and recurrent neural networks have gradually emerged. Among them, some studies use one-dimensional convolution technology to reduce the dimension of time-domain signals and combine time convolutional networks (TCNs) or neural ordinary differential equations (NeuralODEs) to extract local fault features of devices; some teams also introduce self-attention mechanisms to capture relevant information within a global range. Such methods have improved the accuracy of fault detection for short-time series signals. However, in actual deployment, they still face the following challenges: Large environmental noise and load fluctuations: When a device is simultaneously affected by electromagnetic interference, mechanical vibration, temperature and humidity changes, etc., the sensor output signal often shows highly random and sudden fluctuations, which can easily cause the algorithm to misjudge high disturbances under normal operating conditions as faults, or to submerge real fault features in noise, resulting in missed detections; Lack of multi-source data fusion and adaptability: Device monitoring usually involves multiple sensor channels, but traditional methods often process data from each channel independently or only perform simple weighting; There is a lack of in-depth modeling of the correlation between sensors and the ability to automatically adjust the fusion strategy according to the working conditions. Especially when the load suddenly switches or the operating mode changes frequently, fixed fusion schemes often fail; Insufficient online update and continuous learning: Some intelligent fault detection methods are often trained on an offline large dataset and then put into use in the field, making it difficult to track the long-term evolution of device and environmental characteristics in a timely manner. If there is a lack of online learning or adaptive update after deployment, once encountering new interferences or new load patterns outside the distribution range of the training set, the detection accuracy will significantly decrease.

[0004] 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 on key electromechanical equipment in actual places where the noise level fluctuates greatly with time periods and load conditions (such as shopping malls, airports, industrial production lines, etc.); through technical means such as dynamic threshold adjustment, sensor redundancy verification, short-term and long-term feature coupling, and online adaptive learning, solve the detection difficulties caused by environmental noise and unstable load.

[0005] This method is not only applicable to certain vertical transportation equipment (such as elevators in large shopping malls), but also can be widely applied to 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

[0006] (I) Technical problems to be solved Aiming at the deficiencies of the prior art, the present invention provides a data anomaly detection method based on multi-scale feature extraction. By first acquiring and annotating multi-channel signals such as acceleration, current, and electromagnetic, an environmental label is constructed; then a TCN and LSTM fusion model is used to extract short-term shocks and long-term trends; then the threshold is dynamically adjusted according to the elevator working conditions and noise index, and the faults are redundantly verified; subsequently, the false alarm and missed alarm return correction model, sensor weights and thresholds are adjusted; finally, relying on reinforcement learning or meta-learning for cross-season adaptation, a monitoring-detection-feedback-update-adaptive closed loop is constructed, significantly improving the elevator operation safety and maintenance efficiency; thus solving the technical problems recorded in the background art.

[0007] (II) Technical solutions To achieve the above objectives, the present invention is realized through the following technical solutions: A data anomaly detection method based on multi-scale feature extraction, including, when the elevator is in different working conditions, automatically calling multi-source sensor data, using filtering and time synchronization means to unify the sequence reference, generating a multi-dimensional environmental label in combination with passenger flow and scenario factors, and constructing a high-confidence dataset for subsequent multi-scale analysis; After initially cleaning the multi-source sensor data, the LTCN structure is enabled in a multi-noise enhancement scenario. First, the short-term convolutional TCN is used to capture transient shocks, and then the LSTM is used to refine long-term dependencies, outputting multi-scale features and enhancing the sensitivity of fault detection; When multi-scale features are generated and the elevator working conditions and environmental index are obtained, a dynamic threshold is constructed based on the difference determination, and only multi-source consistent anomalies are triggered for alarm through multi-sensor redundancy verification, reducing false alarms caused by noise interference; If a suspicious result is output and real feedback is obtained, the multi-source sensor data and multi-scale features are traced back to re-label the false alarm and missed alarm samples, incrementally revise the decision threshold and sensor weights, and update the multi-scale model; When the environmental season suddenly changes or the need for cross-mall deployment arises, enable the reinforcement learning and meta-learning mechanisms to adaptively correct the decision threshold according to the current state description and reward function and migrate some parameters of the LTCN.

[0008] Furthermore, 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 set, record the elevator operation conditions and related environmental states during the time period, and preprocess the collected data to construct an environmental index function to measure the noise and disturbance levels under different market conditions.

[0009] Furthermore, synchronize the obtained multi-source data streams and environmental indices in time to form a multi-dimensional record containing elevator operation data and environmental index functions, introduce the elevator working state, and generate context elements from the mall operation information; Combine the elevator working state, environmental index, and context elements to generate multi-dimensional environmental labels for the multi-source data records.

[0010] Furthermore, use the multi-source data stream set as the input of the short-term convolution channel and extract transient fault signals at short time scales; define the output of the short-term convolution channel as the short-term feature matrix, and fuse the short-term feature matrix with the environmental index and elevator working state through element-wise weighting or attention mechanism to output the corrected short-term feature vector.

[0011] Furthermore, input the serialized short-term corrected features into the LSTM layer to obtain long-term memory vectors, and at the same time splice or fuse the short-term corrected features and long-term memory vectors with attention weights to obtain the final multi-scale feature vector: Combine the multi-dimensional environmental labels, perform noise enhancement on the normal condition data during the model training stage, and define a multi-objective loss function to enable the model to perform weighted learning in high-noise scenarios and reduce the risk of overfitting to low-noise samples.

