Method for monitoring the safety of an elevator operating state

Abstract

The application discloses an elevator operation state safety monitoring method. The elevator steel wire rope vibration signal and motor driving current signal are collected in real time; based on the start-stop instruction signal of the elevator frequency converter, the filter mode is dynamically switched by the hardware level gate circuit to suppress impact noise in the start-stop stage and capture wear characteristics in the steady state stage; the filtered vibration signal is input into the time domain feature extraction submodule of the lightweight neural network model to extract the time domain impact feature; the current signal is converted into frequency spectrum data and fused to generate a time-frequency feature map; the recurrent neural network in the time series analysis submodule identifies the wear trend feature and outputs a wear probability value; when the value exceeds the dynamically corrected safety threshold, a safety protection action is triggered. The cooperative design of hardware level dynamic filtering and embedded lightweight neural network analysis solves the problems of false alarm in the start-stop stage of the elevator, early identification difficulty of the progressive wear of the steel wire rope and high real-time safety monitoring in the resource limited environment.

Description

Technical Field

[0001] This invention relates to the field of elevator safety monitoring technology, specifically to a method for monitoring the safety status of elevator operation. Background Technology

[0002] Elevators, as indispensable vertical transportation tools in modern buildings, have operational safety and reliability that directly impacts the safety of people's lives and property. Elevator wire ropes, as key components for load-bearing and transmission, are a focus of elevator safety monitoring due to potential failures caused by fatigue, wear, or deformation. Currently, monitoring the operational status of elevators, especially for early wear and fault warnings of wire ropes, mainly relies on the following two types of technical solutions: The first category is the traditional method based on sensor signal threshold determination. This method collects physical signals during elevator operation using vibration sensors, current sensors, etc., and presets a fixed safety threshold. When the signal amplitude exceeds the threshold, the system triggers an alarm. However, this method has limitations: First, the elevator generates huge mechanical impacts during transient processes such as starting and braking, and the signal amplitude can easily exceed the fixed threshold, leading to frequent false alarms; second, the fixed threshold cannot effectively identify the gradual micro-wear of the wire rope, and can only be detected when the wear reaches a severe level, resulting in insufficient timeliness and accuracy of early warning, failing to meet the needs of early fault diagnosis.

[0003] The second category is advanced monitoring methods based on intelligent algorithms. In recent years, some research has attempted to apply artificial intelligence models such as neural networks to elevator fault diagnosis, identifying anomalies by analyzing the time-frequency characteristics of vibration or current signals. While these methods have shown potential in pattern recognition, they face severe challenges in practical engineering implementation. On the one hand, complex neural network models involve large computational loads and high inference latency, making them difficult to deploy in the embedded chips of elevator control cabinets with limited computing and storage resources. Data often needs to be uploaded to cloud servers for processing, introducing communication latency and failing to meet the stringent real-time response requirements of elevator safety systems. On the other hand, existing algorithm solutions typically employ software filtering, whose filtering mode switching depends on the instruction scheduling of the central processing unit, resulting in slow response speeds. In the transient phase of elevator start-up and stop, they cannot quickly suppress impact noise, leading to low signal-to-noise ratios in the input signals for subsequent feature extraction, severely impacting the accuracy of model judgments. Summary of the Invention

[0004] The purpose of this invention is to provide a method for monitoring the safety of elevator operation status, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for monitoring the safety status of an elevator, comprising the following steps: Vibration signals of the elevator wire rope and motor drive current signals are collected in real time using vibration sensors and current sensors. Based on the start and stop command signals of the elevator frequency converter, the filtering mode is dynamically switched through the hardware-level gate control circuit: when the elevator start or stop command is detected, it automatically switches to the high-frequency filtering mode to suppress mechanical impact noise; when the elevator is in the steady-state operation stage, it switches to the low-frequency filtering mode to capture progressive wear characteristics. The filtered vibration signal is input into the time-domain feature extraction submodule of the lightweight neural network model, and the vibration signal is convolved to extract the time-domain impact features. The current signal is synchronously converted into spectral data and fused with the time-domain impulse characteristics to generate a time-frequency feature map. The time-frequency feature map is input into the time-series analysis submodule of the lightweight neural network model, and wear trend features are identified through recurrent neural network units; Based on the wear trend characteristics, a wear probability value is output, and a safety protection action is triggered when the wear probability value exceeds a preset safety threshold. The hardware-level gate control circuit and the inverter start / stop signal line are directly connected physically, without the need for central processing unit intervention; the lightweight neural network model is deployed in the embedded chip of the elevator control cabinet.

[0006] Preferably, the dynamic switching filtering mode of the hardware-level gating circuit specifically includes: During the transient phase triggered by the elevator start command, the gate control circuit activates the high-pass filter characteristic to attenuate mechanical impact noise in the 0-50Hz frequency band, while increasing the sampling density of the vibration signal. During the transient phase triggered by the elevator braking command, the gate control circuit activates the low-pass filter characteristic to suppress high-frequency interference caused by wire rope slack. The switching of the filtering mode is directly driven by the level transition of the inverter start / stop signal, and line interference is eliminated through a differential signal transmission mechanism.

[0007] Preferably, the temporal feature extraction submodule adopts a one-dimensional convolutional neural network architecture, and its specific operations include: The filtered vibration time series data is processed by sliding window convolution, and the window length is dynamically adjusted to adapt to the elevator running speed. Local temporal features of vibration signals are extracted using multi-layer convolutional kernels; The output of the convolutional neural network is compressed by the pooling layer to generate a temporal feature vector, which is then fused with the spectral data using a feature concatenation method.

[0008] Preferably, the time series analysis submodule adopts a recurrent neural network architecture, and its specific operations include: Using time-frequency feature maps as input sequences, historical state information is dynamically filtered through a forget gate mechanism. During the elevator start-up and stop phases, enhance sensitivity to the rate of spectral energy decay; The number of hidden layer units in the recurrent neural network is optimized through pruning to reduce computational complexity.

