Elevator safety intelligent monitoring method and system combining Internet of Things and AI

Abstract

The invention provides an elevator safety intelligent monitoring method and system combining the Internet of Things and AI, and the method comprises the steps: obtaining a real-time sensing data set collected by a plurality of Internet of Things sensors, and carrying out the preprocessing operation, so as to generate a standard sensing data set; the pre-trained elevator state analysis model is used for carrying out feature extraction and safety state evaluation on the standard sensing data set, the real-time running state of the target elevator can be comprehensively and accurately reflected by integrating multi-source sensing data, evaluation deviation caused by single data or non-standard data is avoided, and the safety of the elevator is improved. An equipment maintenance optimization strategy is determined according to the safety state evaluation result, and a maintenance instruction set is generated to trigger maintenance operation execution, so that intelligence, automation and precision of elevator safety monitoring and maintenance are realized, the probability of elevator safety accidents is effectively reduced, the safety and reliability of elevator operation are improved, and the safety and reliability of elevator operation are improved. The manual inspection cost and the maintenance blindness are reduced, and the elevator maintenance management efficiency and quality are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology. Specifically, it relates to an elevator safety intelligent monitoring method and system combining the Internet of Things and AI. Background Art

[0002] In the traditional elevator safety monitoring field, it mainly relies on manual regular inspections and simple sensor alarm mechanisms. Manual inspections have problems such as long time intervals and inability to monitor in real time, and potential safety hazards during elevator operation cannot be detected in a timely manner. And simple sensor alarm mechanisms usually can only issue alarms for a single abnormal situation and lack the ability to comprehensively analyze the overall operating state of the elevator.

[0003] For example, a single sensor may only be able to monitor a certain parameter of the elevator, such as speed, temperature, etc., and will alarm when the parameter exceeds the preset threshold. However, the safety condition of the elevator is a complex system problem affected by multiple factors, and a single sensor cannot comprehensively reflect the true operating state of the elevator. In addition, the data formats and time series collected by different types of sensors are often inconsistent, making it difficult to effectively integrate and analyze, resulting in inaccurate and untimely assessment of the elevator safety state, and thus unable to formulate reasonable equipment maintenance strategies, bringing greater risks to the safe operation of the elevator. Summary of the Invention

[0004] In view of the problems mentioned above, in combination with the first aspect of the present invention, embodiments of the present invention provide an elevator safety intelligent monitoring method combining the Internet of Things and AI. The method includes: Obtain a set of real-time perception data collected by multiple Internet of Things sensors installed on a target elevator. The set of real-time perception data includes multi-source perception data of the target elevator at different time nodes; Perform preprocessing operations on the set of real-time perception data to generate a set of preprocessed standard perception data. The preprocessing operations include data cleaning, format standardization, and time series alignment; Extract features from the set of standard perception data based on a pre-trained elevator state analysis model to generate a set of real-time fusion state features of the target elevator, and perform a safety state assessment on the set of real-time fusion state features to generate a safety state assessment result of the target elevator; Determine a corresponding equipment maintenance optimization strategy according to the safety state assessment result of the target elevator; Generate a set of maintenance instructions according to the equipment maintenance optimization strategy, and send the set of maintenance instructions to a target maintenance terminal to trigger the execution of maintenance operations.

[0005] In another aspect, an embodiment of the present invention further provides an elevator safety intelligent monitoring system combining the Internet of Things and AI, including a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor. The machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0006] Based on the above aspects, in the embodiment of the present invention, by obtaining a set of real-time perception data collected by multiple Internet of Things sensors, generating a set of standard perception data through preprocessing operations, and using a pre-trained elevator state analysis model to perform feature extraction and safety state evaluation on the set of standard perception data, it can comprehensively and accurately reflect the real-time operation state of the target elevator by integrating multi-source perception data, avoiding evaluation biases caused by single data or non-standard data. According to the safety state evaluation results, determine the equipment maintenance optimization strategy and generate a set of maintenance instructions to trigger the execution of maintenance operations, realizing the intelligence, automation and precision of elevator safety monitoring and maintenance, effectively reducing the probability of elevator safety accidents, improving the safety and reliability of elevator operation, reducing the cost of manual inspection and the blindness of maintenance, and enhancing the efficiency and quality of elevator maintenance management. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 is a schematic flowchart of the execution of the elevator safety intelligent monitoring method combining the Internet of Things and AI provided by an embodiment of the present invention.

[0008] Figure 2 is a schematic diagram of an exemplary hardware and software component of the elevator safety intelligent monitoring system combining the Internet of Things and AI provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0009] The present invention will be specifically described below with reference to the accompanying drawings of the specification. Figure 1 is a schematic flowchart of the elevator safety intelligent monitoring method combining the Internet of Things and AI provided by an embodiment of the present invention. The elevator safety intelligent monitoring method combining the Internet of Things and AI will be introduced in detail below.

[0010] Step S110: Obtain a set of real-time perception data collected by multiple Internet of Things sensors installed on the target elevator. The set of real-time perception data includes multi-source perception data of the target elevator at different time nodes.

[0011] In this embodiment, the target elevator is a specific elevator that needs to be monitored for safety and intelligence, and a plurality of different types of Internet of Things sensors are installed on it. These Internet of Things sensors can sense various operating state information of the target elevator at different time nodes. Specifically, the types of sensors include speed sensors for monitoring the running speed of the elevator, acceleration sensors for detecting changes in elevator acceleration, door state sensors for judging the opening and closing state of the elevator doors, temperature sensors for measuring the temperature inside the car, and humidity sensors for detecting the humidity inside the car, etc.

[0012] At each time node t, each sensor collects corresponding sensed data. Let the set of time nodes be T = {t1, t2, t3,...}. For each time node ti ∈ T, the data collected by different types of sensors constitutes a multi-source sensed data vector. For example, at time node t1, the running speed of the elevator collected by the speed sensor is denoted as v1, the acceleration collected by the acceleration sensor is denoted as a1, the door state information collected by the door state sensor is denoted as d1 (the door state can be represented by a binary value, such as 0 for the door closed and 1 for the door open), the temperature inside the car collected by the temperature sensor is denoted as θ1, and the humidity inside the car collected by the humidity sensor is denoted as . Combining these data, the multi-source sensed data vector X1 = [v1, a1, d1, θ1, at time node t1 is obtained.

[0013] As time goes by, the multi-source sensed data vectors at different time nodes are collected in sequence. Arranging the multi-source sensed data vectors at all time nodes in chronological order forms a real-time sensed data set X = {X1, X2, X3,...}, which completely records the multi-source sensed data situation of the target elevator at different time nodes.

[0014] Step S120: Perform a preprocessing operation on the real-time sensed data set to generate a preprocessed standard sensed data set. The preprocessing operation includes data cleaning, format standardization, and time series alignment.

[0015] In order to improve the quality and usability of the data for more accurate subsequent elevator state analysis, it is necessary to perform a preprocessing operation on the obtained real-time sensed data set. The preprocessing operation mainly includes three key steps: data cleaning, format standardization, and time series alignment, which will be described in detail below.

[0016] Step S121: Identify the data missing segments in the real-time sensed data set, and perform interpolation compensation processing on the data missing segments based on the sensed data of adjacent time nodes to generate a compensated continuous sensed data sequence.

[0017] In the real-time perception data set, due to sensor failures, communication interference, or other unforeseen factors, data missing may occur. The primary task of this step is to identify these data missing segments. For example, traverse each multi-source perception data vector in the real-time perception data set and check whether there are missing values in each element of the vector. For example, for the multi-source perception data vector Xj = [vj, aj, dj, θj, , if a certain element, such as the temperature value θj, is missing (which can be represented by a specific identifier, such as NaN), it indicates that the temperature data at this time node is missing.

[0018] After identifying the data missing segments, interpolation compensation processing needs to be performed based on the perception data of adjacent time nodes. Assume that the temperature data θj at time node tj is missing, and the temperature data at adjacent time nodes tj-1 and tj+1 are θj-1 and θj+1 respectively. The linear interpolation method can be used to compensate for the missing temperature value. The basic idea of linear interpolation is to assume that the data changes linearly between adjacent time nodes. The specific calculation formula is: θj = θj-1 + ((θj+1 - θj-1) * (tj - tj-1) / (tj+1 - tj-1)).

[0019] For other types of data missing, a similar linear interpolation method can also be used for compensation. For example, if the speed data vk is missing and the speed data at adjacent time nodes vk-1 and vk+1 are known, the compensated speed value vk can be calculated according to the same linear interpolation formula.

[0020] By performing interpolation compensation processing on all data missing segments in the real-time perception data set, a compensated continuous perception data sequence X' can be generated, and each multi-source perception data vector in it no longer contains missing values.

[0021] Step S122: Uniformly process the formats of the perception data collected by different sensors in the continuous perception data sequence, convert the perception data into a preset standard numerical format, and then perform Z-Score normalization processing to generate a normalized perception data sequence.

[0022] After generating the compensated continuous perception data sequence, since the data formats collected by different sensors may be different, for example, the speed data may be in meters per second, the temperature data in degrees Celsius, the humidity data in percentage, etc., in order to facilitate subsequent analysis and processing, the formats of these perception data need to be uniformly processed.

[0023] First, according to the requirements of the preset standard numerical format, the perception data collected by different sensors is converted. For example, physical quantity data such as speed data and acceleration data is converted into floating-point format, and discrete data such as door status data is converted into binary numerical format.

[0024] Next, the converted perception data is processed by Z-Score normalization. Z-Score normalization is a commonly used data standardization method, which can convert the data into a standard normal distribution with a mean of 0 and a standard deviation of 1. For each multi-source perception data vector Xi' = [vi', ai', di', θi', in the continuous perception data sequence X', Z-Score normalization is performed on each element therein respectively.

[0025] Taking the speed data vi' as an example, let the mean of the speed data at all time nodes be μv and the standard deviation be σv, then the normalized speed data is: .

[0026] Similarly, for acceleration data ai', temperature data θi', humidity data etc., the same Z-Score normalization formula is also used for processing. For the door status data di', since it is binary data itself, no normalization processing is required.

[0027] By performing Z-Score normalization on all elements in the continuous perception data sequence, a normalized perception data sequence is generated. The data in this sequence has a unified format and a standardized numerical range, eliminating the influence of the dimension between different data types.