[0012] Furthermore, to dynamically adjust the alarm sensitivity under different working conditions and environmental interferences, combine the elevator operation state and environmental index, introduce a threshold function that changes with the scenario, and construct an anomaly score function to measure the difference between the multi-scale feature vector and the reference vector to measure the degree of deviation of potential faults; Determine whether to trigger a fault suspicion based on the comparison between the anomaly score function and the threshold function. If the anomaly score function the threshold function, it is determined as a suspected fault, otherwise it is regarded as normal or noise fluctuation.

[0013] Furthermore, define a single-point anomaly score for each sensor to obtain a score vector. After determining a suspected fault, perform multi-source consistency detection on the score vector, and introduce matrix metrics to characterize the overall coupling degree of the anomaly scores of all sensors; If the matrix metric exceeds the redundant fusion threshold, it is determined as a multi-source confirmed fault, triggering an alarm or entering adaptive feedback and handling. Otherwise, this anomaly is regarded as a non-genuine fault situation such as noise interference or single-source sensor failure.

[0014] Furthermore, match the labels of the suspicious faults and multi-source confirmed faults with the real status feedback to generate a decision result table, including correct error reporting, false alarm, and missed alarm, as well as normal or suspected fluctuation records. For the samples with confirmed false alarms or missed alarms, re-label them and supplement the corresponding real fault marks or normal marks to form updated training or calibration data. Retain the samples with correct error reporting or normal determination and conduct a feasible source review.

[0015] Furthermore, after collecting the latest correct error reporting, false alarm, and missed alarm information, update the threshold function. If a single sensor channel frequently conflicts with the determinations of other channels, correspondingly reduce the fusion weight corresponding to this channel in the redundant check. When the false alarm rate of a certain sensor is higher than expected, lower the corresponding weight. If its contribution to fault judgment is highly reliable, increase the corresponding weight.

[0016] Furthermore, for the multi-scale model LTCN, use the re-labeled fault samples and normal samples for incremental or periodic retraining to update the learnable parameters. When updating, define a loss function with a penalty term, and iteratively update the multi-scale model in a mini-batch or online manner, making it continuously approach the optimal state under the guidance of new environmental data and fault cases. Package the updated dynamic threshold function, the sensor weight vector, and the re-trained model parameters here into the decision layer so that the latest configuration can be immediately adopted in the next round of detection.

[0017] Furthermore, combine the dynamically updated multi-dimensional environmental labels with the current threshold configuration, the sensor weight vector, and the key indexes of the model parameters as the state description, use it as the input of the reinforcement learning or meta-learning algorithm, and set up a multi-dimensional reward function. In the reinforcement learning framework, select or update the threshold strategy or model migration method at the next decision-making moment according to the current state description and the reward function.

[0018] Furthermore, make a migration decision for the multi-scale feature extraction model and define the model migration strategy in the meta-learning or reinforcement learning framework. Package the threshold strategy function and the model migration strategy into a unified reinforcement learning or meta-learning update loop, and perform gradient optimization through the obtained reward function of the interaction.

[0019] Furthermore, the updated threshold policy function and the model migration policy are put into operation in the elevator system in real time. Whenever a new operation cycle arrives, it is decided whether to revise the threshold and / or update the local model based on the current state description, and the fault detection effect after each decision is recorded; When facing different shopping malls or seasonal switches, the existing threshold adjustment policy and model migration policy are migrated as meta-models. If there are continuous high false alarm or high missed alarm situations, reminders are sent to the outside or feedback and collaborative updates are triggered.

