Feature crossover elevator trapped people time series prediction model construction method based on deep learning, obtained model and prediction method

Abstract

The present invention relates to a method for constructing a feature cross elevator trapped time series prediction model based on deep learning, a obtained model and a prediction method. The construction method comprises composing a time series of historical information data, passing the input information into a single-layer fully connected layer to convert it into a primary abstract feature, feature interaction, passing it into an LSTM neural network and obtaining a comprehensive time series abstract feature, passing it into a fully connected layer and converting it into a predicted probability of trapped people, calculating the prediction loss with a cross entropy loss function, and updating the LSTM neural network parameters by back propagation to train the model. The prediction method comprises obtaining a plurality of prediction models by adopting a construction method, performing comprehensive processing using a plurality of prediction models with slightly different focuses and obtaining a comprehensive probability of trapped people as a prediction result. The present invention can extract the time series features of the elevator using only the historical record data of the elevator, and predict the probability of trapped people at the next time point in the elevator, under the premise of lacking real-time operation data of the elevator components.

Description

Technical Field

[0001] The present invention relates to a method for constructing a feature-crossing elevator entrapment time series prediction model based on deep learning, a prediction model obtained by the construction method, and a prediction method using the model, belonging to the technical fields of fault prediction, artificial neural networks, and time series prediction. Background Art

[0002] Equipment failure generally refers to an event or phenomenon in which the equipment loses or reduces its specified function, which is manifested as abnormal production and operation of the equipment, or some parts of the equipment lose their original accuracy or performance, causing the equipment to fail to operate normally and technical performance to decrease, thus affecting normal production and causing loss of corporate profits. In the case of elevator equipment, the most prominent manifestation of failure is the incident of people being trapped due to aging of components and failure of control components.

[0003] There are many methods for fault prediction, the most common ones are model-based fault prediction methods and data-driven fault prediction methods.

[0004] The model-based fault prediction method assumes that a mathematical model can be used to simulate the operation of the object system. This method evaluates the degree of damage to key components by calculating functional damage, and is used to evaluate the remaining life of components by establishing physical models or random process modeling. Usually, the fault characteristics of the object system are closely related to the parameters of the model used. As the research on the evolution mechanism of equipment or system failures gradually deepens, the model can be gradually corrected and adjusted to improve its prediction accuracy.

[0005] However, since the data currently available is sparse historical data and there is no real-time operation data of elevator components, it is difficult to use model-based fault prediction methods to predict the probability of elevator entrapment. When it is difficult to obtain detailed operation data for mathematical modeling of equipment systems, long-term records of equipment failures can be used to perform probability analysis on historical fault data and predict the probability of equipment failure using data-driven fault prediction methods. Artificial neural networks are one of the methods.

[0006] Artificial neural networks simulate the operating mechanism of human brain neurons, use computers to simulate the transmission of brain signals, have the ability to imitate continuous nonlinear functions, and can learn from samples. According to the task design and the design of the neural network structure, the neural network can learn the abstract features of the input data and complete tasks such as classification and prediction. However, neural network training requires a large number of data samples, and there are shortcomings such as slow convergence, local minima, and difficult to determine network structure.

[0007] Time series forecasting is to organize the historical data of the forecast object into a time series according to a certain time period, establish a model of the object's change over time based on the characteristic changes between different time periods, and analyze the object's change trends in the short and long term in the future.

[0008] As the interaction point between artificial neural networks and time series prediction, LSTM neural networks can combine the advantages of both parties and have a stronger ability to fit complex inputs. There is a certain gap between the input information and the output probability of people trapped in elevators. We can try to use LSTM neural networks to solve the conversion problem between input and output. Therefore, it is urgent to develop a feature-crossing elevator trapped time series prediction model and prediction method based on deep learning.

[0009] After searching, it was found that the invention patent with application number CN202110977026.1 and application publication number CN113734928A disclosed a neural network-based method for predicting elevator faults in use. Sensors are used to collect elevator operation data, and an elevator fault prediction model is established based on the collected data. After the sensor collects the signal, it is transmitted to the background processing system through the 4G module in the collection device. The background data processing device uses a neural network model to effectively predict the type of faults that may occur in the elevator in the future based on the elevator operation status parameters. However, the existing technology represented by this technical solution cannot achieve effective prediction based solely on the historical record data of the elevator.