[0009] Preferably, the comparison mechanism between the output of the wear probability value and the safety threshold includes: A two-layer decision-making logic is constructed: the first layer triggers an initial warning based on the comparison between the wear probability value and the basic threshold; the second layer dynamically adjusts the threshold based on the elevator operation stage. When the elevator is in the starting or braking phase, the safety threshold is automatically increased, and the decision observation window is extended. The trigger signal for the safety protection action is transmitted to the safety relay through an optocoupler isolation circuit.

[0010] Preferably, the deployment optimization of the lightweight neural network model includes: The model parameters are quantized and compressed, converting floating-point weights into fixed-point representations; A dedicated data buffer is established in the shared memory area of ​​the embedded chip for intermediate feature transfer between the time-domain feature extraction submodule and the time-series analysis submodule; The total number of model parameters is controlled within 70% of the available storage capacity of the embedded device.

[0011] Preferably, the direct physical connection between the hardware-level gating circuit and the inverter start / stop signal line is specifically implemented as follows: The control output port of the frequency converter is directly coupled using a TTL level signal line. A signal shaping module is integrated at the input of the gate control circuit to harden the rising and falling edges of the start and stop commands; The differential signal transmission mechanism is implemented using twisted-pair cables, with a common-mode rejection ratio of not less than 40dB.

[0012] Preferably, the adaptive window length adjustment mechanism of the temporal feature extraction submodule includes: The speed feedback signal from the elevator encoder is analyzed in real time, and the convolution window length is automatically extended when the running speed is below 0.5m / s. During high-speed operation, shorten the window length; The window length has a non-linear inverse relationship with the running speed, and dynamic matching is achieved through a preset mapping function.

[0013] Preferably, the optimization of the forget gate mechanism in the recurrent neural network architecture includes: A learnable forgetting weight coefficient is introduced to dynamically adjust the retention ratio of historical states according to the elevator operation stage; During the transient phase, reduce the dependence on historical states; During the steady-state phase, the fusion weight of historical states is increased.

[0014] Preferably, the dynamic correction method for the safety threshold of the two-layer decision logic includes: The basic threshold is periodically calibrated based on the elevator's service life and load history; During the startup phase, the safety threshold decays over time according to a preset curve: the initial value increases to 1.5 times the base threshold, and then linearly returns to the base threshold after 0.5 seconds. During the braking phase, the increase in the safety threshold is positively correlated with the braking deceleration, which is dynamically calculated by monitoring the deceleration parameters fed back by the frequency converter in real time.

[0015] Compared with the prior art, the beneficial effects of the present invention are: to achieve dynamic switching of filtering modes through hardware-level gating circuit to suppress transient interference, and to combine a lightweight neural network model for time-frequency feature fusion and wear trend identification, effectively solving the problem of false alarms during elevator start-up and stop and the problem of early detection of progressive wear of wire rope, while adapting to the resource constraints of embedded systems. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the monitoring method according to an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 This application proposes a method for monitoring the safety of elevator operation status, including the following steps: Vibration signals of the elevator wire rope and motor drive current signals are collected in real time using vibration sensors and current sensors. Based on the start and stop command signals of the elevator frequency converter, the filtering mode is dynamically switched through the hardware-level gate control circuit: when the elevator start or stop command is detected, it automatically switches to the high-frequency filtering mode to suppress mechanical impact noise; when the elevator is in the steady-state operation stage, it switches to the low-frequency filtering mode to capture progressive wear characteristics. The filtered vibration signal is input into the time-domain feature extraction submodule of the lightweight neural network model, and convolution operation is performed on the vibration signal to extract the time-domain impact features. The current signal is synchronously converted into spectral data and fused with the time-domain impulse characteristics to generate a time-frequency feature map. The time-frequency feature map is input into the time-series analysis submodule of the lightweight neural network model, and the wear trend features are identified through the recurrent neural network unit; Output a wear probability value based on wear trend characteristics, and trigger a safety protection action when the wear probability value exceeds a preset safety threshold. Among them, the hardware-level gate control circuit and the inverter start / stop signal line are directly connected by a physical connection, without the need for central processing unit intervention; the lightweight neural network model is deployed in the embedded chip of the elevator control cabinet.

[0019] Hardware-level gating circuits can be understood as a hardware structure used to rapidly switch signal filtering modes, implemented in various ways. For example, they can employ circuit designs based on analog switches, using the on and off states of the analog switches to select different filtering paths; or they can be implemented using programmable logic devices, where the state machine within the logic device controls the switching of filtering parameters. Their main purpose is to achieve rapid response to filtering needs during the transient processes of elevator start-stop, avoiding untimely signal processing caused by software scheduling delays.

[0020] Lightweight neural network models can be understood as neural network architectures optimized for specific application scenarios, implemented in various ways. For example, pruning techniques can be used to remove redundant network layers, retaining only the core computational units that contribute to target feature extraction; or knowledge distillation techniques can be used to transfer knowledge from complex models to simplified models, thereby reducing computational complexity while maintaining performance. Their primary purpose is to adapt to the limited computing resources of embedded devices, ensuring real-time analysis can be performed locally.

[0021] Specifically, the convolution operation in the temporal feature extraction submodule can be implemented in various ways. For example, a sliding convolution operation with a fixed window length can be used to extract local features by segmenting the signal; or an adaptive window length design can be adopted to dynamically adjust the coverage of the convolution kernel according to the signal characteristics. This is mainly to extract key features from the vibration signal that reflect the abnormal state of the wire rope.

[0022] Furthermore, the generation of time-frequency feature maps can be achieved in various ways. For example, short-time Fourier transform can be used to convert the current signal into spectral data, and then combined with the time-domain features of the vibration signal through matrix concatenation; alternatively, wavelet transform can be used to decompose the current signal into multiple scales, and then the decomposition results can be weighted and fused with the vibration features. The main purpose is to integrate time-domain and frequency-domain information to comprehensively characterize the elevator's operating status.