[0020] (III) Beneficial effects The present invention provides a data anomaly detection method based on multi-scale feature extraction, having the following beneficial effects: Through the organic cooperation of five steps: multi-source data collection and operating environment calibration, multi-scale feature extraction and noise-robust model construction, scenario-aware anomaly detection and sensor redundancy verification, adaptive feedback and collaborative update decision-making, and intelligent threshold learning and scenario adaptation, the accuracy of fault detection can be significantly improved under the complex working conditions of shopping mall elevators, and the following beneficial effects are specifically presented: For the multi-source data stream set carry out environment calibration and generate multi-dimensional environment labels After that, information such as different load weights, speeds, and interference intensities is fully integrated into the data system, enabling the subsequent recognition algorithm to maintain stability in various noise scenarios; Rely on the LTCN model (combining short-term convolutional TCN and long-term dependence LSTM) to obtain multi-scale feature vectors and ensure that real fault signs can still be captured under electromagnetic interference, passenger flow impact, etc. through the noise enhancement strategy; Adopt a scenario-aware dynamic threshold and integrate the elevator operating state and the environment index to adjust the fault determination threshold in real time, significantly reducing the false alarm rate under full load or high noise conditions. At the same time, with the help of the sensor redundancy verification mechanism, false alarms will be automatically suppressed when a single path is abnormal while other sensors are normal, further improving the detection reliability; Through the unified feedback update of false alarms or missed alarms, the threshold and sensor weights can be adaptively revised to ensure the long-term evolution of the system and gradually improve the redundancy fusion strategy; The introduced reinforcement learning and meta-learning can not only iteratively train on existing data, but also quickly migrate or adjust in new scenarios such as across seasons and across shopping malls: on the one hand, rely on the reward function to comprehensively measure accuracy, maintenance cost, and missed alarm risk, and on the other hand, use the policy network to online optimize the dynamic threshold and some parameters of the LTCN to ensure a high recognition rate even under sudden increases in air-conditioning load or high passenger flow environments; Overall, this solution fully covers the entire link of elevator fault detection from data - feature - detection - feedback - adaptation, and can effectively solve the technical problems brought by noise interference, variable load, and scene switching, providing strong technical support for the real - time safety monitoring and subsequent large - scale deployment of shopping mall elevators. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a schematic flow diagram of the data anomaly detection method based on multi - scale feature extraction of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0023] Please refer to Figure 1 , the present invention provides a data anomaly detection method based on multi - scale feature extraction, including, Step 1: When the elevator is in different working conditions such as no - load, full - load, or high - speed, automatically call multi - source sensors, use basic filtering and time synchronization to filter out spike interference and retain key waveform features, and generate multi - dimensional environmental labels in combination with environmental information Then, by inputting context elements such as temperature, humidity, and floor usage frequency enrich the data context to form a high - confidence data set that can be adapted to subsequent multi - scale feature extraction at any time; The content included in the said Step 1 is as follows: Step 101: Environmental basic collection and environmental index extraction Install and activate various types of sensors (including acceleration, vibration, current, voltage, temperature, humidity, and noise, etc.), uniformly number the data output by each sensor to form a preliminary multi - source data stream set , where, represents the current acquisition time, and M represents the number of sensors; During different typical periods of shopping mall operation (such as peak passenger flow period, off - peak period), continuously collect the above - mentioned multi - source data streams and record relevant environmental states (such as passenger flow, electromagnetic interference intensity, indoor temperature and humidity, etc.). For the convenience of subsequent calibration and analysis, record that the elevator is in various operating conditions such as no - load, full - load, acceleration, braking, etc. during the recording period; Perform necessary processing on the collected data to remove extremely high - frequency noise (to avoid the impact of large - amplitude sensor spike values on the overall analysis), and define the environmental index function , It can comprehensively capture the interference characteristics of transient and steady states to measure the noise and disturbance levels under different market conditions, and is defined by the following formula: In the formula: = , representing the M-channel sensor signals collected at time, such as acceleration, current, and voltage; is the subset of weight coefficients corresponding to each sensor (which does not conflict with the adaptive coefficients of other steps in this solution), and is used to highlight the attention to the main noise source or key fault signal in the environmental index; and respectively represent the translation amount and scale in the continuous wavelet transform; and are the lower and upper limits of the value range of the translation amount, and are the lower and upper limits of the scale range; is the matrix representation based on multi-source signals in the time-frequency domain, and can be defined as: Among them, represents the signal of sensor after wavelet transform at the translation amount = , scale to obtain a multi-component vector (for example, splicing complex numbers or multi-channel coefficients into a column vector); represents performing a matrix logarithm operation on , where is the identity matrix (with the same dimension as ), and the purpose is to avoid logarithmic singularity and improve the recognition of small-amplitude perturbations; is a weighted matrix of the same dimension (which can be a diagonal matrix or a band-pass coupling), and is used to further emphasize or weaken the influence of the specified frequency band or sensor channel at the matrix logarithm level; is the matrix trace (Trace) operator, which can compress the result of the matrix logarithm into the scalar domain; is the global scaling factor, which is used to amplify or compress the result after integral calculation; When in use, to establish an objective and quantifiable environmental benchmark for subsequent situation awareness and multi-scale fault detection, different weights can be set for different frequency bands and sensors , it can adaptively highlight key interference sources in the complex environment of shopping mall elevators, such as vibrations during peak and valley passenger flows, electromagnetic interference during the switching of air conditioners and power equipment, etc. By combining frequency-domain analysis with the time domain, transient noise and steady-state disturbance information can be obtained simultaneously, providing a solid data foundation for more effectively separating environmental noise from real anomalies in subsequent fault feature extraction.

[0024] Step 102: Data synchronization and generation of multi-dimensional environment labels Perform time synchronization operations on the obtained multi-source data streams and environmental indices to uniformly timestamp the data of each sensor at the same moment or within the same operation cycle segment, forming multi-dimensional records containing elevator operation data and environmental index functions ; Introduce the working state of the elevator , for example, use the symbol to represent the operating state of the elevator at time (the values can be no load, full load, acceleration, braking, etc.). According to the mall operation information, synchronously record situational elements such as passenger flow, temperature and humidity, and electromagnetic interference sources, and use to uniformly represent these situational elements , for example: Combined with the working state of the elevator , environmental index and situational elements , generate multi-dimensional environment labels for the multi-source data records, such as: ; This label can not only reflect the operating conditions of the elevator itself, but also effectively distinguish different interference levels and situational information in the environmental dimension; Store the multi-dimensional environment labels and the synchronized data together in the environmental label library, providing an integrated source of environmental context + data input that can be directly referenced for subsequent multi-scale feature extraction and scenario-aware anomaly detection steps; At this time, the environmental index extracted in step 101 and the multi-dimensional environment labels generated in step 102 are all uniformly stored in the data structure, ensuring seamless query and use in subsequent steps; When in use, through time synchronization and the establishment of multi-dimensional environment labels, it can be ensured that data under different sensors, different frequency bands, and different operating conditions can be uniformly managed at the same moment Complete information such as the operating status of the elevator, the strength of environmental disturbances, and the passenger flow density is directly retrieved, providing precise context support for subsequent steps (such as differentiating the acceleration differences between no-load and full-load conditions during fault determination), reducing misjudgments or missed judgments caused by inconsistent data records or incomplete working condition annotations, and greatly improving the accuracy of subsequent fault detection and diagnosis of the system.

[0025] In this step (Step 1), through the progressive processing of Steps 101 and 102, a systematic environmental calibration system is established: on the one hand, the interference degree of multi-source data at different frequency bands is characterized using the environmental index ; on the other hand, combined with the actual working condition of the elevator and the elements of the mall scenario, multi-dimensional environmental tagging management is carried out on the collected data. This environmental calibration system can not only provide high-precision input data and context support for subsequent fault detection, but also lay an application foundation that can adapt to noise, passenger flow fluctuations, and electromagnetic interference as a whole, helping to achieve accurate and reliable elevator fault monitoring and identification in a noisy and changing mall environment.