[0010] In addition, the invention patent with application number CN202010861302.3 and application publication number CN112001548A discloses a deep learning-based OD passenger flow prediction method. Although this type of prediction method also uses deep learning technology, it does not teach how to improve the technology to achieve effective prediction based only on the historical record data of the elevator. Summary of the invention

[0011] The main purpose of the present invention is to overcome the problems existing in the prior art and provide a method for constructing a feature-crossing elevator entrapment time series prediction model based on deep learning. The model constructed by this method and the prediction method using the model make full use of the historical record data of the elevator to obtain high-order information between elevator features, thereby realizing the prediction of the probability of entrapment in the elevator at the next time point.

[0012] The technical solution of the present invention to solve the technical problem is as follows:

[0013] A method for constructing a feature-crossing elevator trapped person time series prediction model based on deep learning, characterized in that it includes the following steps:

[0014] Step 1: The deep learning model organizes the historical information data of the elevator into time series for input at each time point; the dimension of the input tensor is [batch_size, seq_len, features], where batch_size is the number of time series input into the deep learning model at a time, seq_len is the length of the time series, and features is the number of features contained in the time series at each time point;

[0015] The second step is to pass the input information into a single fully connected layer and convert it into primary abstract features;

[0016] The third step is to perform feature interaction on the primary abstract features themselves to obtain feature interaction results;

[0017] Step 4: Pass the feature interaction results into the LSTM neural network and obtain comprehensive temporal abstract features;

[0018] Step 5: Pass the comprehensive temporal abstract features into two fully connected layers and convert them into predicted probability of being trapped;

[0019] Step 6: Pass the predicted probability of people being trapped and the actual value of people being trapped into the cross entropy loss function, calculate the prediction loss, and update the LSTM neural network parameters through back propagation to train the model; the actual value of people being trapped comes from the historical information data of the elevator; the prediction model construction is completed.

[0020] This construction method inputs the historical information data of the elevator into the deep learning model, and then passes through a single fully connected layer, a feature cross layer and an LSTM neural network in sequence. The features are further abstracted, and finally the fully connected layer predicts the probability of people being trapped in the elevator at the next time point. After the model is trained, it can be put into actual deployment.

[0021] The technical solution further improved by the present invention is as follows:

[0022] Preferably, in the first step, the historical information data is selected from or includes the model, speed limit, weight limit, location, manufacturing unit, weather, temperature, entrapment record, and fault record of the elevator.

[0023] By adopting the above preferred scheme, the specific technical features of the first step can be further optimized.

[0024] Preferably, in the second step, the conversion is performed according to the following formula:

[0025] f x =W FC ·f input +b FC I

[0026] In Formula I, W FC is the weight matrix of the fully connected layer, f inputFor input information, b FC is the bias of the fully connected layer, f x It is a primary abstract feature.

[0027] More preferably, by setting the matrix W FC The number of columns f determines the primary abstract feature f x characteristic length.

[0028] By adopting the above preferred scheme, the specific technical features of the second step can be further optimized.

[0029] Preferably, in the third step, the specific process of feature interaction is as follows:

[0030] First, the primary abstract feature f x Expand from the original dimension [batch_size, seq_len, features] to f x1 , and f x1 The dimension is [batch_size, seq_len, 1, features]; f x1 Copy and expand to f x2 , and f x2 The dimension is [batch_size, seq_len, features, features];

[0031] Then, press the formula to x1 and f x2 Do matrix multiplication to interact features:

[0032] f xx =f x1 ·f x2 II

[0033] In Formula II, f xx is the result of feature interaction, and f xx The dimension is [batch_size, seq_len, 1, features];

[0034] After that, first f xx The dimension is reduced to [batch_size, seq_len, features] for subsequent calculations.

[0035] By adopting the above preferred scheme, the specific technical features of the third step can be further optimized.