[0028] Step S123: Extract the timestamp information of each data point in the normalized perception data sequence, and perform time window alignment processing on the multi-source perception data according to the timestamp information to generate a time-synchronized multi-channel perception data set.

[0029] After obtaining the normalized perception data sequence, in order to further analyze the correlation between the data collected by different sensors, time synchronization processing of the multi-source perception data is required. This step first extracts the timestamp information of each data point in the normalized perception data sequence. The timestamp information records the specific time when each data point is collected, which is the key basis for time synchronization.

[0030] Then, time-window alignment processing is performed on the multi-source perception data according to the timestamp information. The purpose of time-window alignment is to match and integrate the data collected by different sensors within the same time range. The specific approach is to set a fixed time-window size Δt and divide the entire time range into multiple consecutive time windows. For example, with a 1-minute time-window size, the time range from t1 to tN is divided into multiple 1-minute time windows.

[0031] For each time window, the normalized perception data collected by different sensors within that window is integrated. Assume that within a certain time window [tj, tj+Δt], the normalized speed data collected by the speed sensor is , , …, the normalized acceleration data collected by the acceleration sensor is âj1, âj2, …, the door status data collected by the door status sensor is , , …, the normalized temperature data collected by the temperature sensor is , …, the normalized humidity data collected by the humidity sensor is , , … These data are arranged and integrated in chronological order to form a multi-channel perception data vector Yj = , , …, âj1, âj2, …, , , …, , …, , , …].

[0032] As the time window progresses, the multi-source perception data within each time window is integrated in turn, generating a time-synchronized multi-channel perception data set Y = {Y1, Y2, Y3, …}. Each multi-channel perception data vector in this set contains the normalized perception data collected by different sensors within the same time window, achieving the time synchronization of multi-source perception data.

[0033] Step S124: Perform noise filtering processing on the multi-channel perception data set to remove abnormal data points exceeding the preset fluctuation threshold, generating the preprocessed standard perception data set.

[0034] After generating the time-synchronized multi-channel perception data set, due to the errors of the sensors themselves or the interference of the external environment, there may be some noise and abnormal data points in the data. These abnormal data points may have an adverse impact on the subsequent elevator state analysis, so it is necessary to perform noise filtering processing on the multi-channel perception data set.

[0035] In this step, a method with a preset fluctuation threshold is adopted to identify and remove abnormal data points. For each multi-channel perception data vector in the multi-channel perception data set Y , each element therein is analyzed separately.

[0036] Taking the normalized speed data as an example, a preset fluctuation threshold δv is set. Calculate the difference between this element and its adjacent elements. If the absolute value of the difference exceeds the fluctuation threshold δv, then this element is considered an abnormal data point. For example, for , if and , then is marked as an abnormal data point.

[0037] For other types of normalized data, such as the normalized acceleration data âi, the normalized temperature data , the normalized humidity data , etc., also according to the same method, according to their respective preset fluctuation thresholds (such as δa, δθ, ), the identification of abnormal data points is carried out.

[0038] When the abnormal data points are identified, they are removed from the multi-channel perception data vector. For the multi-channel perception data vector after removing the abnormal data points, interpolation processing is performed according to the adjacent normal data points before and after to ensure the continuity of the data.

[0039] By performing noise filtering processing on all multi-channel perception data vectors in the multi-channel perception data set, removing the abnormal data points therein and performing interpolation compensation, a preprocessed standard perception data set Z is generated. The data in this standard perception data set has undergone a series of preprocessing operations such as data cleaning, format standardization, time series alignment, and noise filtering, and the data quality has been significantly improved.

[0040] Step S130: Based on the pre-trained elevator state analysis model, feature extraction is performed on the standard perception data set to generate the real-time fusion state feature set of the target elevator, and a safety state assessment is performed on the real-time fusion state feature set to generate the safety state assessment result of the target elevator.

[0041] After obtaining the preprocessed standard perception data set, the pre-trained elevator state analysis model is used to perform feature extraction and safety state assessment on it. The elevator state analysis model is a complex model trained with a large amount of historical data, which can extract the key features reflecting the running state of the target elevator from the standard perception data set and evaluate the safety state of the elevator according to these features. The specific processes of feature extraction and safety state assessment will be described in detail below.

[0042] Step S131: Input the standard perception data set into a pre-trained elevator state analysis model, and call the convolutional feature extraction layer of the elevator state analysis model to perform a convolutional operation on the multi-channel perception data to extract the local operation state features of the target elevator.

[0043] First, input the preprocessed standard perception data set Z into a pre-trained elevator state analysis model. The convolutional feature extraction layer of the elevator state analysis model is the first key module of the model, and its main function is to perform a convolutional operation on the multi-channel perception data to extract the local operation state features of the target elevator.

[0044] The convolutional operation extracts local features in the data by performing a sliding calculation on the multi-channel perception data using a convolutional kernel. For each multi-channel perception data vector Zi in the standard perception data set Z, the convolutional feature extraction layer performs a convolutional operation using multiple convolutional kernels with different sizes and parameters.

[0045] Assume that the size of the convolutional kernel is k×k. When performing a convolutional operation on the multi-channel perception data vector Zi, the convolutional kernel slides point by point on Zi, and each time it slides, it multiplies the convolutional kernel with the data at the corresponding position element by element and sums them to obtain a convolutional result value. As the convolutional kernel slides, a series of convolutional result values are obtained, and these convolutional result values form a feature map.

[0046] For different convolutional kernels, different feature maps are obtained. Combining the feature maps obtained from all convolutional kernels forms the local operation state feature set L of the target elevator. Each feature map in the local operation state feature set L reflects the operation state information of the target elevator within a certain local range, such as the speed change and acceleration fluctuation of the elevator in a short period of time.

[0047] Step S132: Based on the time series analysis layer of the elevator state analysis model, perform time dependence modeling on the local operation state features to generate a global state feature sequence reflecting the elevator operation trend.

[0048] After extracting the local operation state feature set L of the target elevator, it is necessary to further analyze the time dependence between these features to generate a global state feature sequence reflecting the elevator operation trend. This task is completed by the time series analysis layer of the elevator state analysis model.

[0049] Step S1321: Divide the local operation state features into multiple consecutive feature subsequences according to a time window, and each feature subsequence contains multi-channel feature vectors with a preset time step length.

[0050] First, divide the local running state feature set L according to time windows. Set a preset time step s, arrange the feature maps in the local running state feature set L in chronological order, and then divide them into multiple consecutive feature subsequences at intervals of s. For example, assume that the local running state feature set L contains N feature maps, namely L1, L2, …, LN. Taking the time step s = 5 as an example, the first feature subsequence S1 contains the feature maps L1, L2, L3, L4, L5, the second feature subsequence S2 contains the feature maps L6, L7, L8, L9, L10, and so on. Each feature subsequence contains multi-channel feature vectors of the preset time step, and these feature vectors reflect the local running state information of the target elevator over a period of time.

[0051] Step S1322: Call the bidirectional gated recurrent unit of the time series analysis layer to perform forward time propagation calculation on each feature subsequence, and generate the hidden state vectors at each time step during the forward propagation process.

[0052] Next, for each feature subsequence, call the bidirectional gated recurrent unit (Bi-GRU) of the time series analysis layer to perform forward time propagation calculation. The bidirectional gated recurrent unit is a neural network unit that can capture the time dependence in sequence data, and it controls the flow and memory of information through a gating mechanism.

[0053] Taking the feature subsequence Si as an example, during the forward time propagation calculation, the Bi-GRU starts from the first time step of the feature subsequence and processes the multi-channel feature vectors at each time step in turn. At each time step t, the Bi-GRU updates the hidden state vector at the current time step through a series of gating calculations based on the input feature vector at the current time step and the hidden state vector at the previous time step. Specifically, the Bi-GRU includes an update gate and a reset gate. The update gate is used to control how much information of the hidden state vector at the previous time step needs to be passed to the current time step, and the reset gate is used to control the degree of fusion between the input feature vector at the current time step and the hidden state vector at the previous time step.

[0054] After the forward time propagation calculation, for each time step t in the feature subsequence Si, a corresponding forward hidden state vector hFt is generated. Arranging the forward hidden state vectors at all time steps in the feature subsequence Si in chronological order, we obtain the forward hidden state vector sequence HF = {hF1, hF2, …, hFs} of this feature subsequence during the forward propagation process.

[0055] Step S1323: Synchronously call the bidirectional gated recurrent unit to perform reverse-time propagation calculation on each feature subsequence, generating the hidden state vectors at each time step during the reverse propagation process.

[0056] After completing the forward-time propagation calculation, synchronously call the bidirectional gated recurrent unit to perform reverse-time propagation calculation on each feature subsequence. The process of reverse-time propagation calculation is similar to that of forward-time propagation calculation, except for the opposite direction. Starting from the last time step of the feature subsequence, process the multi-channel feature vectors at each time step in turn.

[0057] Taking the feature subsequence Si as an example again, during the reverse-time propagation calculation, the Bi-GRU updates the hidden state vector at the current time step through gated calculation based on the input feature vector at the current time step and the hidden state vector at the next time step. After the reverse-time propagation calculation, for each time step t in the feature subsequence Si, a corresponding reverse hidden state vector hBt is generated. Arranging the reverse hidden state vectors of all time steps in the feature subsequence Si in reverse time order, we obtain the reverse hidden state vector sequence HB = {hB1, hB2,..., hBs} of this feature subsequence during the reverse propagation process.

[0058] Step S1324: Concatenate the forward hidden state vector and the reverse hidden state vector at the same time step in the feature dimension to generate a composite hidden state vector that fuses the bidirectional time dependence relationship.

[0059] After obtaining the forward hidden state vector sequence HF and the reverse hidden state vector sequence HB, concatenate the forward hidden state vector and the reverse hidden state vector at the same time step in the feature dimension. For each time step t in the feature subsequence Si, concatenate the forward hidden state vector hFt and the reverse hidden state vector hBt in the feature dimension to obtain a composite hidden state vector ht = [hFt; hBt].

[0060] In this way, the information captured by the forward-time propagation and the reverse-time propagation is fused, enabling the composite hidden state vector to simultaneously reflect the operating state information of the target elevator before and after this time step, thereby better capturing the bidirectional time dependence relationship in the sequence data. Arranging the composite hidden state vectors of all time steps in the feature subsequence Si in time order, we obtain the composite hidden state vector sequence H = {h1, h2,..., hs} of this feature subsequence.