[0026] Step 2: When the environmental calibration is completed and the multi-source data stream set and the multi-dimensional environmental tags are obtained, the LTCN model is constructed by calling the short-term convolution TCN and the long-term memory LSTM under multiple noise levels and with the effect of data augmentation. First, the transient impact signal is captured by the convolutional kernel and then the steady trend is obtained. The output of multi-scale features can improve the fault sensitivity; The content of Step 2 is as follows: Step 201: Short-term convolution channel and environmental weighted fusion Taking the multi-source data stream set as the input of the short-term convolution channel (hereinafter abbreviated as the TCN channel), using one-dimensional or two-dimensional convolutional kernels (specifically depending on the channel structure of the sensor data) to extract features at short time scales and capture transient fault signals such as impacts and collisions; Define the output of the short-term convolution channel as the short-term feature matrix , where S represents Short-Term Features. Within each receptive field, the convolutional kernel can combine hyperparameters such as the stride and dilation rate to achieve sensitive capture of abnormal impacts; To improve the noise suppression ability of the short-term convolution, the environmental index and the operating status are regarded as dynamic weighting factors, enabling the output of the short-term convolution channel to adaptively adjust the sensitivity to different noise environments; Specifically, a fusion function can be defined, and the short-term feature matrix is weighted element-wise or through an attention mechanism with the environmental index , elevator operating state are fused to output a corrected short-term feature vector , and its operation form can be written as: In the formula: represents element-wise multiplication is a all-ones tensor matching the shape of for controllable amplification / scale on the reference value; is an environmental exponential amplification factor, which can be set according to the environmental noise sensitivity, and the value can be between 0.5 and 2.0; is a correction term for the current operating state of the elevator, and different additions or weakenings can be given to certain states (such as high-speed full load) through look-up tables or simple mapping functions; This fusion process can enhance the robustness to environmental disturbances: when the environmental index is high (strong noise), the model (short-term convolutional channel (TCN channel) and its environmental weighted fusion sub-module) automatically reduces the over-response to transient spikes. When the operating state is high load, the fault trigger sensitivity is appropriately increased, etc.; When in use, make full use of the environmental index obtained in the first step and the elevator operating state to dynamically weight the short-term convolutional features, which can distinguish noise shocks from real fault shocks: combined with the convolutional operation within a short time range, the model can capture sudden fault events more precisely, achieving a balance between short-term high sensitivity and noise suppression. Explicitly introducing the environmental index and operating condition labels into the weighting mechanism of the convolutional output layer makes the model more adaptable in a noisy environment.

[0027] Step 202, Long-term Memory Integration and Noise Robustness Training After completing Step 201, the short-term corrected features are obtained. Next, LSTM (Long Short-Term Memory Network) is introduced to capture the smooth operating mode and slow-changing trend of the elevator under different loads and speeds; The serialized short-term corrected features are input into the LSTM layer to obtain the long-term memory vector , where L represents Long-TermFeatures. In this process, LSTM can learn the periodic or trend features of elevator operation across multiple time steps and distinguish the difference between normal slow wear and potential fault signs; At the same time, the short-term corrected features With long-term memory vector Perform concatenation or attention weight fusion to obtain the final multi-scale feature vector : in represents a mapping function (e.g., linear projection or attention fusion) that integrates short-term shock information with long-term trend information; Combine the multi-dimensional environment labels in the first step During the model training phase, noise enhancement is performed on normal operating data, that is, interference fragments from actual measurements or simulations are randomly superimposed, so that the model can still maintain the stability of feature extraction in high-noise scenarios; 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: Where: is the reference feature or annotation vector (which can come from the statistical features 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; 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; When used, LSTM combines the characterization of long-term trends with the capture of short-term shocks by the short-term convolution channel TCN, effectively distinguishing between transient random fluctuations and trend-based fault evolution; the noise enhancement strategy and environmental weighted training enable the model to maintain the ability to extract real fault features when facing 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 focus of the loss function on different noise scenarios.

[0028] By constructing a multi-scale feature extraction model that can capture both 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 with the operating status Perform dynamic weighting so that transient fault signals can still be captured in a noisy background. The LSTM module in step 202 models long-term temporal variations under different load and speed conditions and highlights the discrimination ability for high-noise and real-fault samples with a multi-objective loss function. The multi-scale features output after the two are fused Not only has significant advantages in noise suppression, but also provides high feature confidence and discrimination for subsequent scenario-aware anomaly detection. In the noisy and variable environment of shopping mall elevators, it truly realizes short-term sensitivity + long-term robustness + environment-adaptive fault feature extraction.