[0023] The current signal is converted into spectral data through a short-time Fourier transform using a Hamming window with a length of 256 points and an overlap of 128 points, and a sampling rate of 2kHz. The generated spectral matrix is ​​concatenated with the vibration time-domain feature vector along the feature dimension to form a time-frequency feature map, which is then input into the time-series analysis submodule. The time-series analysis submodule employs a pruned recurrent neural network (RNN), with 48 hidden layer units after pruning optimization. A learnable weight coefficient is introduced in the forget gate mechanism, and a retention ratio between 0 and 1 is output through the sigmoid function. During the elevator start-up or braking phase, the system automatically reduces this ratio to reduce reliance on historical states, while increasing the ratio during the steady-state phase to enhance trend memory. After training, the entire neural network model is quantized and compressed, with the weights converted from 32-bit floating-point to 8-bit fixed-point numbers, and stored in a dedicated data buffer reserved in the shared memory of the embedded chip. The buffer size is set to 16KB and managed using a circular queue to ensure efficient transfer of intermediate features between the time-domain feature extraction and the time-series analysis submodule.

[0024] The innovation of this application lies in its collaborative design of hardware-level dynamic filtering and lightweight neural networks. This addresses the issues of false alarms caused by mechanical impact noise during elevator operation's start-up or braking transients, insufficient early detection of progressive wire rope wear, and achieving high real-time safety monitoring under conditions of limited embedded chip resources in elevator control cabinets. Compared to existing technologies that rely on fixed thresholds or cloud computing, this application achieves rapid switching of filtering modes through hardware-level gating circuits, improving the real-time performance of signal processing. Simultaneously, the optimized deployment of the lightweight neural network model ensures efficient analysis in the embedded environment, avoiding the latency risks associated with cloud transmission. This end-to-end optimized design effectively overcomes the shortcomings of existing technologies in terms of anti-interference, real-time performance, and embedded adaptability.

[0025] The working principle of this application embodiment is as follows: Vibration signals of the elevator wire rope and motor drive current signals are collected in real time by vibration sensors and current sensors, providing basic data support for subsequent analysis. Based on the start and stop command signals of the elevator inverter, the hardware-level gating circuit dynamically switches the filtering mode. During the elevator start or braking phase, it automatically switches to the high-frequency filtering mode to suppress mechanical impact noise and avoid false alarms caused by transient impacts; during the steady-state operation phase, it switches to the low-frequency filtering mode, thereby effectively capturing the characteristics of progressive wear of the wire rope. The hardware-level gating circuit and the inverter start and stop signal lines are directly connected physically, requiring no intervention from the central processing unit, shortening the response time of filtering mode switching and enhancing anti-interference capability.

[0026] Furthermore, the filtered vibration signal is input into the time-domain feature extraction submodule of the lightweight neural network model. Time-domain impact features are extracted by convolutional operations on the vibration signal, quantifying the abnormal vibration of the wire rope. Simultaneously, the current signal is converted into spectral data and fused with the time-domain impact features to generate a time-frequency feature map, integrating time-domain and frequency-domain information to comprehensively characterize the elevator's operating status. This time-frequency feature fusion design overcomes the limitations of single-dimensional analysis, capturing both explicit anomalies caused by transient impacts or wear in the wire rope and implicit anomalies in the load matching between the motor drive system and the wire rope.

[0027] The time-frequency feature map is input into the time-series analysis submodule of the lightweight neural network model. Recurrent neural network units identify wear trend characteristics and capture the gradual evolution of wear. Based on the wear trend characteristics, a wear probability value is output, and a safety protection action is triggered when the wear probability value exceeds a preset safety threshold, achieving trend-based decision-making. The lightweight neural network model is deployed in an embedded chip in the elevator control cabinet, ensuring localized real-time analysis under limited resource conditions and avoiding the latency risks and data privacy leakage issues associated with cloud transmission.

[0028] The hardware filtering front end quickly eliminates transient noise interference, ensuring the quality of the input signal; the feature extraction and fusion stage combines time-domain impact and spectral information to enhance the sensitivity to micro-wear features; the time-series analysis submodule uses a recurrent neural network to capture historical state changes, improving the reliability of trend recognition; the entire process is executed in a closed loop in an embedded environment, and the real-time response capability is optimized through hardware and software collaboration, thereby systematically solving the problems of false alarms, early identification, and resource constraints.

[0029] The filtered vibration signal is input into the temporal feature extraction submodule of the lightweight neural network model. This submodule adopts a one-dimensional convolutional neural network architecture, containing two convolutional layers and one max-pooling layer. The first convolutional layer uses 32 convolutional kernels of length 8, and the second convolutional layer uses 64 convolutional kernels of length 4, both with a stride of 1. The activation function is ReLU. The convolutional window length is dynamically adjusted according to the speed signal fed back by the elevator encoder, specifically through a mapping function. Implementation, in which The real-time speed is in m / s. , and Based on parameters fitted from measured data, the window length is extended to 128 sampling points when the speed is below 0.5 m / s, and shortened to 32 sampling points when running at high speed. The pooling layer uses max pooling with a length of 2, and the output feature vector is concatenated and fused with the current spectrum data.

[0030] , and Their physical meanings are scale factor, velocity offset, and base window length, respectively. These parameters were determined through the following experimental calibration method: During the elevator commissioning phase, the elevator was controlled to run at several constant speeds, such as 0.1, 0.5, 1.0, 1.5, and 2.0 m / s, and corresponding standard vibration signals were collected. The length of the signal segment that best characterizes the wire rope wear at each speed was analyzed and determined as the optimal window length at that speed. Subsequently, the least squares method was used to analyze the data points. By performing curve fitting, a set of specific parameter values ​​can be obtained. For example, in a certain type of elevator, the fitting yields a = 120, b = 0.3, c = 10. At this time, when the speed v = 0.5 m / s, the window length L ≈ 128 sampling points; when v = 2.0 m / s, L ≈ 32 sampling points.