[0061] Step S1325: Perform a moving average filtering process on the composite hidden state vector to eliminate short-term fluctuation noise and retain the trend component in the time series.

[0062] To further improve the quality of the composite hidden state vector sequence H, eliminate the short-term fluctuation noise that may exist in it, and retain the trend components in the time series, it is necessary to perform a moving average filtering process on the composite hidden state vector. Moving average filtering is a commonly used signal processing method that smooths data by calculating the average value of data within a certain time window.

[0063] Set a moving window size w. For each composite hidden state vector ht in the composite hidden state vector sequence H, centered on ht, take a total of w composite hidden state vectors before and after it (if ht is at the beginning or end of the sequence and there are less than w vectors, take as many vectors as possible), and calculate the average value of these vectors. Specifically, for ht, the vector after its moving average filtering is calculated as follows: Add the w composite hidden state vectors centered on ht, and then divide by w.

[0064] For example, when w is odd, the w composite hidden state vectors centered on ht are ht-(w-1) / 2,..., ht-1, ht, ht+1,..., ht+(w-1) / 2, then =(ht-(w-1) / 2+…+ht-1+ht+ht+1+…+ht+(w-1) / 2) / w.

[0065] By performing a moving average filtering process on each composite hidden state vector in the composite hidden state vector sequence H, the filtered composite hidden state vector sequence is obtained. The vectors in this sequence are smoother, reducing the influence of short-term fluctuation noise and better reflecting the trend changes in the operating state of the target elevator.

[0066] Step S1326: Input the filtered composite hidden state vector into the fully connected mapping layer of the time series analysis layer, and generate a dimensionality-reduced feature vector that matches the local operating state feature dimension through a non-linear activation function.

[0067] After obtaining the filtered composite hidden state vector sequence it is input into the fully connected mapping layer of the time series analysis layer. The role of the fully connected mapping layer is to perform non-linear transformation and dimensionality reduction processing on the filtered composite hidden state vector so that its dimension matches the dimension of the local operating state feature.

[0068] The fully connected mapping layer contains multiple neurons, and each neuron is connected to each element of the input filtered composite hidden state vector. When an input filtered composite hidden state vector When this happens, the fully connected mapping layer will perform a linear weighted summation on it, and then perform a non-linear transformation through a non-linear activation function. Common non-linear activation functions include the ReLU (Rectified Linear Unit) function, etc.

[0069] Suppose the fully connected mapping layer has m neurons. For the input filtered composite hidden state vector , each neuron will perform a weighted summation on the elements of to obtain an intermediate value. Then these intermediate values are input into the non-linear activation function, and after calculation, m output values are obtained. Combining these output values forms a dimensionality-reduced feature vector vt.

[0070] Through the processing of the fully connected mapping layer and the non-linear activation function, the high-dimensional information of the filtered composite hidden state vector is compressed and transformed to obtain a dimensionality-reduced feature vector that matches the dimension of the local operating state characteristics. For each vector in the sequence of filtered composite hidden state vectors , such processing is performed to obtain a sequence of dimensionality-reduced feature vectors V = {v1, v2,..., vs}.

[0071] Step S1327: Concatenate the dimensionality-reduced feature vectors in the order of time windows into a continuous time series to generate the global state feature sequence reflecting the elevator operation trend.

[0072] Finally, the dimensionality-reduced feature vectors in the dimensionality-reduced feature vector sequence V are concatenated in the order of time windows to form a continuous time series. Since each vector in the dimensionality-reduced feature vector sequence V corresponds to the elevator operation state information within a time window, after concatenating them in time order, it can fully reflect the operation trend of the target elevator over a period of time.

[0073] By sequentially concatenating the dimensionality-reduced feature vectors v1, v2,..., vs, the obtained continuous time series is the global state feature sequence G reflecting the elevator operation trend. The global state feature sequence G synthesizes the operation state information of the target elevator in multiple time windows and provides an important basis for subsequent environmental correlation analysis and anomaly detection.

[0074] Step S133: Invoke the environmental correlation module of the elevator state analysis model, perform a correlation analysis on the global state feature sequence and the preset environmental parameters, and generate the environmental correlation features of the target elevator under different environmental conditions.

[0075] After obtaining the global state feature sequence G reflecting the elevator operation trend, the environment correlation module of the elevator state analysis model is called to perform correlation analysis on the global state feature sequence G and the preset environment parameters. The preset environment parameters may include information such as the environmental temperature, humidity, and air pressure of the building where the elevator is located, and these environmental parameters may affect the operation state of the elevator.

[0076] The environment correlation module will first preprocess the preset environment parameters to make them match the dimension and dimension of the global state feature sequence G. For example, if the global state feature sequence G is normalized, then the environment parameters also need to be normalized.

[0077] Then, the environment correlation module will perform correlation calculations on each feature vector in the global state feature sequence G and the corresponding environment parameters. Specifically, for the i-th feature vector Gi in the global state feature sequence G and the corresponding environment parameter vector Ei, the environment correlation module will perform weighted splicing on them through a learnable weight matrix W. The splicing method can be to splice Gi and the weighted Ei in the feature dimension to obtain a new vector Ci.

[0078] The calculation process of Ci is as follows: First, multiply the environment parameter vector Ei by the weight matrix W to obtain the weighted environment parameter vector Ei', and then splice Gi and Ei' in the feature dimension, that is, Ci = [Gi; Ei'].

[0079] By performing such correlation calculations on each feature vector in the global state feature sequence G, a series of environment correlation feature vectors C = {C1, C2,...} are obtained. These environment correlation feature vectors reflect the operation state characteristics of the target elevator under different environmental conditions and provide important information for more comprehensively evaluating the safety state of the elevator.

[0080] Step S134: Identify the fluctuation pattern of the global state feature sequence through the anomaly detection layer of the elevator state analysis model, and generate the anomaly fluctuation feature of the target elevator, where the anomaly fluctuation feature carries the corresponding anomaly time interval.

[0081] Next, use the anomaly detection layer of the elevator state analysis model to identify the fluctuation pattern of the global state feature sequence G. The main purpose of the anomaly detection layer is to detect whether there are abnormal fluctuation patterns in the global state feature sequence G, and these abnormal fluctuation patterns may indicate potential safety hazards of the elevator.

[0082] The anomaly detection layer processes the global state feature sequence G using a sliding window approach. A sliding window size k is set and slides over the global state feature sequence G. For the feature vector sequence within each sliding window, the anomaly detection layer performs dynamic time warping (DTW) matching with the templates in the predefined anomaly pattern library.

[0083] Dynamic time warping is a method for measuring the similarity between two time series, which allows stretching and warping of the sequences on the time axis to find the best matching path. When performing DTW matching, the anomaly detection layer can calculate the similarity score between the feature vector sequence within the sliding window and each template in the predefined anomaly pattern library.

[0084] If the similarity score between the feature vector sequence within a certain sliding window and a certain template in the predefined anomaly pattern library exceeds the preset similarity threshold, it is considered that there is an abnormal fluctuation in the feature vector sequence within that sliding window. Record the time interval corresponding to this abnormal fluctuation, as well as the abnormal fluctuation feature obtained after a certain processing of the feature vector sequence within that sliding window.

[0085] By performing such processing on all sliding windows of the global state feature sequence G, the abnormal fluctuation feature set A of the target elevator can be generated, and each abnormal fluctuation feature carries the corresponding abnormal time interval. These abnormal fluctuation features and abnormal time intervals provide an important basis for subsequent safety state assessment.

[0086] Step S135: Integrate the local operation state features, global state feature sequence, environment correlation features, and abnormal fluctuation features to generate the real-time integrated state feature set.

[0087] After obtaining the local operation state feature set L, global state feature sequence G, environment correlation feature set C, and abnormal fluctuation feature set A respectively, they need to be integrated to generate the real-time integrated state feature set R.

[0088] The integration method can be concatenation in the feature dimension. First, ensure that the dimensions and dimensions of the local operation state feature set L, global state feature sequence G, environment correlation feature set C, and abnormal fluctuation feature set A are unified. If there are non-uniform situations, corresponding preprocessing, such as normalization processing, etc., needs to be carried out.

[0089] Then, the feature maps in the local operation state feature set L, the feature vectors in the global state feature sequence G, the environment - related feature vectors in the environment - related feature set C, and the abnormal fluctuation feature vectors in the abnormal fluctuation feature set A are concatenated in sequence in the feature dimension. For example, for the local operation state feature vector Ll, the global state feature vector Gl, the environment - related feature vector Cl, and the abnormal fluctuation feature vector Al corresponding to a certain time point, they are concatenated to obtain a real - time fusion state feature vector Rl = [Ll; Gl; Cl; Al].

[0090] Such concatenation processing is performed on the feature vectors corresponding to all time points, and the real - time fusion state feature set R is obtained. The real - time fusion state feature set R synthesizes information from multiple aspects such as the local operation state, the global operation trend, the environment - related information, and the abnormal fluctuation situation of the target elevator.

[0091] Step S136: Perform a safety state assessment on the real - time fusion state feature set to generate a safety state assessment result of the target elevator.

[0092] After obtaining the real - time fusion state feature set R, it is necessary to perform a safety state assessment on it to determine the current safety state of the target elevator. This process is mainly completed through the multi - feature fusion fully - connected layer, the attention weight assignment layer, and the classifier layer of the elevator state analysis model.

[0093] Step S1361: Input the local operation state features, the global state feature sequence, the environment - related features, and the abnormal fluctuation features in the real - time fusion state feature set into the multi - feature fusion fully - connected layer of the elevator state analysis model, and map the multi - dimensional features to a unified feature space through a non - linear transformation to generate a fusion state vector.

[0094] First, input the local operation state features, the global state feature sequence, the environment - related features, and the abnormal fluctuation features in the real - time fusion state feature set R into the multi - feature fusion fully - connected layer of the elevator state analysis model. The role of the multi - feature fusion fully - connected layer is to fuse and transform these multi - dimensional features and map them into a unified feature space.