[0029] Step 3: When the multi-scale feature vector is output and the elevator working state and the environmental index are already available, for the multi-scale feature vector and the reference vector use the difference to evaluate the anomaly score and calculate the threshold function . If the anomaly score exceeds the limit and the multi-sensor channels are redundantly consistent, it is determined as a fault and an alarm is output; The said step 3 includes the following contents: Step 301: Dynamic threshold setting and scenario-aware anomaly discrimination Take the multi-scale feature vector output in the second step , to measure the degree of deviation of potential faults at this moment , calculate the difference with the reference vector (such as the statistical benchmark of historical normal working conditions characteristics, or the normal state subspace projection marked by maintenance personnel); Define the anomaly score function to measure the difference degree between the multi-scale feature vector and the reference vector , and ensure sufficient resolution in a high-noise environment. The formula can be written as: where: represents -norm, and (such as is the Euclidean distance when is learned from the initial stage or historical normal data and can be updated during deployment or periodic maintenance; To dynamically adjust the alarm sensitivity under different working conditions and environmental interferences, introduce a threshold function that changes with the scenario , combined with the elevator operation status defined in the first step and the environmental index , it can take the following form: In the formula: is the reference threshold mapping function, used to select or interpolate a reference threshold according to the current elevator operation status (such as no load, full load, high speed, etc.); is a smooth non-linear activation function, such as the Sigmoid function, used to map the environmental index to a domain suitable for amplification or attenuation; is a coefficient for adjusting the influence of the environmental index, and its value can be between 0.5 and 1.5; represents the overall scaling factor, and its value can be between 0.1 and 0.3, used to control the threshold value level; When the environmental index is high, also rises accordingly, so as to increase the threshold and avoid a large number of false alarms; when the elevator state represents low speed or no load, may be small, making the model more sensitive to capture minor fault symptoms; Finally, at time , according to the comparison between the anomaly score function and the threshold function to determine whether to trigger a fault suspicion; If , it is determined as a suspicious fault, otherwise it is regarded as normal or noise fluctuation; When in use, the threshold dynamically rises under high-noise or high-load conditions, effectively reducing frequent false alarms caused by normal fluctuations. Under low-noise or special sensitive conditions, the threshold automatically tightens to avoid potential hidden dangers of missed alarms. Through the difference measurement with the reference vector , the intuitive quantification of the fault tendency is realized, and it can adaptively change with the environment; by forming an explicit anomaly score - dynamic threshold mechanism, and using logarithms and exponents to construct, good discrimination is achieved in different amplitude ranges.

[0030] Step 302, Multi-sensor redundant verification and comprehensive judgment In addition to using the multi-scale feature vector output by the fusion in the second step for detection, the local anomaly scores of each sensor at different scales can be retained. For example, the short-term convolutional channel (TCN) unit or the output of the LSTM is split into sensor components, which are correspondingly denoted as , ; for each sensor , define the single-point anomaly score Obtain a set of score vectors : In this way, when a suspicious fluctuation is detected at a certain point, it is possible to judge whether there is consistency by comparing the scores of other sensors; among them, for each sensor Calculate the absolute deviation between the observed value and the predicted (or reference) value, amplify and logarithmically smooth it to obtain the single-point anomaly score; for the score vector Perform multi-source consistency detection and introduce matrix metrics to characterize the overall coupling degree of the anomaly scores of all sensors: In the formula: is the outer product matrix, which can reflect the correlation and amplitude size between the scores of different sensors; is the global scaling coefficient, and its value can be between 0.1 and 2; is the matrix trace (Trace) operator, which can compress the result of the matrix logarithm into the scalar domain for comparison with the threshold; If the matrix metric is significantly greater than a certain set trigger threshold, it indicates that multiple sensors have all generated relatively high anomaly scores, and the confidence level increases accordingly; if only the score of a single sensor is high while others are low, it means that the matrix metric is difficult to reach the trigger threshold, thus reducing the false alarm probability.

[0031] After determining the suspicious fault, perform multi-source consistency verification on the score vector If the matrix metric exceeds the redundancy fusion threshold, it is determined as a multi-source confirmed fault, triggering an alarm or entering adaptive feedback and handling: otherwise, this 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 for updating the reliability weights of each sensor or the fault mode library.

[0032] When in use, by performing the outer product matrix logarithm operation and taking the trace on the single-sensor anomaly score, the coupling characteristics of multi-source simultaneous high anomalies can be captured, effectively reducing false alarms caused by single-source channel noise or fault illusions; the threshold function in step 301 Combined with the multi-source redundancy check in this step, the determination can balance sensitivity and reliability. It can not only adapt to different working conditions but also significantly reduce the risks of missed alarms and false alarms. By using the scenario-aware threshold adjustment and multi-sensor redundancy check mechanism, a flexible and highly reliable fault and anomaly determination process is constructed. It adapts to different passenger flow loads and environmental interference conditions and significantly reduces the false alarm and missed alarm rates through multi-source mutual verification, providing higher reliability for the safety diagnosis of mall elevator failures.

[0033] Step 4: When outputting suspicious or confirmed fault results and obtaining real operation and maintenance feedback, recycle the multi-source data stream set , multi-scale feature vectors and missed alarm and false alarm samples to the training library. By revising the dynamic threshold and the sensor weight w and incrementally updating the multi-scale model parameters Ω, iteratively improve the fault detection accuracy step by step and suppress the frequent false alarms caused by noise fluctuations; The said Step 4 includes the following contents: Step 401: Fault determination result feedback and data re-labeling Match the labels of the suspicious faults and multi-source confirmed faults generated in the third step, as well as the real status feedback given by subsequent manual inspections or maintenance systems, to generate a determination result table; If the system determines a fault and it is confirmed as a fault by manual or maintenance means, it is recorded as a correct alarm, that is, the elevator fault detection and diagnosis system; If the system determines a fault but there is actually no anomaly, it is recorded as a false alarm; If the system determines as normal but there is actually a fault, it is recorded as a missed alarm; Other situations are recorded as normal or suspected fluctuations; For the samples of confirmed false alarms or missed alarms, it is necessary to re-label them in the multi-source environment perception training data set (including the multi-source data stream set , multi-scale feature vectors and multi-dimensional environment labels ) used in the first and second steps, and supplement the corresponding real fault marks or normal marks to form updated training or calibration data; Retain the samples of correct alarms or normal determinations to continue expanding the boundary information between normal and faults; 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 this sensor can be reviewed in this step to exclude determination errors caused by sensor failures or serious deviations. If it is found that false alarms are mostly concentrated in a certain type of working condition (such as full load and high-speed operation), it indicates that there are deficiencies in the model or threshold setting in this working condition scenario, and subsequent steps need to focus on optimizing this scenario.