[0031] This application further proposes a dynamic switching filtering mode for the hardware-level gate control circuit, specifically including: during the transient phase triggered by the elevator start command, the gate control circuit activates the high-pass filtering characteristic to attenuate mechanical impact noise in the 0-50Hz frequency band, while increasing the sampling density of the vibration signal; during the transient phase triggered by the elevator brake command, the gate control circuit activates the low-pass filtering characteristic to suppress high-frequency interference caused by wire rope slack; the switching of the filtering mode is directly driven by the level transition of the inverter start / stop signal, and line interference is eliminated through the differential signal transmission mechanism.

[0032] High-pass filtering refers to a filtering method that effectively attenuates signal components below a set cutoff frequency. It is implemented using RC high-pass filter circuits or digital filtering algorithms, aiming to specifically remove low-frequency vibration noise generated by the sudden load on the motor during elevator startup. Low-pass filtering refers to a filtering method that effectively attenuates signal components above a set cutoff frequency. It is implemented using LC low-pass filter circuits or moving average filtering algorithms, aiming to suppress high-frequency oscillation interference caused by sudden changes in wire rope tension during elevator braking. Differential signal transmission is a technique that uses two signal lines to transmit signals with opposite phases, implemented through twisted-pair cables and differential amplifiers, to eliminate the influence of external electromagnetic noise on start / stop signal transmission.

[0033] During the transient phase triggered by the elevator start command, the gate control circuit quickly activates the high-pass filter characteristic by detecting level transitions in the inverter start / stop signal. This immediate response ensures effective suppression of mechanical impact noise in its early stages. Simultaneously, the system automatically increases the sampling density of the vibration signal, allowing for the complete capture of subtle wear characteristics even during transient impacts. During the transient phase triggered by the elevator brake command, the gate control circuit similarly switches quickly to low-pass filter mode via level transition detection. This switching requires no central processing unit intervention, ensuring rapid response. The differential signal transmission mechanism plays a crucial role in this process. It isolates external electromagnetic noise through the common-mode rejection characteristics of twisted-pair cables, ensuring accurate transmission of start / stop signals and preventing erroneous switching of filter modes due to line interference. By distinguishing the noise characteristics during the elevator's transient phase, dynamic adaptation of filtering behavior is achieved, improving the reliability of the monitoring system under complex operating conditions.

[0034] It solves the problem of insufficient signal-to-noise ratio caused by the difference in noise characteristics during the start-up and braking transient phases, and also ensures the timeliness of filter mode switching through hardware-level fast response, ultimately improving the accuracy of subsequent wear feature extraction.

[0035] This application further proposes a one-dimensional convolutional neural network architecture for the temporal feature extraction submodule. The specific operations include: performing sliding window convolution processing on the filtered vibration time series data, with the window length dynamically adjusted to adapt to the elevator running speed; extracting local temporal features of the vibration signal through multi-layer convolutional kernels; and generating a temporal feature vector by compressing the output of the convolutional neural network through a pooling layer. The fusion of this vector with the spectral data adopts a feature concatenation method.

[0036] One-dimensional convolutional neural network architecture refers to a deep learning model specifically designed for processing time-series data, implemented using multiple convolutional layers and activation functions. Sliding window convolution is a technique that progressively moves a fixed-size window across the time series and performs convolution operations, aiming to capture local features in vibration signals. Multi-layer convolutional kernels refer to multiple convolutional filters with different sizes or parameters, implemented using different weight matrices, to extract features from vibration signals at multiple scales. Pooling layers are a dimensionality reduction operation, implemented through max pooling or average pooling, reducing data volume while preserving key features. Feature concatenation refers to combining feature vectors from different sources into new feature representations according to certain rules, achieved through simple vector concatenation or weighted fusion methods.

[0037] By analyzing the speed feedback signal from the elevator encoder in real time, the length of the sliding window is dynamically adjusted. This allows the window to be extended at low speeds to capture more complete progressive wear characteristics, while shortening at high speeds to avoid noise interference. This adaptive mechanism ensures that the convolution operation always matches the vibration signal characteristics at the current speed. The multi-layer convolution kernel design can identify impact patterns at different scales hierarchically, such as the weak pulse initiation point and decay process in the early stages of wear, providing a key criterion for distinguishing between normal mechanical vibration and abnormal wear. The introduction of pooling layers not only reduces data dimensionality but also retains the most representative features, thereby alleviating the computational burden on embedded devices. The feature stitching method fully preserves the correlation between time-domain impact features and frequency-domain energy distribution, enabling subsequent time-series analysis to comprehensively integrate multi-dimensional information and accurately identify wear trends. High-precision feature extraction with speed adaptation is achieved under limited embedded resource constraints, solving the problems of incomplete feature capture and noise interference caused by elevator speed fluctuations, laying a reliable foundation for real-time safety monitoring.

[0038] This application further proposes specific operations for the time series analysis submodule using a recurrent neural network architecture, including: using the time-frequency feature map as the input sequence and dynamically filtering historical state information through a forget gate mechanism; enhancing the sensitivity to the spectral energy decay rate during the elevator start-stop phase; and optimizing the number of hidden layer units in the recurrent neural network through pruning to reduce computational complexity.

[0039] The time-frequency feature map refers to a two-dimensional feature representation that integrates the time-domain impact characteristics of vibration signals and the spectral data of current signals. It is generated using methods such as short-time Fourier transform or wavelet transform, aiming to provide richer feature information for subsequent wear trend identification. The forget gate mechanism can be understood as a gating structure that controls the proportion of historical state information retained. It achieves dynamic adjustment by introducing learnable weight coefficients, selectively filtering invalid information at different stages of elevator operation. Furthermore, the pruning optimization of the number of hidden layer units refers to simplifying the network structure by removing redundant neurons, achieved using gradient-based methods or sparse constraints, thereby reducing the computational load of the model.