[0095] The multi - feature fusion fully - connected layer contains multiple neurons, and each neuron is connected to each feature element in the input real - time fusion state feature set R. When these features are input, the multi - feature fusion fully - connected layer will perform a linear weighted sum on them and then perform a non - linear transformation through a non - linear activation function. Common non - linear activation functions include the Sigmoid function, the Tanh function, etc.

[0096] Through the processing of the multi-feature fusion fully connected layer, the multi-dimensional features in the real-time fusion state feature set R are integrated and transformed to generate a fusion state vector F. The fusion state vector F contains multi-faceted operating state information of the target elevator in a unified feature space.

[0097] Step S1362: Invoke the attention weight assignment layer of the elevator state analysis model, calculate the cosine similarity between each feature dimension in the fusion state vector and the pre-trained abnormal pattern library based on the learnable query-key-value pair mechanism, generate attention weights reflecting the importance of potential abnormal features, and perform dynamic weighted recombination on the fusion state vector through the attention weights to generate an enhanced feature vector focused on abnormal sensitive dimensions.

[0098] Next, invoke the attention weight assignment layer of the elevator state analysis model to process the fusion state vector F. The attention weight assignment layer calculates the cosine similarity between each feature dimension in the fusion state vector F and the pre-trained abnormal pattern library based on the learnable query-key-value pair mechanism.

[0099] The pre-trained abnormal pattern library contains feature templates of various known elevator abnormal patterns. For each feature dimension in the fusion state vector F, the attention weight assignment layer uses it as a query vector to calculate the cosine similarity with each feature template (key vector) in the pre-trained abnormal pattern library. Cosine similarity is a method to measure the cosine value of the angle between two vectors, which can reflect the similarity degree between two vectors.

[0100] The cosine similarity scores between each feature dimension in the fusion state vector F and each feature template in the pre-trained abnormal pattern library are calculated. Then, based on these similarity scores, attention weights reflecting the importance of potential abnormal features are generated. Specifically, a learnable weight matrix can be used to transform the similarity scores to obtain the attention weight vector W.

[0101] Each element in the attention weight vector W corresponds to a feature dimension in the fusion state vector F, and its value represents the importance of this feature dimension in abnormal detection. Perform dynamic weighted recombination on the fusion state vector F through the attention weight vector W, that is, multiply each feature dimension in the fusion state vector F by the corresponding attention weight to obtain a new vector. Combine these new vectors to generate an enhanced feature vector E focused on abnormal sensitive dimensions. The enhanced feature vector E highlights the feature dimensions related to abnormal patterns more prominently, which helps to improve the accuracy of subsequent classification.

[0102] Step S1363: Input the enhanced feature vector into the classifier layer of the elevator state analysis model. Through the predefined safety description label mapping table in the classifier layer, match the enhanced feature vector with the multi-dimensional decision boundaries corresponding to each safety description label, and calculate the projection distances of the enhanced feature vector in the directions of each safety description label.

[0103] After obtaining the enhanced feature vector E, input it into the classifier layer of the elevator state analysis model. The main task of the classifier layer is to classify the safety state of the target elevator based on the enhanced feature vector E.

[0104] A safety description label mapping table is predefined in the classifier layer. This table contains various possible elevator safety state description labels and the corresponding multi-dimensional decision boundaries for each label. The multi-dimensional decision boundary is a region in a high-dimensional space used to distinguish different safety states.

[0105] For the enhanced feature vector E, the classifier layer matches it with the multi-dimensional decision boundaries corresponding to each safety description label. Specifically, calculate the projection distances of the enhanced feature vector E in the directions of each safety description label. The projection distance refers to the projection length of the enhanced feature vector E in the direction of the decision boundary corresponding to each safety description label. By calculating the projection distance, the closeness of the enhanced feature vector E to each safety description label can be measured.

[0106] Step S1364: According to the comparison result between the projection distance and a preset label determination threshold, select the top N safety description labels with the smallest projection distances as the candidate label set, and calculate the confidence probabilities of each candidate label through the softmax function of the classifier layer.

[0107] Based on the calculated projection distance, compare it with a preset label determination threshold. If the projection distance corresponding to a certain safety description label is less than the label determination threshold, then this label is considered a possible candidate label.

[0108] Select the top N safety description labels with the smallest projection distances as the candidate label set. N is a preset parameter used to control the number of candidate labels. Then, calculate the confidence probabilities of each candidate label through the softmax function of the classifier layer. The softmax function is a commonly used classification probability calculation function that can convert the projection distances of each candidate label into probability values, such that the sum of the confidence probabilities of all candidate labels is 1.

[0109] For each candidate label in the candidate label set, the softmax function calculates a confidence probability value according to its projection distance. The higher the confidence probability value, the more the enhanced feature vector E matches the safety state corresponding to this candidate label.

[0110] Step S1365: Sort the candidate tags from high to low based on the confidence probability, select all the security description tags whose confidence probability exceeds the preset probability threshold, and combine them to generate a security status evaluation result including the abnormal type identifier and the associated component information.

[0111] Finally, sort the candidate tags from high to low based on the calculated confidence probability of each candidate tag. Then, select all the security description tags whose confidence probability exceeds the preset probability threshold. The preset probability threshold is a pre-set parameter used to screen out the security description tags with higher confidence.

[0112] Combine the selected security description tags to generate a security status evaluation result S including the abnormal type identifier and the associated component information. The abnormal type identifier can clearly indicate the possible abnormal types of the target elevator, such as elevator speed abnormality, door failure, etc.; the associated component information can indicate the elevator components related to the abnormality, such as motors, door machines, etc.

[0113] Step S140: Determine the corresponding equipment maintenance optimization strategy according to the security status evaluation result of the target elevator.

[0114] After obtaining the security status evaluation result S of the target elevator, it is necessary to determine the corresponding equipment maintenance optimization strategy according to this result. This process mainly involves steps such as the generation of a maintenance requirement parameter set, the matching of maintenance operation items, priority evaluation, the generation and adjustment of a maintenance sequence, and resource conflict detection.

[0115] Step S141: Generate a maintenance requirement parameter set according to the abnormal type identifier and the associated component information in the security status evaluation result, in combination with the time-aligned sequence of the real-time operation data stream and the historical maintenance log of the target elevator.

[0116] First, generate a maintenance requirement parameter set P according to the abnormal type identifier and the associated component information in the security status evaluation result S, in combination with the time-aligned sequence of the real-time operation data stream and the historical maintenance log of the target elevator. The real-time operation data stream records the current operation status information of the target elevator, and the historical maintenance log records the past maintenance situation of the target elevator.

[0117] Perform an association analysis on the abnormal type identifier and the associated component information with the real-time operation data stream and the historical maintenance log. For example, if the security status evaluation result S indicates that the elevator speed is abnormal and the associated component is the motor, then the current operation parameters of the motor, such as rotational speed, current, etc., can be obtained from the real-time operation data stream, and the past maintenance records of the motor, such as maintenance time, replaced components, etc., can be obtained from the historical maintenance log.

[0118] Based on this information, a series of maintenance-related parameters are generated, such as the urgency of maintenance, the maintenance tools required, the parts that may need to be replaced, etc. Combining these parameters forms a set P of maintenance requirement parameters.

[0119] Step S142: Match the corresponding standard maintenance operation items from a preset maintenance knowledge base according to the abnormal type identifier, and extract the associated supplementary maintenance operation items in the maintenance knowledge base based on the associated component information.

[0120] After generating the set P of maintenance requirement parameters, according to the abnormal type identifier in the safety status evaluation result S, match the corresponding standard maintenance operation items from a preset maintenance knowledge base. The preset maintenance knowledge base is a database containing standard maintenance operations corresponding to various elevator abnormal types.

[0121] For each abnormal type identifier, there is a corresponding standard maintenance operation item in the maintenance knowledge base. For example, if the abnormal type identifier is elevator speed abnormality, the maintenance knowledge base may contain standard maintenance operation items such as checking the motor speed and adjusting the control system parameters.

[0122] At the same time, based on the associated component information in the safety status evaluation result S, extract the associated supplementary maintenance operation items from the maintenance knowledge base. The associated component information can help further determine the specific components related to the abnormality, so as to obtain the supplementary maintenance operation items for these components from the maintenance knowledge base. For example, if the associated component is the motor, then supplementary maintenance operation items such as checking the motor winding and lubricating the motor bearings can be extracted from the maintenance knowledge base.

[0123] Combine the matched standard maintenance operation items and the extracted supplementary maintenance operation items to form a preliminary set M of maintenance operation items.

[0124] Step S143: Evaluate the priority of the set of maintenance requirement parameters to generate the time sensitivity weights and execution order constraint conditions of each maintenance operation item.

[0125] Evaluating the priority of the set of maintenance requirement parameters is to determine the importance degree and execution order of each maintenance operation item. The set of maintenance requirement parameters contains various parameters related to maintenance, such as the urgency of maintenance, the maintenance tools required, the parts that may need to be replaced, etc. These parameters can reflect the priority of the maintenance operation items from multiple aspects.

[0126] First, consider the urgency of maintenance. If the safety status assessment result shows that there are serious safety hazards in the elevator and immediate maintenance is required, then the urgency of the corresponding maintenance operation item is relatively high. For example, if the abnormal type identifier is a fault in the elevator braking system, this kind of fault may cause the elevator to fail to stop normally, seriously threatening the safety of passengers. Therefore, the urgency of the maintenance operation item for the braking system is very high.

[0127] Secondly, consider the maintenance tools required and the components that may need to be replaced. If some maintenance operation items require special maintenance tools or need to replace key components, and the acquisition of these tools or components may take a certain amount of time, then the priorities of these maintenance operation items also need to be adjusted accordingly. For example, if the elevator motor needs to be replaced and the inventory of the motor is insufficient and needs to be ordered from the supplier, then the priority of the maintenance operation item for replacing the motor may be affected.

[0128] Based on these factors, a comprehensive evaluation is carried out on each maintenance operation item in the set of maintenance requirement parameters to generate the time sensitivity weight of each maintenance operation item. The time sensitivity weight reflects the urgency of each maintenance operation item in terms of time. The higher the weight, the more urgently the maintenance operation item needs to be executed.

[0129] At the same time, it is also necessary to determine the execution order constraint conditions between the maintenance operation items. There may be requirements for the order of execution between some maintenance operation items. For example, before replacing a certain component of the elevator, it may be necessary to disassemble and inspect it first. These execution order constraint conditions can be determined by analyzing the nature of the maintenance operation items and the structural principle of the elevator.