[0034] During use, by comparing the real fault information with the system determination result, the limitations of the current model can be discovered in a timely manner, and the new information can be fed back into the dataset for re-annotation, laying a foundation for subsequent threshold adjustment and model training. Establishing a dynamic update mechanism for the fault case library can continuously enrich the environment-fault comparison samples in the real shopping mall environment and also enable continuous learning for rare working conditions that are difficult to simulate.

[0035] Step 402: Update of dynamic threshold and sensor weights After collecting the latest correct error reports, false alarms, and missed alarm information, adjust the threshold function used in the third step For example, introduce a differential learning rate based on historical accuracy feedback , and define a new threshold update relationship: Where: is the dynamic threshold; is the error metric representing the current determination result (for example, for a correct error report, can be made less than a certain threshold, while for a false alarm or missed alarm, is made larger); is the threshold correction rate at this moment, which is dynamically determined according to multi-dimensional environmental labels such as passenger flow and electromagnetic interference intensity and the determination accuracy in recent times; is the offset constant, and its value can be between 0.1 and 0.5, which is used to stabilize the correction process; When the determination result has a large deviation multiple times ( exceeds a certain threshold), will deviate significantly from 1, forcing to make corresponding adjustments (increase or decrease); In the process of multi-source sensor fusion, if a single sensor channel frequently conflicts with the determinations of other channels, accordingly reduce the fusion weight corresponding to this channel in the redundancy check; in specific implementation, a weight vector can be defined and combined with the process of calculating the outer product matrix , such as: In the formula: represents weighted adjustment of the abnormal score vector of each sensor; is the identity matrix; is at time the sensor weight After adjustment, the measurement result of the coupling strength of the abnormal scores of multiple sensors, is the global scaling factor, which can be between 0.1 and 2; For the Road sensor at all times The single-point anomaly score of is the matrix trace operator; 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 highly reliable, then the corresponding weight is increased ; This enables differentiated adjustment of multi-source redundant verification, and maintains a high overall accuracy rate even in scenarios with significant noise interference.

[0036] When used, the dynamic threshold correction mechanism can quickly compensate for false alarms or missed alarms, allowing the system to seek a balance between fewer false alarms and fewer missed alarms in continuous iterations; sensor weight redistribution can not only prevent overall misjudgments caused by software and hardware defects or frequent abnormal interference in a certain channel, but 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.

[0037] Step 403: Multi-scale model parameter retraining and collaborative decision output 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. ; When updating, a type of loss function with a penalty term can be defined to highlight the extreme concern for false positive and false negative scenarios, for example: in: Indicated in the parameter Next, the model predicts the characteristic value of the input signal at the moment; According to the real features or expected output after re-annotation; 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; It is a weighting coefficient, which can be between 1.0 and 10. It can give higher weights to false positive / missing samples, so that they can generate larger gradient updates during training. Denote the training mini-batch sample set; Iteratively update the model in mini-batch ( ) or online mode, so that it continuously approaches the optimal state under the guidance of new environmental data and fault cases.

[0038] Pack the dynamically threshold function updated in step 402 together with the sensor weight vector and the retrained model parameters here into the decision layer, so that the latest configuration can be immediately adopted in the next round of detection; If there are structural changes in the mall environment (such as a large number of newly added large-scale equipment, seasonal changes in the passenger flow pattern, etc.), the adaptability of the model to new working conditions can be ensured through regular batch retraining or online fine-tuning; the collaborative update results are continuously recycled in the next round of monitoring-detection-feedback-update closed-loop to form a continuously evolving fault detection system.

[0039] During use, through the retraining of the parameters of the multi-scale model (such as TCN, etc.), the coordinated improvement between the recognition ability and the threshold strategy is realized. When there are new changes in the external environment and the elevator operation mode, this mechanism can enable the model to learn new data distributions and fault symptom characteristics in a timely manner, avoiding continuous misjudgment due to old models; Through the triple linkage of threshold + sensor fusion weight + deep model parameters, the system can co-evolve in different dimensions, maximizing the detection stability and safety under the changing working conditions of mall elevators, enabling the system to continuously optimize the sensitivity and robustness of fault recognition while continuously receiving real operation and maintenance information and environmental changes.

[0040] Combined with the multi-source data acquisition, noise-robust feature extraction and scenario-aware anomaly detection constructed in the previous three steps, this fourth-step closed-loop update mechanism can minimize the false alarm and missed alarm probabilities, ensuring that mall elevators still maintain a high level of safe operation and fault prevention in noisy and changing environments.