[0040] Deep optimization of the recurrent neural network architecture enables effective monitoring of elevator wire rope wear trends. First, a time-frequency feature map containing rich information is used as the input sequence. A forget gate mechanism is employed to adjust the retention ratio of historical state information in real time according to the elevator's operating stage. This dynamic filtering method effectively suppresses mechanical shock noise interference during the start-up and braking phases, ensuring the purity of the input data. Simultaneously, the sensitivity to the spectral energy decay rate is specifically enhanced during the elevator start-up and stop phases. This design fully utilizes the correlation between wire rope wear and spectral energy changes, enabling the system to more sensitively capture early micro-wear signals. By optimizing the number of hidden layer units in the recurrent neural network, the computational complexity of the model is reduced, making it adaptable to the resource constraints of embedded chips and ensuring efficient real-time processing under limited hardware conditions. The above optimization measures, combined with a lightweight neural network model deployment scheme, solve the problem of transient noise interference and overcome the bottleneck of embedded deployment, forming a complete solution from algorithm optimization to hardware adaptation.

[0041] Deploying a recurrent neural network in the embedded chip of the elevator control cabinet to identify wear trends ensures the accuracy of fault warnings and meets the requirements of real-time system response.

[0042] This application further proposes a comparison mechanism between the output of the wear probability value and the safety threshold, including: A two-layer decision-making logic is constructed: the first layer triggers an initial warning based on the comparison between the wear probability value and the basic threshold; the second layer dynamically adjusts the threshold based on the elevator operation stage. When the elevator is in the starting or braking phase, the safety threshold is automatically increased, and the decision observation window is extended. The trigger signal for the safety protection action is transmitted to the safety relay through an optocoupler isolation circuit.

[0043] The two-layer decision logic refers to a hierarchical judgment mechanism implemented using software algorithms or hardware circuit modules. The first layer of this mechanism is used for rapid anomaly detection, while the second layer adjusts the judgment criteria based on the operating status, aiming to improve the accuracy of early warnings and reduce false alarms. Dynamically corrected thresholds can be adjusted in real time according to different stages of elevator operation, achieved through preset rule tables or adaptive algorithms, designed to adapt to changes in transient signal characteristics. The decision observation window refers to the continuous monitoring period before the system makes a final judgment, implemented through a timer circuit or software timer, aiming to provide a more stable evaluation basis after transient impacts.

[0044] Optocoupler isolation circuits are electrical isolation devices that use optocouplers to transmit signals, blocking noise interference paths and ensuring the reliability of trigger signals. Safety relays are key components for performing safety protection actions; they are implemented using electromagnetic relays or solid-state relays, and their purpose is to quickly respond to trigger signals and execute protective actions.

[0045] A two-layer decision-making logic enables tiered judgment of wear probability values. The first layer performs rapid screening based on a basic threshold, triggering an initial warning immediately upon detecting an anomaly. The second layer dynamically adjusts the threshold based on the elevator's operating phase, automatically raising the safety threshold and extending the observation time during start-up or braking, effectively avoiding false alarms caused by transient impacts. This mechanism allows the system to flexibly adjust judgment criteria according to different operating states, improving the accuracy of warnings. Simultaneously, the application of optocoupler isolation circuits ensures the purity of the trigger signal, preventing external interference from affecting the operation of the safety relays, guaranteeing the reliability of the entire safety protection system, solving the problem of false alarms caused by fixed thresholds, and improving the overall performance of the system in complex operating environments.

[0046] After the wear probability value is output, it enters a two-layer decision-making logic for safety assessment: The first layer compares the probability value with a basic threshold; if it exceeds the threshold, a preliminary warning is triggered. The second layer dynamically adjusts the safety threshold based on the elevator's operating stage. The basic threshold is calibrated monthly based on the elevator's cumulative operating time and historical load data, using the following calibration formula: ,in As the initial threshold, For the number of years of operation, Average load factor , This is the attenuation coefficient. During the startup phase, the safety threshold is set according to... The curve decays over time, where The time after startup (in seconds) is 0.5 seconds, after which it returns to the baseline threshold; during the braking phase, the increase in the safety threshold is related to the braking deceleration. Positive correlation, specifically ,in The proportional coefficient represents the deceleration. The probability is calculated based on the real-time current change rate fed back from the frequency converter. When the probability value exceeds the dynamically corrected threshold, the system transmits a trigger signal to the safety relay through an optocoupler isolation circuit to execute a safety protection action.

[0047] The age-related degradation factor represents the percentage reduction in the basic threshold of an elevator due to overall equipment aging, which occurs every year of operation. The typical value range is 0.01 to 0.05, and it can be determined based on the mean time between failures (MTBF) of that elevator model. This is the load impact factor, representing the percentage increase in the base threshold for every 10% increase in average load rate. A typical value range is 0.02 to 0.10, and it can be determined by analyzing the relationship between historical load and fault data. For example, it can be taken as... =0.02, =0.05. This coefficient establishes the braking deceleration. (Unit: m / s²) and safety threshold increase The linear relationship between the two was determined through braking tests: within a safe range, braking with different decelerations was applied, and the maximum vibration disturbance amplitude caused by the slack of the wire rope during braking was measured. This disturbance amplitude showed an approximately linear relationship with the deceleration. This is the slope of the linear relationship, which can be obtained through linear regression. For example, in a certain test, =0.1 means that for every 1 m / s² increase in deceleration, the safety threshold is temporarily raised by 0.1 units.

[0048] This application further proposes deployment optimizations for lightweight neural network models, including: The model parameters are quantized and compressed, converting floating-point weights into fixed-point representations; A dedicated data buffer is established in the shared memory area of ​​the embedded chip for intermediate feature transfer between the time-domain feature extraction submodule and the time-series analysis submodule; The total number of model parameters is controlled within 70% of the available storage capacity of the embedded device.

[0049] Model parameters refer to the set of adjustable variables in a neural network used to represent the mapping relationship between inputs and outputs, implemented using floating-point or fixed-point numbers. Converting floating-point weights to fixed-point representation is to adapt to the computational characteristics of embedded environments, reducing storage space usage and improving computational efficiency to meet real-time requirements. A dedicated data buffer refers to a specific area allocated in shared memory, implemented using structures such as circular buffers or deques, to improve the efficiency of intermediate feature transfer and avoid latency issues caused by general memory management. The total number of model parameters is controlled within 70% of the available storage capacity of the embedded device. This limitation is based on the practical needs of embedded systems to accommodate multi-tasking, reserving sufficient memory space to ensure system stability.