[0130] Step S144: Use a topological sorting algorithm to analyze the dependency relationship of the standard maintenance operation items and supplementary maintenance operation items, and generate an initial maintenance sequence that meets the execution order constraints.

[0131] After obtaining the time sensitivity weights and execution order constraint conditions of each maintenance operation item, use a topological sorting algorithm to analyze the dependency relationship of the standard maintenance operation items and supplementary maintenance operation items. The topological sorting algorithm is an algorithm used to sort a directed acyclic graph (DAG). It can arrange the nodes in the graph in order of precedence, so that for any directed edge (u, v) in the graph, node u is always arranged before node v.

[0132] Regard the standard maintenance operation items and supplementary maintenance operation items as the nodes in the graph, and regard the execution order constraint conditions between them as the directed edges in the graph. For example, if maintenance operation item A must be executed before maintenance operation item B, then there is a directed edge from node A to node B.

[0133] Sort the directed acyclic graph through a topological sorting algorithm to obtain an initial maintenance sequence that satisfies the execution order constraint. The maintenance operation items in the initial maintenance sequence are arranged in sequence, ensuring that before each maintenance operation item is executed, its preceding operation items have been completed.

[0134] Step S145: Dynamically adjust the priority of the initial maintenance sequence according to the time sensitivity weights to generate a maintenance operation sequence that takes into account both the dependency relationship and the weights.

[0135] After obtaining the initial maintenance sequence, it is also necessary to dynamically adjust its priority according to the time sensitivity weights of each maintenance operation item. Although the initial maintenance sequence satisfies the execution order constraint conditions, it may not fully consider the time urgency of each maintenance operation item.

[0136] Sort the maintenance operation items in the initial maintenance sequence from high to low according to the time sensitivity weights. During the sorting process, it is necessary to ensure that the execution order constraint conditions are still satisfied. For example, if the time sensitivity weight of maintenance operation item A is higher than that of maintenance operation item B, but maintenance operation item B is a preceding operation item of maintenance operation item A, then maintenance operation item B still needs to be ranked before maintenance operation item A.

[0137] Through dynamic priority adjustment, generate a maintenance operation sequence that takes into account both the dependency relationship and the weights. This maintenance operation sequence not only ensures the execution order between maintenance operation items but also preferentially arranges maintenance operation items with higher time sensitivity weights, improving the maintenance efficiency.

[0138] Step S146: Detect resource occupancy conflicts in the adjusted maintenance operation sequence, and generate a conflict-free maintenance operation sequence by dynamically adjusting the overlapping intervals of time windows and verifying the resource occupancy continuity constraint.

[0139] After generating a maintenance operation sequence that takes into account both the dependency relationship and the weights, it is necessary to detect resource occupancy conflicts in it. Maintenance operation items may occupy various resources, such as maintenance personnel, maintenance tools, elevator operation time, etc. If multiple maintenance operation items occupy the same resource at the same time, resource occupancy conflicts will occur.

[0140] First, analyze the resources required for each maintenance operation item and the resource occupancy time. For example, a certain maintenance operation item requires two maintenance personnel and specific maintenance tools, and needs to be carried out when the elevator is stopped, with a duration of a certain period.

[0141] Then, check the resource occupancy of each maintenance operation item in the maintenance operation sequence to see if there are resource occupancy conflicts. If conflicts are found, they need to be resolved by dynamically adjusting the overlapping intervals of the time windows. For example, if two maintenance operation items both need to use the same set of maintenance tools and their time windows overlap, then the time window of one of the maintenance operation items can be adjusted to stagger it from the time window of the other maintenance operation item.

[0142] During the process of adjusting the time windows, it is also necessary to verify the resource occupancy continuity constraint. Some resources need to be used continuously during the operation. For example, when a maintenance personnel is performing a certain maintenance operation, they cannot interrupt halfway to perform other operations. Therefore, when adjusting the time windows, it is necessary to ensure that the resource occupancy is continuous and there will be no interruption.

[0143] By detecting resource occupancy conflicts and dynamically adjusting the time windows, a conflict-free maintenance operation sequence is generated. The maintenance operation items in this conflict-free maintenance operation sequence will not conflict in terms of resource occupancy and meet the resource occupancy continuity constraint, ensuring the smooth progress of the maintenance operations.

[0144] Step S147: Generate a parallel execution instruction set according to the dependency relationships of the operation items in the conflict-free maintenance operation sequence to form the final equipment maintenance optimization strategy.

[0145] After obtaining the conflict-free maintenance operation sequence, generate a parallel execution instruction set according to the dependency relationships of the operation items in it. There may be no dependency relationships between some maintenance operation items, or the dependency relationships can be arranged reasonably to enable parallel execution.

[0146] Analyze the dependency relationships of the operation items in the conflict-free maintenance operation sequence to find out the combinations of operation items that can be executed in parallel. For example, if there is no dependency relationship between maintenance operation item A and maintenance operation item B, and the resources they require do not conflict, then they can be arranged to be executed in parallel at the same time.

[0147] For the combinations of operation items that can be executed in parallel, generate corresponding parallel execution instructions. The parallel execution instructions contain information such as the execution time and required resources of each operation item to ensure that each operation item can be coordinated during the parallel execution process.

[0148] Combine all the parallel execution instructions to form the final equipment maintenance optimization strategy. The final equipment maintenance optimization strategy comprehensively considers factors such as the dependency relationships of the maintenance operation items, the time sensitivity weights, and the resource occupancy situations, and can effectively improve the efficiency and quality of elevator maintenance.

[0149] Step S150: Generate a set of maintenance instructions according to the device maintenance optimization strategy, and send the set of maintenance instructions to the target maintenance terminal to trigger the execution of maintenance operations.

[0150] After determining the final device maintenance optimization strategy, it is necessary to generate a set of maintenance instructions according to this strategy and send them to the target maintenance terminal to trigger the execution of maintenance operations.

[0151] Step S151: Analyze the parallel execution instruction set in the device maintenance optimization strategy, and extract the execution time window parameters, target component identifiers, and associated resource allocation identifiers of each operation item.

[0152] First, analyze the parallel execution instruction set in the final device maintenance optimization strategy. The parallel execution instruction set contains detailed information about each maintenance operation item, such as execution time, required resources, etc.

[0153] Extract the execution time window parameters of each operation item from the parallel execution instruction set. The execution time window parameters specify the time range during which each operation item can be executed, including the start time and end time. This helps maintenance personnel reasonably arrange their working hours and ensure that maintenance operations can be completed on time.

[0154] At the same time, extract the target component identifiers of each operation item. The target component identifiers clarify the elevator components targeted by each maintenance operation item. For example, motors, door machines, control systems, etc. This enables maintenance personnel to accurately find the components that need to be maintained and improve maintenance efficiency.

[0155] In addition, it is also necessary to extract the associated resource allocation identifiers of each operation item. The associated resource allocation identifiers specify the resources required for each operation item, such as maintenance personnel, maintenance tools, spare parts, etc. Through the associated resource allocation identifiers, it can be ensured that each operation item can obtain the required resources and guarantee the smooth progress of maintenance operations.

[0156] Step S152: Verify the real-time available quantity of the corresponding maintenance resources from the maintenance resource scheduling database according to the associated resource allocation identifier. If the real-time available quantity is lower than the preset threshold, trigger a cross-terminal resource allocation instruction to supplement the resource gap and update the associated resource allocation identifier.

[0157] After extracting the associated resource allocation identifiers of each operation item, verify the real-time available quantity of the corresponding maintenance resources from the maintenance resource scheduling database according to these identifiers. The maintenance resource scheduling database records the current status and available quantity of all maintenance resources.

[0158] For each associated resource allocation identifier, look up the corresponding maintenance resource in the maintenance resource scheduling database and obtain its real-time available quantity. If the real-time available quantity of a certain maintenance resource is lower than the preset threshold, it indicates that the resource may not meet the requirements of the current maintenance operation and needs to be replenished.

[0159] When a resource gap is detected, trigger a cross-terminal resource allocation instruction. The cross-terminal resource allocation instruction can allocate resources from other available terminals to the current required terminal to replenish the resource gap. For example, if the available quantity of a certain maintenance tool is insufficient at the current terminal, the tool can be allocated from other terminals.

[0160] After completing the resource allocation, update the associated resource allocation identifier to ensure that it reflects the latest resource allocation situation. This can ensure that subsequent maintenance operations can accurately obtain the required resources.

[0161] Step S153: Invoke the predefined component maintenance instruction template library based on the target component identifier, map the operation logic of the operation item to the standard steps in the component maintenance instruction template library, and generate an initial instruction unit containing an operation code, an execution timing mark, and resource binding information.

[0162] After verifying the availability of the maintenance resources and updating the associated resource allocation identifier, invoke the predefined component maintenance instruction template library based on the target component identifier. The component maintenance instruction template library contains standard maintenance operation steps and instruction templates for different elevator components.

[0163] For each operation item, look up the corresponding standard steps in the component maintenance instruction template library according to its target component identifier. Then, map the operation logic of the operation item to the standard steps. For example, if the operation item is to inspect the motor, and the standard steps include specific steps such as inspecting the motor winding and inspecting the motor bearings, then map the operation logic of the operation item to these standard steps to determine the specific parameters of each standard step.

[0164] Through parameter mapping, generate an initial instruction unit containing an operation code, an execution timing mark, and resource binding information. The operation code is used to uniquely identify each operation item, facilitating subsequent management and tracking. The execution timing mark specifies the execution order of each operation step to ensure that the maintenance operation can be carried out in the correct order. The resource binding information binds each operation step to the required maintenance resources to ensure that the operation steps can obtain the required resource support.

[0165] Step S154: Perform parallel timestamp synchronization processing on the initial instruction unit according to the execution time window parameter, generate a discretized timestamp sequence that matches the time interval endpoints of each operation item, and perform timeline sorting on the initial instruction unit based on the timestamp sequence.

[0166] After generating the initial instruction unit, perform parallel timestamp synchronization processing on it according to the execution time window parameter. The execution time window parameter specifies the execution time range of each operation item. Through parallel timestamp synchronization processing, these time ranges can be converted into a discretized timestamp sequence.