[0041] Step Five: When new seasons or sudden changes in the mall environment still occur after multiple rounds of feedback updates, the reinforcement learning or meta-learning module, according to the state description and the reward function , uses the threshold strategy function to coordinate and determine the threshold and schedule and combine the model migration strategy to achieve cross-load and cross-noise adaptive fault detection and shorten the retraining cycle; The fifth step includes the following contents: Step 501: Scene state representation and reward mechanism definition The dynamically updated multi-dimensional environmental labels (including elevator status , environmental index , as well as external passenger flow, electromagnetic interference intensity, etc.) and the current threshold configuration , sensor weight vector , key indexes of model parameters, etc. are combined into a state description ; This state description is the input of the reinforcement learning or meta-learning algorithm, used to depict the environmental situation of the current system and its own detection configuration; to guide the system to spontaneously learn excellent threshold and model migration strategies, it is necessary to set up a multi-dimensional reward function , combined with multiple factors such as the accuracy index, false alarm / missed alarm distribution, and maintenance cost obtained after real-time operation: In the formula: represents the comprehensive accuracy measurement of fault detection in the current time period; and represent the missed alarm rate and false alarm rate respectively; represents the cost or expense in the current scenario adaptation process (such as the additional resource consumption caused by overly frequent threshold adjustment, sensor weighting, etc.); is a positive real number weight, used to balance accuracy, false alarm / missed alarm, and system overhead; In the reinforcement learning framework, each period (such as half an hour, one hour, or other appropriate windows) can be regarded as a decision-making moment. According to the current state description and the reward function select or update the threshold strategy or model migration method at the next decision-making moment; in the meta-learning framework, the learning experience of similar working condition scenarios in historical operation data can be used to achieve fast adaptation to new scenarios or new seasons, as follows: First, input the current scenario feature vector (including environmental labels, existing thresholds, sensor weights, and recent performance indicators) into the policy network , and it outputs two actions: one is the continuous threshold correction amount , and the other is the discrete model update instruction ("no update", "fine-tuning", or "retraining"); after executing these actions, the system calculates the immediate reward according to the detection accuracy, false alarm / missed alarm rate, and update cost in the next period, and then uses the policy gradient algorithm to perform iterative optimization, so as to achieve the adaptive joint decision-making of threshold and model migration.

[0042] When in use, by combining the environmental state and the adaptive configuration into a unified state description , so that the current noise level and working condition characteristics can be more accurately identified in the changing shopping mall situation, and the reward function is introduced 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.

[0043] Step 502: Joint learning of threshold strategy and model migration strategy The dynamic threshold 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.); The updated threshold is , which can be expressed as: 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; 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; For example: If the current state describes Reflecting the new strong seasonal characteristics (air conditioning overload), model migration strategy Perhaps a high-level retraining request will be output to fine-tune some channels of the convolutional layer; if there is only a small fluctuation, the output is a decision to keep 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); 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: Wherein: is the historical period or mini-batch for review; is the time discount factor or meta-learning step size; is the learning rate; is the multi-dimensional reward function; Through this update rule, continuous iteration can be carried out in time series or across scenarios, enabling the threshold policy and the model transfer policy to cooperate with each other and continuously approaching the global optimal solution.

[0044] When in use, incorporating threshold adjustment and model retraining into the same decision-making framework significantly reduces the limitations brought about by solely relying on fixed threshold functions or offline training: when there are seasonal or scenario mutations, the system can quickly detect performance degradation and respond autonomously. By combining time discounting and meta-learning rules, a flexible balance can be achieved between short-term benefits and long-term adaptability, providing higher flexibility for elevator fault detection in complex mall environments.

[0045] Step 503, Online execution and cross-scenario adaptive output Bring the threshold policy function obtained in step 502 and the model transfer policy online in the elevator system in real time. Whenever a new operation cycle arrives, it is determined whether to perform threshold revision and / or local model update through the current state description and at the same time, record the fault detection effect (accuracy, false negatives and false positives) after each decision, and calculate the new reward function continuously enrich the learning samples; represents the dynamic fault determination threshold after being corrected by the reinforcement learning (RL) strategy; When the system faces different malls or seasonal switches, the existing threshold adjustment policy can be migrated as a meta-model together with the model transfer policy

[0046] : in the new scenario, only a small amount of online tuning or small-scale training is required to achieve the available threshold configuration and feature extraction capabilities; for completely unfamiliar extreme scenarios, certain safety policies (such as increasing alarm redundancy) can also be retained to minimize the risk of false negatives, and gradually converge to a more reasonable threshold through learning; during the execution process, if there are continuous high false positives or high false negatives, remind the maintenance personnel or trigger the detailed feedback and collaborative update process in the fourth step again.

[0046] In use, the online execution mechanism ensures that the system can adjust the thresholds and model strategies in a timely manner according to the latest situation during the actual operation of the elevator, shortening the latency of feedback-improvement. Through the meta-model migration method, it can be applied to different shopping malls or seasons with one key without retraining from scratch, further improving the generality of the detection scheme. Combining real-time online execution with cross-scenario migration to obtain the maximum adaptive effect with minimal manual intervention, enabling the elevator fault detection system to transform from passive update to active learning in the Internet of Things application of shopping mall elevators, thus effectively improving the overall safety and operation and maintenance efficiency and having the potential to be extended to other complex electromechanical equipment monitoring scenarios.

[0047] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of this application.

[0048] Those skilled in the art can 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 foregoing method embodiments and will not be elaborated herein.

[0049] In 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 illustrative. For example, the division of the units is only for some logical function divisions, and there can be other division methods in actual implementation. For example, 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 displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces, and the indirect couplings or communication connections of the devices or units can be in electrical, mechanical, or other forms.

[0050] The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they can be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0051] As described above, it is only the specific implementation manner of the present application. However, the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims described above.

Claims

1. A data anomaly detection method based on multi-scale feature extraction, characterized in that: including When the elevator is in different working conditions, it automatically calls multi-source sensor data, adopts a unified sequence benchmark for filtering and time synchronization, combines passenger flow and scenario factors to generate multi-dimensional environment labels, and constructs a high-confidence dataset for subsequent multi-scale analysis; After initially cleaning the multi-source sensor data, the LTCN structure is enabled in a multi-noise enhancement scenario. First, the short-term convolutional TCN is used to capture transient shocks, and then the LSTM is used to refine long-term dependencies, output multi-scale features, and enhance the sensitivity of fault detection; When multi-scale features are generated and the elevator working conditions and environmental indices are obtained, a dynamic threshold is determined based on the difference judgment, and only multi-source consistent anomalies are triggered for alarm through multi-sensor redundant verification, reducing false alarms caused by noise interference; If a suspicious result is output and real feedback is obtained, the multi-source sensor data and multi-scale features are traced back to re-label misreported and missed reported samples, incrementally revise the judgment threshold and sensor weights, and update the multi-scale model; When environmental seasonal mutations or cross-mall deployment requirements occur, the reinforcement learning and meta-learning mechanisms are enabled to adaptively correct the judgment threshold according to the current state description and reward function and transfer some LTCN parameters.