[0050] The deployment optimization scheme employs multiple technical means to ensure the efficient and stable operation of the lightweight neural network model within the resource-constrained elevator control cabinet chip. First, quantization compression technology converts floating-point weights into fixed-point representations, reducing storage requirements while adapting to the chip's fixed-point computing capabilities, thus accelerating the inference process without affecting the accuracy of key feature recognition. Second, a dedicated data buffer is established in the shared memory area to address the intermediate features output by the time-domain feature extraction submodule and the real-time transmission requirements, avoiding the random access latency of general memory management, specializing the feature transmission path, improving the data exchange efficiency between submodules, and ensuring uninterrupted continuous analysis of wear trend features. Finally, by controlling the total number of model parameters to within 70% of the available storage capacity of the device, sufficient memory space is reserved for critical operations such as triggering safety protection actions, preventing system crashes due to memory overflow, thereby ensuring the reliability and real-time response capability of the monitoring method throughout the entire elevator operation cycle. These measures synergistically optimize the storage efficiency, data flow efficiency, and system stability of the model deployment, enabling the safety monitoring method to achieve accurate wear probability assessment under limited resources.

[0051] The lightweight neural network model is obtained through the following training process: A training dataset was constructed, drawing historical operational data from multiple elevators of the same model under normal operating conditions, known slight wear conditions, and severe wear conditions. The collected raw signals included wire rope vibration signals and motor drive current signals. Professional maintenance personnel labeled each data segment with corresponding wear status tags, such as normal, slight wear, and severe wear, based on maintenance records and manual inspection results. The raw data underwent the same preprocessing as online monitoring, including hardware filtering and short-time Fourier transform, to generate time-frequency feature map samples and their corresponding labels.

[0052] Secondly, the network structure and training objective are defined. A network architecture is used as the initial model, and the training objective is to minimize the difference between the predicted wear state and the true label. The cross-entropy loss function is used as the optimization objective, and the loss function formula is as follows: ,in, The true class probability distribution of the sample. The model predicts the wear probability distribution. Model training is then performed using stochastic gradient descent (SGD) or the Adam optimizer, with an initial learning rate of 0.001 and a batch size of 32. The dataset is divided into training, validation, and test sets in a 7:2:1 ratio. Forward propagation is performed on the training set to calculate the loss, and backpropagation updates the network weights. After each training epoch, the model performance is evaluated on the validation set, and the learning rate is adjusted or early stopping is implemented to prevent overfitting. Training continues until the model's accuracy on the validation set stabilizes and the loss function no longer decreases significantly, typically requiring 100-200 training epochs.

[0053] Finally, model optimization and deployment preparation are performed. The model after training convergence is pruned, removing redundant neurons in the hidden layers whose weights have absolute values ​​close to zero, compressing the number of hidden layer units in the recurrent neural network to 48. The model parameters are quantized, converting the 32-bit floating-point weights to an 8-bit fixed-point representation using linear quantization. The quantization formula is as follows:

[0054] The total number of quantized model parameters is controlled within 70% of the available storage capacity of the target embedded chip, and the accuracy loss of the quantized model is verified to be within an acceptable range through test sets. Furthermore, the above approach works closely with other characteristics of lightweight neural network models. For example, by reducing storage requirements through quantization compression, the model can run more efficiently on embedded chips, while the introduction of a dedicated data buffer further improves the efficiency of intermediate feature transfer, making the entire monitoring process smoother. This design not only solves the problems of excessive model parameter storage and low efficiency of intermediate feature transfer, but also effectively avoids deployment failures or real-time performance degradation due to insufficient resources, thus meeting the rapid response requirements of elevator safety systems.

[0055] This application further proposes a specific implementation method for the direct physical connection between the hardware-level gate control circuit and the inverter start / stop signal line in the above-mentioned elevator operation status safety monitoring method. A TTL level signal line is used to directly couple the inverter's control output port; a signal shaping module is integrated at the input of the gate control circuit to harden the rising and falling edges of the start / stop commands; the differential signal transmission mechanism is implemented through twisted-pair cable, with a common-mode rejection ratio of not less than 40dB.

[0056] TTL level signal lines refer to signal transmission lines that adopt a transistor-to-transistor logic level standard. They are implemented using single-ended or differential signal transmission to ensure that start / stop commands are quickly transmitted to the gate circuit in a standardized level format, avoiding delays and distortions caused by signal conversion. Signal shaping modules can be understood as circuit units used to optimize digital signal waveforms. Implemented through structures such as Schmitt triggers or monostable multivibrators, they eliminate signal jitter and noise interference, making level transition edges steeper and clearer, thus accurately capturing transient command changes. Differential signal transmission mechanisms utilize two signal lines to transmit signals with opposite phases, implemented through twisted pairs or shielded cables. They cancel external electromagnetic interference and isolate common-mode noise, thereby improving signal integrity.

[0057] By directly coupling the inverter's control output port with TTL level signal lines, instantaneous drive of start / stop commands is achieved, ensuring real-time response capability for filter mode switching. Simultaneously, a signal shaping module is integrated at the gate control circuit input, effectively eliminating signal jitter and noise interference by hardening the rising and falling edges of start / stop commands, preventing erroneous switching due to line fluctuations. Furthermore, a differential signal transmission mechanism is implemented using twisted-pair cables, with a common-mode rejection ratio (CMRR) set to no less than 40dB. The balanced transmission characteristics of twisted-pair cables are used to cancel external electromagnetic interference, significantly improving signal integrity. These technologies work together to ensure stable execution of filter mode switching under complex operating conditions, ultimately achieving reliable noise suppression during transient phases, solving the problem of signal transmission being susceptible to line noise interference, and further enhancing the system's anti-interference capability and real-time performance through hardware-level optimization, providing more reliable technical support for the safe monitoring of elevator operation.