[0167] First, determine the time interval endpoints of each operation item, that is, the start time and the end time. Then, generate a discretized timestamp sequence based on these time interval endpoints. The discretized timestamp sequence divides the timeline into multiple small time intervals, and each time interval corresponds to a timestamp.

[0168] Match the initial instruction unit with the discretized timestamp sequence to determine the position of each initial instruction unit on the timeline. Then, perform timeline sorting on the initial instruction unit based on the timestamp sequence, so that the initial instruction units are arranged in the order of execution time. This can ensure that maintenance operations can be carried out in the correct time order and avoid time conflicts.

[0169] Step S155: Dynamically bind the sorted initial instruction unit with the updated associated resource allocation identifier to generate a set of maintenance instructions including a timestamp consistency check code, a resource allocation status code, and an operation logic relationship chain.

[0170] After performing timeline sorting on the initial instruction unit, dynamically bind the sorted initial instruction unit with the updated associated resource allocation identifier. The purpose of dynamic binding is to ensure that each initial instruction unit can accurately obtain the required maintenance resources and can coordinate with other operation items during execution.

[0171] During the binding process, generate a set of maintenance instructions including a timestamp consistency check code, a resource allocation status code, and an operation logic relationship chain. The timestamp consistency check code is used to verify whether the timestamps of the instructions in the set of maintenance instructions are consistent, ensuring that maintenance operations can be carried out in the correct time order. The resource allocation status code reflects the resource allocation situation of each operation item, facilitating maintenance personnel to understand the usage status of resources. The operation logic relationship chain describes the logical relationship between each operation item, ensuring that maintenance operations can be carried out in the correct order and manner.

[0172] Step S156: Encode the maintenance instruction set into a data message stream according to a preset communication protocol through an encrypted transmission channel, and distribute it according to the address identifier of the target maintenance terminal and the message priority rule, triggering the target maintenance terminal to parse and execute the parallel operation items in the maintenance instruction set.

[0173] Finally, encode the maintenance instruction set into a data message stream according to a preset communication protocol through an encrypted transmission channel. The encrypted transmission channel can ensure the security of the maintenance instruction set during the transmission process and prevent data from being stolen or tampered with. The preset communication protocol stipulates the format and transmission rules of the data message stream to ensure that the maintenance instruction set can be accurately transmitted to the target maintenance terminal.

[0174] Distribute the data message stream according to the address identifier of the target maintenance terminal and the message priority rule. The address identifier of the target maintenance terminal is used to determine the sending target of the maintenance instruction set, and the message priority rule determines the priority of the data message stream according to factors such as the urgency of the maintenance operation.

[0175] When the target maintenance terminal receives the data message stream, it parses the maintenance instruction set therein and executes the parallel operation items. The target maintenance terminal performs maintenance operations according to the operation codes, execution timing marks, resource binding information, etc. in the maintenance instruction set to ensure that the maintenance work of the elevator can be successfully completed.

[0176] Step S210: Obtain a historical elevator operation data set, where the historical elevator operation data set includes historical perception data of multiple elevators and their corresponding safety status labels.

[0177] To train the pre-trained elevator status analysis model, it is first necessary to obtain a historical elevator operation data set. The historical elevator operation data set includes the historical perception data of multiple elevators over a past period of time and their corresponding safety status labels.

[0178] The historical perception data is collected by Internet of Things sensors installed on the elevator, including information such as the running speed, acceleration, door opening and closing status, temperature and humidity inside the car of the elevator. These data record the running status of the elevator at different time nodes and are important inputs for training the model.

[0179] The safety status label is used to label the safety status of the elevator at each time node. The safety status label can be divided into different categories such as normal status and abnormal status, and for the abnormal status, it can be further subdivided, such as speed anomaly, door failure, etc. The safety status label is the target output for training the model and is used to guide the model to learn how to identify the safety status of the elevator from the historical perception data.

[0180] By collecting the historical perception data of multiple elevators and their corresponding safety status labels, a large-scale historical elevator operation dataset is formed. This historical elevator operation dataset can cover elevators of different types and different operating environments, providing rich data samples for the training of the model.

[0181] Step S220: Perform sample augmentation processing on the historical elevator operation dataset to generate an augmented training dataset containing noise injection samples and time series perturbation samples.

[0182] After obtaining the historical elevator operation dataset, in order to improve the generalization ability and robustness of the model, it is necessary to perform sample augmentation processing on it. Sample augmentation processing can increase the diversity of training data, enabling the model to better adapt to different actual situations.

[0183] Step S221: Add random Gaussian noise to each perception data sequence in the historical elevator operation dataset to generate noise injection samples.

[0184] First, add random Gaussian noise to each perception data sequence in the historical elevator operation dataset. Random Gaussian noise is a random noise that conforms to the Gaussian distribution, with a mean of zero and a standard deviation that can be adjusted according to the actual situation.

[0185] For each perception data sequence in the historical elevator operation dataset, such as the elevator operation speed sequence, acceleration sequence, etc., add a randomly generated Gaussian noise value to each data point. For example, for a data point v in the elevator operation speed sequence, add a random Gaussian noise value ε to obtain a new data point v' = v + ε.

[0186] By adding random Gaussian noise to each perception data sequence, noise injection samples are generated. The noise injection samples simulate the noise interference that may occur when sensors collect data in actual situations, enabling the model to learn how to accurately identify the safety status of the elevator in a noisy environment.

[0187] Step S222: Randomly select some perception data sequences in the historical elevator operation dataset, and perform local segment displacement or scaling operations on the time axis of the selected perception data sequences to generate time series perturbation samples.

[0188] In addition to adding random Gaussian noise, it is also necessary to perform local segment displacement or scaling operations on the time axis of some perception data sequences in the historical elevator operation dataset to generate time series perturbation samples.

[0189] Randomly select some perception data sequences in the historical elevator operation dataset. For the selected perception data sequences, select a local segment on the time axis. Then, perform displacement or scaling operations on this local segment.

[0190] The displacement operation is to move a local segment forward or backward on the time axis by a certain time interval. For example, move a local segment in a certain perception data sequence from time t1 to t2 to time t3 to t4. The scaling operation is to stretch or compress a local segment on the time axis. For example, scale the time length of a local segment in a certain perception data sequence from t1 to t2 to t5 to t6.

[0191] By performing local segment displacement or scaling operations on the time axis for part of the perception data sequence, time series perturbation samples are generated. The time series perturbation samples simulate the possible temporal changes in the actual elevator operation state, enabling the model to learn how to accurately identify the safety state of the elevator under different time patterns.

[0192] Step S223: Combine the noise injection samples and time series perturbation samples with the original historical elevator operation data set to generate the enhanced training data set.

[0193] After generating the noise injection samples and time series perturbation samples, they are combined with the original historical elevator operation data set to generate the enhanced training data set. The enhanced training data set includes the original historical elevator operation data set, noise injection samples, and time series perturbation samples, and the diversity of the data is significantly improved.

[0194] It should be noted that the noise amplitude of the noise injection samples does not exceed the preset percentage threshold of the standard numerical format of the perception data, and the displacement or scaling range of the time series perturbation samples does not exceed the preset ratio threshold of the original time series length. This can ensure that the generated samples still have a certain degree of rationality and authenticity and will not have a negative impact on the training of the model.

[0195] Step S230: Construct an initial elevator state analysis model, where the initial elevator state analysis model includes a sample convolution feature extraction layer, a sample time series analysis layer, a sample environment association module, a sample anomaly detection layer, a sample multi-feature fusion fully connected layer, a sample attention weight assignment layer, and a sample classifier layer connected in sequence.

[0196] The initial elevator state analysis model consists of multiple modules and levels. The sample convolutional feature extraction layer is similar to the convolutional feature extraction layer in step S131. It performs a convolutional operation on the input sample multi-channel perception data to extract local features and generate a sample initial feature map. The sample time series analysis layer refers to step S132 and uses a bidirectional gated recurrent unit to model the time dependence of the sample initial feature map, capture the forward and backward dependence relationships, and generate a sample global state feature sequence. The sample environment association module refers to step S133, associates the sample global state feature sequence with a preset sample environment parameter matrix, and generates a sample environment association feature vector through weighted splicing. The sample anomaly detection layer refers to step S134, uses a sliding window and dynamic time warping matching to detect anomalies, and extracts sample anomaly fluctuation features and time interval markers. The sample multi-feature fusion fully connected layer normalizes and splices the sample initial feature map, the sample global state feature sequence, the sample environment association feature vector, and the sample anomaly fluctuation features, and then performs multi-layer non-linear transformation to generate a reduced-dimensional sample intermediate feature vector. The sample attention weight assignment layer refers to step S1362, calculates the cosine similarity between the sample intermediate feature vector and the pre-trained anomaly pattern library, generates a sample attention weight matrix, and generates a sample enhanced attention feature vector through weighted focusing. The sample classifier layer refers to steps S1363 - S1365, performs multi-label classification on the sample enhanced attention feature vector, and generates a sample prediction probability distribution of each safety state label through a fully connected mapping and a softmax activation function.

[0197] Step S240: Train the initial elevator state analysis model based on the enhanced training dataset to generate the pre-trained elevator state analysis model.

[0198] Use the enhanced training dataset to train the initial model and adjust the trainable parameters.

[0199] Step S241: Perform a multi-scale convolutional kernel sliding calculation on the sample multi-channel perception data in the enhanced training dataset through the sample convolutional feature extraction layer to generate a sample initial feature map.

[0200] Input the sample multi-channel perception data in the enhanced training dataset into the sample convolutional feature extraction layer, perform a sliding calculation with a multi-scale convolutional kernel, and generate a sample initial feature map with reference to step S131.

[0201] Step S242: Input the sample initial feature map into the sample time series analysis layer for bidirectional gated recurrent unit modeling, capture the forward and backward dependence relationships in the sample time series, and generate a sample global state feature sequence.

[0202] Input the sample initial feature map into the sample time series analysis layer, perform modeling with a bidirectional gated recurrent unit with reference to step S132, and generate a sample global state feature sequence.

[0203] Step S243: Invoke the sample environment association module to perform element-by-element weighted concatenation of the sample global state feature sequence and a preset sample environment parameter matrix, and adjust the influence coefficient of the sample environment parameters through a learnable weight matrix to generate a sample environment association feature vector.

[0204] The sample global state feature sequence is input into the sample environment association module, and element-by-element weighted concatenation is performed with a preset sample environment parameter matrix with reference to Step S133 to generate a sample environment association feature vector.