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, continuously collect multi-source data streams at different typical time periods to form a multi-source data stream set, record the elevator operating conditions and related environmental states during the time period; after preprocessing the collected data, construct an environmental index function to measure the noise and disturbance levels under different market conditions; Synchronize the obtained multi-source data streams and environmental indices in time to form a multi-dimensional record including elevator operation data and environmental index function, introduce the elevator working state, and generate context elements from the mall operation information; Combining the elevator working state, environmental index and context elements, generate multi-dimensional environment labels for the multi-source data record.

3. The data anomaly detection method based on multi-scale feature extraction according to claim 2, characterized in that: Use the multi-source data stream set as the input of the short-term convolutional channel, and extract transient fault signals at short time scales; Define the output of the short-term convolutional channel as the short-term feature matrix, and fuse the short-term feature matrix with the environmental index and elevator working state through element-wise weighting or attention mechanism to output a corrected short-term feature vector.

4. The data anomaly detection method based on multi-scale feature extraction according to claim 3, characterized in that: Input the serialized short-term corrected features into the LSTM layer to obtain long-term memory vectors, and at the same time splice or fuse the short-term corrected features with the long-term memory vectors through attention weights to obtain the final multi-scale feature vector; Combined with multi-dimensional environment labels, noise enhancement is performed on normal working condition data during the model training stage, and a multi-objective loss function is defined to enable the model to perform weighted learning in a high-noise scenario and reduce the risk of overfitting to low-noise samples.

5. The data anomaly detection method based on multi-scale feature extraction according to claim 4, characterized in that: To dynamically adjust the alarm sensitivity under different working conditions and environmental disturbances, a threshold function that varies with the scenario is introduced by combining the elevator operation status and the environmental index, and 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 deviation of potential faults. Whether a fault suspicion is triggered is determined based on the comparison between the anomaly score function and the threshold function. If the anomaly score function is greater than the threshold function, it is determined as a suspicious fault; otherwise, it is regarded as normal or noise fluctuation.

6. The data anomaly detection method based on multi-scale feature extraction according to claim 5, wherein: After defining the single-point anomaly score for each sensor to obtain a score vector, after determining the suspicious fault, multi-source consistency detection is performed on the score vector, and a matrix metric is introduced to characterize the overall coupling degree of the anomaly scores of all sensors; If the matrix metric exceeds the redundancy fusion threshold, it is determined as a multi-source confirmed fault, triggering an alarm or entering adaptive feedback and handling; otherwise, this anomaly is regarded as a non-genuine 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, wherein: Match the labels of the suspicious fault and the multi-source confirmed fault with the true state feedback to generate a decision result table, including correct error reporting, false alarm, and missed alarm, normal or suspected fluctuation records; for the samples with confirmed false alarm or missed alarm, re-label them and supplement the corresponding true fault label or normal label to form updated training or calibration data; retain the samples with correct error reporting or normal determination, and conduct a review of feasible sources.

8. The data anomaly detection method based on multi-scale feature extraction according to claim 7, wherein: After collecting the latest correct error reporting, false alarm, and missed alarm information, update the threshold function; If there is a frequent conflict between the determination of a single sensor channel and other channels, correspondingly reduce the fusion weight of this channel in the redundancy check; when the false alarm rate of a certain sensor is higher than expected, lower the corresponding weight, and if its contribution to the fault judgment is highly reliable, increase the corresponding weight; For the multi-scale model LTCN, use the re-labeled fault samples and normal samples for incremental or periodic re-training to update the learnable parameters; When updating, define a loss function with a penalty term, and iteratively update the multi-scale model in a mini-batch or online manner, so that it continuously approaches the optimal state under the guidance of new environmental data and fault cases; Package the updated dynamic threshold function, the sensor weight vector, and the re-trained model parameters here into the decision layer so that the latest configuration can be immediately adopted in the next round of detection.

9. The data anomaly detection method based on multi-scale feature extraction according to claim 8, wherein: Combine the dynamically updated multi-dimensional environmental tags with the current threshold configuration, sensor weight vector, and key indices of model parameters into a state description, which serves as the input for reinforcement learning or meta-learning algorithms and sets up a multi-dimensional reward function; in the reinforcement learning framework, select or update the threshold policy or model migration method at the next decision-making moment based on the current state description and the reward function; make migration decisions for the multi-scale feature extraction model and define the model migration strategy under the meta-learning or reinforcement learning framework. Encapsulate the two strategies of the threshold policy function and the model migration strategy in a unified reinforcement learning or meta-learning update loop, and perform gradient optimization through the reward function obtained by interaction.

10. The data anomaly detection method based on multi-scale feature extraction according to claim 9, wherein: Put the updated threshold policy function and model migration strategy into real-time operation in the elevator system. Whenever a new operation cycle arrives, determine whether to revise the threshold and / or update the local model based on the current state description, and record the fault detection effect after each decision; when facing different shopping malls or seasonal switches, use the existing threshold adjustment strategy and model migration strategy as the meta-model for migration. If there are continuous high false alarm or high miss alarm situations, send a reminder to the outside or trigger feedback and collaborative update.

Images