[0058] This application further proposes an adaptive window length adjustment mechanism for the temporal feature extraction submodule, including: The speed feedback signal from the elevator encoder is analyzed in real time, and the convolution window length is automatically extended when the running speed is below 0.5m / s. During high-speed operation, shorten the window length; The window length has a non-linear inverse relationship with the running speed, and dynamic matching is achieved through a preset mapping function.

[0059] An elevator encoder is a sensor device used to monitor the real-time speed of an elevator. It is implemented through photoelectric encoders or magnetoelectric encoders, providing accurate speed feedback signals as the basis for window adjustment. The convolution window length refers to the time span covered when performing sliding convolution operations on the vibration signal during time-domain feature extraction. It is dynamically adjusted according to a preset speed range using piecewise functions or lookup tables to adapt to the signal characteristic requirements under different operating conditions.

[0060] By analyzing the speed signal fed back from the elevator encoder in real time, a dynamic correlation mechanism between window length and operating speed is established. When the elevator is operating at a low speed (below 0.5 m / s), the system automatically extends the convolution window length. This design is specifically tailored to the vibration signal characteristics during elevator start-up, braking, or low-speed crawling, accumulating more time-series data points and effectively enhancing the ability to capture the progressive wear characteristics of the wire rope. During high-speed operation, the system automatically shortens the window length. This not only reduces the amount of data processed per convolution cycle and improves the real-time response speed of feature extraction, but also suppresses the high-frequency noise aliasing effect caused by an excessively long window. The nonlinear inverse relationship between window length and operating speed is dynamically matched through a preset mapping function. This design fully considers the nonlinear coupling relationship between speed and vibration characteristics during elevator operation, enabling the window length to accurately adapt to the signal variation patterns in different speed ranges.

[0061] Based on this, the proposed scheme is closely integrated with the time-domain feature extraction process in the aforementioned elevator operation status safety monitoring method. By establishing a dynamic correlation mechanism between speed feedback and window length, the adaptability problem of vibration signal feature extraction under varying elevator operating conditions is solved, avoiding feature distortion caused by fixed windows or coarse adjustments, thus laying a reliable foundation for subsequent wear trend identification.

[0062] This application further proposes an optimization of the forget gate mechanism in the recurrent neural network architecture, including: introducing learnable forget weight coefficients to dynamically adjust the retention ratio of historical states according to the elevator operation stage; reducing the dependence on historical states in the transient stage; and increasing the fusion weight of historical states in the steady-state stage.

[0063] Learnable forgetting weight coefficients refer to parameters that can be automatically adjusted through training data. Implemented using a gradient descent-based optimization algorithm, they aim to free the system from the constraints of preset fixed parameters, thereby adapting to individual differences and long-term performance degradation in different elevator equipment. The elevator operation phase refers to the state division of the elevator under different operating conditions such as starting, braking, and normal operation, identified by real-time monitoring of inverter start / stop signals or encoder feedback signals. The retention ratio of historical states can be understood as the degree of influence of the previous hidden layer output on the current state in a recurrent neural network. Specifically, this can be achieved by using a sigmoid function to map the forgetting weight coefficients to the 0-1 range, thereby controlling the degree of retention of historical information.

[0064] By reconstructing the decision logic of the forget gate mechanism, the core contradiction of mismatch between historical data processing and operating condition changes in elevator operation status monitoring is resolved. During actual operation, when the elevator is in the transient phase of starting or braking, the system automatically reduces its reliance on historical states. This mechanism quickly weakens the contamination of current analysis by instantaneous noise generated by mechanical impact, avoiding misjudging transient disturbances as abnormal wear of the wire rope. In the steady-state operation phase, the system increases the fusion weight of historical states. This strategy allows the neural network to continuously accumulate weak, progressive wear characteristics in vibration and current signals, thereby identifying early fault symptoms that are difficult to capture using traditional thresholding methods. This differentiated processing mechanism based on operating phases ensures both the anti-interference capability of transient processes and enhances the sensitivity of steady-state analysis.

[0065] Meanwhile, this solution works seamlessly with the deployment optimization of the aforementioned lightweight neural network model. By quantizing and compressing the model parameters and establishing a dedicated data buffer in the embedded chip, the computational efficiency of the optimized forget gate mechanism is ensured. Furthermore, the dynamic correction method for the safety threshold, combined with a two-layer decision logic, further improves the accuracy of wear trend identification, ultimately achieving optimization across the entire chain from the signal acquisition front-end to the analysis and decision-making back-end.

[0066] This application further proposes a dynamic correction method for the safety threshold based on a two-layer decision-making logic, including: the base threshold is periodically calibrated based on the elevator's service life and load history; during the startup phase, the safety threshold decays over time according to a preset curve: the initial value increases to 1.5 times the base threshold, and then linearly returns to the base threshold after 0.5 seconds; during the braking phase, the increase in the safety threshold is positively correlated with the braking deceleration, and is dynamically calculated by monitoring the deceleration parameters fed back by the frequency converter in real time.

[0067] The base threshold refers to a benchmark reference value adjusted based on the elevator's long-term operating status. It is implemented using a weighted algorithm based on multi-dimensional data such as usage time and load frequency distribution. This ensures the threshold dynamically reflects the actual health status of the equipment, avoiding misjudgments caused by fixed thresholds. The dynamic decay mechanism of the safety threshold can be understood as a piecewise function control strategy, defined by a preset time constant and decay slope, adapting to the natural decay law of mechanical impact energy during the start-up phase. Furthermore, the safety threshold adjustment mechanism during the braking phase is an adaptive algorithm based on physical quantity feedback. It establishes a mapping relationship between deceleration and threshold increase, precisely matching the impact of braking intensity on signal interference.

[0068] By correlating a base threshold with the elevator's historical operating data, dynamic compensation for equipment aging effects is achieved. During startup, raising the initial threshold effectively shields against transient mechanical shock noise, while the subsequent linear decay process ensures the system can quickly regain sensitivity to progressive wear characteristics after the impact weakens. During braking, by introducing the deceleration parameter fed back from the frequency converter as a dynamic input, the adjustment of the safety threshold can respond in real-time to changes in braking intensity, solving the problem of false triggering caused by mechanical shock or wire rope slack during startup and braking. Furthermore, by combining it with the wear probability value output mechanism in the aforementioned elevator operating status safety monitoring method, the accuracy of fault warnings is improved. Especially under complex operating conditions, this scheme can flexibly adjust the threshold according to the characteristics of different operating stages, thereby improving detection reliability while ensuring real-time performance.