[0205] Step S244: Input the sample environment association feature vector into the sample anomaly detection layer for dynamic time warping matching within a sliding window, calculate the sample similarity scores with each template in a predefined anomaly pattern library, and extract the sample anomaly fluctuation features and their time interval markers that exceed the similarity threshold.

[0206] The sample environment association feature vector is input into the sample anomaly detection layer, and dynamic time warping matching within a sliding window is performed with reference to Step S134 to extract the sample anomaly fluctuation features and time interval markers.

[0207] Step S245: Perform min-max normalization on the sample initial feature map, sample global state feature sequence, sample environment association feature vector, and sample anomaly fluctuation features, and then concatenate them along the feature dimension to generate a sample multi-source fusion feature tensor with a unified dimension.

[0208] Perform min-max normalization on the above features and then concatenate them along the feature dimension to generate a sample multi-source fusion feature tensor with a unified dimension.

[0209] Step S246: Perform multi-layer nonlinear transformation on the sample multi-source fusion feature tensor through the sample multi-feature fusion fully connected layer to generate a reduced-dimensional sample intermediate feature vector.

[0210] The sample multi-source fusion feature tensor is input into the sample multi-feature fusion fully connected layer, and a reduced-dimensional sample intermediate feature vector is generated through multi-layer nonlinear transformation.

[0211] Step S247: Invoke the sample attention weight assignment layer to calculate the cosine similarity between the sample intermediate feature vector and each feature template in a pre-trained anomaly pattern library, generate a sample attention weight matrix according to the sample similarity distribution, and perform weighted focusing on the sample intermediate feature vector to generate a sample enhanced attention feature vector.

[0212] The sample intermediate feature vector is input into the sample attention weight assignment layer, calculate the cosine similarity with reference to Step S1362, generate a sample attention weight matrix, and perform weighted focusing to generate a sample enhanced attention feature vector.

[0213] Step S248: Input the sample enhanced attention feature vector into the sample classifier layer for multi-label classification boundary calculation, and generate the sample prediction probability distribution of each safety status label through fully connected mapping and the softmax activation function.

[0214] Input the sample enhanced attention feature vector into the sample classifier layer, and perform multi-label classification with reference to steps S1363 - S1365 to generate the sample prediction probability distribution.

[0215] Step S249: Based on the loss function value between the sample prediction probability distribution and the true safety status label, use the gradient descent algorithm to update the trainable parameters of each layer in the initial elevator state analysis model until the loss function value converges to a preset threshold, and generate the trained elevator state analysis model.

[0216] After obtaining the sample prediction probability distribution of each safety status label, calculate the loss function value between the sample prediction probability distribution and the true safety status label. The loss function is used to measure the difference between the prediction result of the model and the true result. Common loss functions include the cross-entropy loss function, etc.

[0217] According to the calculated loss function value, use the gradient descent algorithm to update the trainable parameters of each layer in the initial elevator state analysis model. The gradient descent algorithm is an optimization algorithm that calculates the gradient of the loss function with respect to the trainable parameters and then updates the trainable parameters in the opposite direction of the gradient, so that the loss function value gradually decreases.

[0218] After each update of the trainable parameters, calculate the loss function value again and determine whether it converges to the preset threshold. The preset threshold is a pre-set small value. When the loss function value is less than the preset threshold, it is considered that the model has converged and the training process ends.

[0219] After multiple iterations of updating the trainable parameters until the loss function value converges to the preset threshold, at this time, the trained elevator state analysis model is generated.

[0220] Among them, in each iteration, according to the calculated gradients, the gradient descent algorithm is used to update the trainable parameters of each layer. There are various variants of the gradient descent algorithm, such as Stochastic Gradient Descent (SGD), Momentum Stochastic Gradient Descent (MomentumSGD), Adagrad, Adadelta, Adam, etc. Different algorithms have different characteristics when updating the trainable parameters. Taking Stochastic Gradient Descent as an example, for each trainable parameter, subtract the learning rate multiplied by the gradient of this parameter from its current value to obtain the updated parameter value. The learning rate is a pre-set hyperparameter that controls the step size of each update. If the learning rate is too large, it may cause the initial elevator state analysis model to oscillate near the optimal solution and fail to converge; if the learning rate is too small, the convergence speed of the initial elevator state analysis model will be very slow.

[0221] In practical applications, it is necessary to consider data privacy protection and anti-disclosure issues. When collecting real-time perception data of the elevator, this data may contain some privacy-sensitive information. For example, the elevator usage time may indirectly reflect the activity patterns of the people in the building. To protect this privacy-sensitive data, various technical means can be adopted.

[0222] First, in the data collection stage, anonymize the data collected by the sensors. For example, for information related to people, replace it with anonymous identifiers. In this way, even if the data is obtained during transmission or storage, attackers cannot directly identify the specific personnel information.

[0223] Second, during the data transmission process, use encryption technology. Use symmetric encryption algorithms (such as AES) or asymmetric encryption algorithms (such as RSA) to encrypt the data to ensure that the data is in ciphertext form during transmission, and only authorized recipients can decrypt and use this data. At the same time, use a secure transmission protocol, such as the SSL / TLS protocol, to establish a secure communication channel to prevent the data from being stolen or tampered with during transmission.

[0224] In terms of data storage, adopt a secure storage system. Store the data in an encrypted database, and only authorized users can access it. At the same time, conduct regular security audits and vulnerability scans on the storage system to promptly discover and repair potential security hazards.

[0225] In addition, establish a strict access control mechanism. Only authorized personnel can access and process the perception data of the elevator. Through authentication and authorization management, ensure that only personnel with corresponding permissions can view, analyze, and use this data.

[0226] Figure 2FIG. 0 shows a schematic diagram of exemplary hardware and software components of an elevator safety intelligent monitoring system 100 that combines the Internet of Things and AI and can implement the ideas of the present application. For example, a processor 120 can be used on the elevator safety intelligent monitoring system 100 that combines the Internet of Things and AI and is used to execute the functions in the present application.

[0227] The elevator safety intelligent monitoring system 100 that combines the Internet of Things and AI can be a general-purpose server or a special-purpose server, both of which can be used to implement the method for elevator safety intelligent monitoring that combines the Internet of Things and AI in the present application. Although only one server is shown in the present application, for convenience, the functions described in the present application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0228] For example, the elevator safety intelligent monitoring system 100 that combines the Internet of Things and AI can include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and different forms of storage media 140, such as disks, ROM, or RAM, or any combination thereof. Exemplarily, the elevator safety intelligent monitoring system 100 that combines the Internet of Things and AI can also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The method of the present application can be implemented according to these program instructions. The elevator safety intelligent monitoring system 100 that combines the Internet of Things and AI also includes an I / O interface 150 between the computer and other input and output devices.

[0229] For ease of explanation, only one processor is described in the elevator safety intelligent monitoring system 100 that combines the Internet of Things and AI. However, it should be noted that the elevator safety intelligent monitoring system 100 that combines the Internet of Things and AI in the present application can also include multiple processors. Therefore, the steps executed by one processor described in the present application can also be jointly executed or separately executed by multiple processors. For example, if the processor of the elevator safety intelligent monitoring system 100 that combines the Internet of Things and AI executes step A and step B, it should be understood that step A and step B can also be jointly executed by two different processors or separately executed in one processor. For example, the first processor executes step A, the second processor executes step B, or the first processor and the second processor jointly execute steps A and B.

[0230] In addition, an embodiment of the present invention also provides a readable storage medium in which computer-executable instructions are preset. When the processor executes the computer-executable instructions, the above method for elevator safety intelligent monitoring that combines the Internet of Things and AI is implemented.

[0231] It should be noted that, in order to simplify the description of the present invention disclosure and thus help the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, sometimes multiple features are incorporated into one embodiment, drawing or description thereof.

Claims

1. An elevator safety intelligent monitoring method combining the Internet of Things and AI, characterized in that, The method includes: Obtaining a set of real-time perception data collected by multiple Internet of Things sensors installed in a target elevator, where the set of real-time perception data includes multi-source perception data of the target elevator at different time nodes; Performing a preprocessing operation on the set of real-time perception data to generate a preprocessed set of standard perception data, where the preprocessing operation includes data cleaning, format standardization, and time series alignment; Based on a pre-trained elevator status analysis model, extracting features from the set of standard perception data to generate a set of real-time fusion status features of the target elevator, and performing a safety status assessment on the set of real-time fusion status features to generate a safety status assessment result of the target elevator; Determining a corresponding device maintenance optimization strategy according to the safety status assessment result of the target elevator; Generating a set of maintenance instructions according to the device maintenance optimization strategy, and sending the set of maintenance instructions to a target maintenance terminal to trigger the execution of a maintenance operation.

2. The elevator safety intelligent monitoring method combining the Internet of Things and AI according to claim 1, characterized in that The performing a preprocessing operation on the set of real-time perception data to generate a preprocessed set of standard perception data includes: Identifying data missing segments existing in the set of real-time perception data, and performing interpolation compensation processing on the data missing segments based on the perception data of adjacent time nodes to generate a compensated continuous perception data sequence; Performing a unification process on the formats of the perception data collected by different sensors in the continuous perception data sequence, converting the perception data into a preset standard numerical format and then performing Z-Score normalization processing to generate a normalized perception data sequence; Extracting the timestamp information of each data point in the normalized perception data sequence, and performing time window alignment processing on the multi-source perception data according to the timestamp information to generate a set of time-synchronized multi-channel perception data; Performing noise filtering processing on the set of multi-channel perception data to remove abnormal data points exceeding a preset fluctuation threshold, and generating the preprocessed set of standard perception data.