[0069] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for monitoring the safety status of an elevator, characterized in that, Includes the following steps: Vibration signals of the elevator wire rope and motor drive current signals are collected in real time using vibration sensors and current sensors. Based on the start and stop command signals of the elevator frequency converter, the filtering mode is dynamically switched through the hardware-level gate control circuit: when the elevator start or stop command is detected, it automatically switches to the high-frequency filtering mode to suppress mechanical impact noise; when the elevator is in the steady-state operation stage, it switches to the low-frequency filtering mode to capture progressive wear characteristics. The filtered vibration signal is input into the time-domain feature extraction submodule of the lightweight neural network model, and the vibration signal is convolved to extract the time-domain impact features. The current signal is synchronously converted into spectral data and fused with the time-domain impulse characteristics to generate a time-frequency feature map. The time-frequency feature map is input into the time-series analysis submodule of the lightweight neural network model, and wear trend features are identified through recurrent neural network units; Based on the wear trend characteristics, a wear probability value is output, and a safety protection action is triggered when the wear probability value exceeds a preset safety threshold. The hardware-level gate control circuit and the inverter start / stop signal line are directly connected physically, without the need for central processing unit intervention; the lightweight neural network model is deployed in the embedded chip of the elevator control cabinet.

2. The elevator operation status safety monitoring method according to claim 1, characterized in that, The dynamic switching filtering modes of the hardware-level gating circuit specifically include: During the transient phase triggered by the elevator start command, the gate control circuit activates the high-pass filter characteristic to attenuate mechanical impact noise in the 0-50Hz frequency band, while increasing the sampling density of the vibration signal. During the transient phase triggered by the elevator braking command, the gate control circuit activates the low-pass filter characteristic to suppress high-frequency interference caused by wire rope slack. The switching of the filtering mode is directly driven by the level transition of the inverter start / stop signal, and line interference is eliminated through a differential signal transmission mechanism.

3. The elevator operation status safety monitoring method according to claim 1, characterized in that, The temporal feature extraction submodule adopts a one-dimensional convolutional neural network architecture, and its specific operations include: The filtered vibration time series data is processed by sliding window convolution, and the window length is dynamically adjusted to adapt to the elevator running speed. Local temporal features of vibration signals are extracted using multi-layer convolutional kernels; The output of the convolutional neural network is compressed by the pooling layer to generate a temporal feature vector, which is then fused with the spectral data using a feature concatenation method.

4. The elevator operation status safety monitoring method according to claim 1, characterized in that, The time series analysis submodule adopts a recurrent neural network architecture, and its specific operations include: Using time-frequency feature maps as input sequences, historical state information is dynamically filtered through a forget gate mechanism. During the elevator start-up and stop phases, enhance sensitivity to the rate of spectral energy decay; The number of hidden layer units in the recurrent neural network is optimized through pruning to reduce computational complexity.

5. The elevator operation status safety monitoring method according to claim 1, characterized in that, The comparison mechanism between the output of the wear probability value and the safety threshold includes: A two-layer decision-making logic is constructed: the first layer triggers an initial warning based on the comparison between the wear probability value and the basic threshold; the second layer dynamically adjusts the threshold based on the elevator operation stage. When the elevator is in the starting or braking phase, the safety threshold is automatically increased, and the decision observation window is extended. The trigger signal for the safety protection action is transmitted to the safety relay through an optocoupler isolation circuit.

6. The elevator operation status safety monitoring method according to claim 1, characterized in that, The deployment optimization of the lightweight neural network model includes: The model parameters are quantized and compressed, converting floating-point weights into fixed-point representations; A dedicated data buffer is established in the shared memory area of ​​the embedded chip for intermediate feature transfer between the time-domain feature extraction submodule and the time-series analysis submodule; The total number of model parameters is controlled within 70% of the available storage capacity of the embedded device.

7. The elevator operation status safety monitoring method according to claim 2, characterized in that, The direct physical connection between the hardware-level gating circuit and the inverter start / stop signal line is specifically implemented as follows: The control output port of the frequency converter is directly coupled using a TTL level signal line. A signal shaping module is integrated at the input of the gate control circuit to harden the rising and falling edges of the start and stop commands; The differential signal transmission mechanism is implemented using twisted-pair cables, with a common-mode rejection ratio of not less than 40dB.

8. The elevator operation status safety monitoring method according to claim 3, characterized in that, The adaptive window length adjustment mechanism of the temporal feature extraction submodule includes: The speed feedback signal from the elevator encoder is analyzed in real time, and the convolution window length is automatically extended when the running speed is below 0.5m / s. During high-speed operation, shorten the window length; The window length has a non-linear inverse relationship with the running speed, and dynamic matching is achieved through a preset mapping function.

9. The elevator operation status safety monitoring method according to claim 4, characterized in that, The optimization of the forget gate mechanism in the recurrent neural network architecture includes: A learnable forgetting weight coefficient is introduced to dynamically adjust the retention ratio of historical states according to the elevator operation stage; During the transient phase, reduce the dependence on historical states; During the steady-state phase, the fusion weight of historical states is increased.

10. The elevator operation status safety monitoring method according to claim 5, characterized in that, The dynamic correction method for the safety threshold of the two-layer decision logic includes: The basic threshold is periodically calibrated based on the elevator's service life and load history; During the startup phase, the safety threshold decays over time according to a preset curve: the initial value increases to 1.5 times the base threshold, and then linearly returns to the base threshold after 0.5 seconds. During the braking phase, the increase in the safety threshold is positively correlated with the braking deceleration, which is dynamically calculated by monitoring the deceleration parameters fed back by the frequency converter in real time.

Images