3. The elevator safety intelligent monitoring method combining the Internet of Things and AI according to claim 1, characterized in that, The extracting features from the set of standard perception data based on a pre-trained elevator status analysis model to generate a set of real-time fusion status features of the target elevator includes: Inputting the set of standard perception data into a pre-trained elevator status analysis model, and invoking the convolutional feature extraction layer of the elevator status analysis model to perform a convolutional operation on the multi-channel perception data to extract the local operation status features of the target elevator; Based on the time series analysis layer of the elevator status analysis model, performing time dependence modeling on the local operation status features to generate a global status feature sequence reflecting the elevator operation trend; Invoking the environment association module of the elevator status analysis model, and performing association analysis on the global status feature sequence and preset environment parameters to generate environment association features of the target elevator under different environmental conditions; Performing fluctuation pattern recognition on the global status feature sequence through the anomaly detection layer of the elevator status analysis model to generate anomaly fluctuation features of the target elevator, where the anomaly fluctuation features carry corresponding anomaly time intervals; Fuse the local operation state features, global state feature sequence, environment correlation features, and abnormal fluctuation features to generate the real-time fusion state feature set; The time series analysis layer based on the elevator state analysis model performs time dependence modeling on the local operation state features to generate a global state feature sequence reflecting the elevator operation trend, including: Divide the local operation state features into multiple consecutive feature subsequences according to a time window, and each feature subsequence contains multi-channel feature vectors with a preset time step; Call the bidirectional gated recurrent unit of the time series analysis layer to perform forward time propagation calculation on each feature subsequence to generate hidden state vectors at each time step during the forward propagation process; Synchronously call the bidirectional gated recurrent unit to perform backward time propagation calculation on each feature subsequence to generate hidden state vectors at each time step during the backward propagation process; Perform feature dimension concatenation on the forward hidden state vector and the backward hidden state vector at the same time step to generate a composite hidden state vector that fuses bidirectional time dependence relationships; Perform a moving average filtering process on the composite hidden state vector to eliminate short-term fluctuation noise and retain the trend components in the time series; Input the filtered composite hidden state vector into the fully connected mapping layer of the time series analysis layer, and generate a dimensionality-reduced feature vector that matches the dimension of the local operation state features through a non-linear activation function; Concatenate the dimensionality-reduced feature vectors in the order of time windows into a continuous time series to generate the global state feature sequence reflecting the elevator operation trend.

4. The elevator safety intelligent monitoring method combining the Internet of Things and AI according to claim 3, characterized in that, Perform a safety state assessment on the real-time fusion state feature set to generate the safety state assessment result of the target elevator, including: Input the local operation state features, global state feature sequence, environment correlation features, and abnormal fluctuation features in the real-time fusion state feature set into the multi-feature fusion fully connected layer of the elevator state analysis model, and map the multi-dimensional features to a unified feature space through non-linear transformation to generate a fusion state vector; Call the attention weight assignment layer of the elevator state analysis model, calculate the cosine similarity between each feature dimension in the fusion state vector and the pre-trained abnormal pattern library based on the learnable query-key-value pair mechanism, generate attention weights reflecting the importance of potential abnormal features, and perform dynamic weighted recombination on the fusion state vector through the attention weights to generate an enhanced feature vector focusing on the abnormal sensitive dimension; Input the enhanced feature vector into the classifier layer of the elevator state analysis model, and match the enhanced feature vector with the multi-dimensional decision boundaries corresponding to each safety description label through the pre-defined safety description label mapping table in the classifier layer, and calculate the projection distance of the enhanced feature vector in the direction of each safety description label; According to the comparison result between the projection distance and the preset label determination threshold, select the top N safety description labels with the smallest projection distance as the candidate label set, and calculate the confidence probability of each candidate label through the softmax function of the classifier layer; Rank the candidate tags from high to low based on the confidence probability, and select all the security description tags whose confidence probability exceeds the preset probability threshold, and combine them to generate a security status evaluation result including the abnormal type identifier and the associated component information.

5. The elevator safety intelligent monitoring method combining the Internet of Things and AI according to claim 1, characterized in that, Determine the corresponding device maintenance optimization strategy according to the security status evaluation result of the target elevator, including: Generate a set of maintenance requirement parameters according to the abnormal type identifier and the associated component information in the security status evaluation result, combined with the time-aligned sequence of the real-time operation data stream and the historical maintenance log of the target elevator; Match the corresponding standard maintenance operation items from the preset maintenance knowledge base according to the abnormal type identifier, and extract the associated supplementary maintenance operation items in the maintenance knowledge base based on the associated component information; Evaluate the priority of the set of maintenance requirement parameters to generate the time sensitivity weight and the execution order constraint condition of each maintenance operation item; Use the topological sorting algorithm to analyze the dependency relationship of the standard maintenance operation items and the supplementary maintenance operation items, and generate an initial maintenance sequence that meets the execution order constraint; Dynamically adjust the priority of the initial maintenance sequence according to the time sensitivity weight to generate a maintenance operation sequence that takes into account the dependency relationship and the weight; Detect the resource occupancy conflict of the adjusted maintenance operation sequence, and generate a conflict-free maintenance operation sequence by dynamically adjusting the overlapping interval of the time window and verifying the resource occupancy continuity constraint; Generate a parallel execution instruction set according to the dependency relationship of each operation item in the conflict-free maintenance operation sequence to form the final device maintenance optimization strategy.

6. The elevator safety intelligent monitoring method combining the Internet of Things and AI according to claim 3, characterized in that, The training method of the pre-trained elevator state analysis model includes: Obtain a historical elevator operation data set, which includes the historical perception data of multiple elevators and their corresponding security status labels; Perform sample augmentation processing on the historical elevator operation data set to generate an augmented training data set including noise injection samples and time series perturbation samples; Construct an initial elevator state analysis model, which includes a sample convolution feature extraction layer, a sample time series analysis layer, a sample environment association module, a sample anomaly detection layer, a sample multi-feature fusion fully connected layer, a sample attention weight assignment layer, and a sample classifier layer connected in sequence; Train the initial elevator state analysis model based on the augmented training data set to generate the pre-trained elevator state analysis model.

7. The elevator safety intelligent monitoring method combining the Internet of Things and AI according to claim 6, characterized in that Training the initial elevator state analysis model based on the augmented training data set to generate the pre-trained elevator state analysis model includes: Perform multi-scale convolution kernel sliding calculation on the sample multi-channel perception data in the augmented training data set through the sample convolution feature extraction layer to generate a sample initial feature map; Input the sample initial feature map into the sample time series analysis layer for bidirectional gated recurrent unit modeling to capture the forward and backward dependency relationships in the sample time series, and generate a sample global state feature sequence; Call the sample environment association module to perform element-by-element weighted splicing on the sample global state feature sequence and a preset sample environment parameter matrix, and adjust the influence coefficient of the sample environment parameters through a learnable weight matrix to generate a sample environment association feature vector; Input the sample environment association feature vector into the sample anomaly detection layer for dynamic time warping matching within a sliding window, calculate the sample similarity scores with each template in the predefined anomaly pattern library, and extract the sample anomaly fluctuation features and their time interval markers that exceed the similarity threshold; Perform min-max normalization on the sample initial feature map, sample global state feature sequence, sample environment association feature vector, and sample anomaly fluctuation features, and then splice them along the feature dimension to generate a sample multi-source fusion feature tensor with a unified dimension; Perform multi-layer non-linear transformation on the sample multi-source fusion feature tensor through the sample multi-feature fusion fully connected layer to generate a reduced-dimensional sample intermediate feature vector; Call the sample attention weight assignment layer to calculate the cosine similarity between the sample intermediate feature vector and each feature template in the pre-trained anomaly pattern library, generate a sample attention weight matrix according to the sample similarity distribution, and perform weighted focusing on the sample intermediate feature vector to generate a sample enhanced attention feature vector; Input the sample enhanced attention feature vector into the sample classifier layer for multi-label classification boundary calculation, and generate the sample prediction probability distribution of each safety state label through full connection mapping and the softmax activation function; Based on the loss function value of the sample prediction probability distribution and the true safety state label, use the gradient descent algorithm to update the trainable parameters of each layer in the initial elevator state analysis model until the loss function value converges to a preset threshold, and generate a trained elevator state analysis model; 8. The elevator safety intelligent monitoring method combining the Internet of Things and AI according to claim 6, characterized in that, The sample enhancement processing of the historical elevator operation data set to generate an enhanced training data set including noise injection samples and time series perturbation samples includes: Adding random Gaussian noise to each perception data sequence in the historical elevator operation data set to generate noise injection samples; Randomly select some perception data sequences in the historical elevator operation data set, and perform local segment displacement or scaling operations on the time axis of the selected perception data sequences to generate time series perturbation samples; Merge the noise injection samples and time series perturbation samples with the original historical elevator operation data set to generate the enhanced training data set; Among them, the noise amplitude of the noise injection sample does not exceed the preset percentage threshold of the standard numerical format of the perception data, and the displacement or scaling range of the time series perturbation sample does not exceed the preset ratio threshold of the original time series length; 9. The elevator safety intelligent monitoring method combining the Internet of Things and AI according to claim 1, characterized in that, Generating a maintenance instruction set according to the equipment maintenance optimization strategy and sending the maintenance instruction set to the target maintenance terminal to trigger the execution of maintenance operations, including: Parsing the parallel execution instruction set in the equipment maintenance optimization strategy, and extracting the execution time window parameters, target component identifiers, and associated resource allocation identifiers of each operation item; Verify the real-time available quantity of the corresponding maintenance resources from the maintenance resource scheduling database according to the associated resource allocation identifier. If the real-time available quantity is lower than the preset threshold, trigger a cross-terminal resource allocation instruction to supplement the resource gap and update the associated resource allocation identifier; Call the predefined component maintenance instruction template library based on the target component identifier, perform parameter mapping on the operation logic of the operation item and the standard steps in the component maintenance instruction template library, and generate an initial instruction unit containing operation codes, execution time sequence marks, and resource binding information; Perform parallel timestamp synchronization processing on the initial instruction unit according to the execution time window parameter, generate a discretized timestamp sequence matching the time interval endpoints of each operation item, and perform timeline sorting on the initial instruction unit based on the timestamp sequence; Dynamically bind the sorted initial instruction unit to the updated associated resource allocation identifier to generate a maintenance instruction set containing timestamp consistency verification codes, resource allocation status codes, and operation logic relationship chains; Encode the maintenance instruction set into a data message stream according to the preset communication protocol through an encrypted transmission channel, and distribute it according to the address identifier of the target maintenance terminal and the message priority rule, triggering the target maintenance terminal to parse and execute the parallel operation items in the maintenance instruction set.

10. An elevator safety intelligent monitoring system combining the Internet of Things and AI, characterized in that, Comprising a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the elevator safety intelligent monitoring method combining the Internet of Things and AI according to any one of claims 1-9 above.

Images