[0036] Preferably, in the fourth step, the LSTM neural network is composed of a plurality of LSTM units; each LSTM unit is composed of a forget gate, an input gate, and an output gate;

[0037] First, the forget gate is processed as follows to obtain the forget weight f t :

[0038] f t =σ(W f ·[h t-1 ,x t ]+b f ) III

[0039] In formula III, h t-1 Represents the short-term memory output feature obtained at the previous time point, x t Represents the feature interaction result f at the current time point xx , W f represents the weight matrix of the forget gate, b f Represents the bias matrix of the forget gate, σ represents the Sigmod activation function;

[0040] Secondly, the input gate is processed as follows and the input weight i is obtained t and candidate features

[0041] i t =σ(W i ·[h t-1 ,x t ]+b i ) IV

[0042]

[0043] In Formula IV, W i represents the weight matrix of the input gate, b i Represents the bias matrix of the input gate, σ represents the Sigmod activation function;

[0044] In formula V, W C Represents the weight matrix of candidate features, b C represents the bias matrix of the candidate feature, tanh is the tanh activation function;

[0045] After that, combined with the forget weight f of the forget gate t , the input weight i of the input gate t and candidate features The current LSTM unit updates the long-term memory C as follows t :

[0046]

[0047] In Formula VI, C t-1 Represents the long-term memory obtained at the previous point in time;

[0048] Next, the output gate is processed as follows and obtains the short-term memory output feature h of the current LSTM unit t :

[0049] o t =σ(W o ·[h t-1 ,x t ]+b o ) VII

[0050] h t =o t *tanh(C t ) VIII

[0051] In Formula VII, o t Represents the output weight, W o represents the weight matrix of the output gate, b o Represents the bias matrix of the output gate, σ represents the Sigmod activation function;

[0052] In Formula VIII, C t is long-term memory, and tanh is the tanh activation function.

[0053] More preferably, in the fourth step, after the LSTM neural network extracts the incoming feature interaction results, it outputs the short-term memory output features extracted by each LSTM unit in a time sequence, and takes the short-term memory output features at the last time point as the comprehensive time series abstract feature s t .

[0054] By adopting the above preferred scheme, the specific technical features of the fourth step can be further optimized.

[0055] Preferably, in the fifth step, the conversion is performed according to the following formula:

[0056] f 1 =W 1 ·s t +b 1 IX

[0057]

[0058] In Formula IX, W 1 is the weight matrix of the first fully connected layer, b 1 is the bias matrix of the first fully connected layer; in formula X, W 2 is the weight matrix of the second fully connected layer, b 2 is the bias matrix of the second fully connected layer, To predict the probability of being trapped.

[0059] In the sixth step, the sample balanced cross entropy loss function is used, as shown in the following formula:

[0060]

[0061]

[0062]

[0063] In Formula XI and Formula XII, α true is the weight coefficient of the positive sample, α false is the weight coefficient of negative samples, n true is the number of positive samples, n false is the number of negative samples;

[0064] In Formula XIII, is the actual situation of someone being trapped in the elevator, and takes the value 0 or 1; is the predicted probability of being trapped; L is the predicted loss.

[0065] By adopting the above preferred scheme, the specific technical features of the fifth and sixth steps can be further optimized.

[0066] The present invention also provides:

[0067] The feature cross-elevator entrapment time series prediction model is constructed using the construction method described above.

[0068] The prediction model can be formally deployed to predict the probability of people being trapped in a target elevator.

[0069] The present invention also provides:

[0070] A feature-crossing elevator trapped person time series prediction method based on deep learning, characterized in that it includes the following steps:

[0071] S1. Adopting the deep learning-based feature crossover elevator trapped people time series prediction model construction method described above, by using the same data source and saving the models at different stages in the same training process, a plurality of prediction models with different weight parameters are obtained;

[0072] S2. Collect the original data of the target elevator, input them into the multiple prediction models of S1 respectively and obtain multiple predicted entrapment probabilities, filter them and obtain the comprehensive entrapment probability as the prediction result; the filtering process is selected from or includes taking the average value and taking the minimum value; the prediction method ends.

[0073] This prediction method uses multiple prediction models with slightly different focuses to perform comprehensive processing and obtain a comprehensive probability of entrapment, so that better prediction performance can be achieved under the premise of limited effective data.

[0074] Compared with the prior art, after the present invention inputs the historical information data of the elevator into the deep learning model, the model normalizes all the information, extracts the time series features through the LSTM network and the self-attention mechanism, and then predicts the probability of people being trapped in the elevator through the fully connected layer. The present invention is suitable for predicting people being trapped in an elevator with only historical record data of the elevator; the present invention can extract the time series features of the elevator using only the historical record data of the elevator, and predict the probability of people being trapped in the elevator at the next time point, under the premise of lacking the real-time operation data of the elevator components, and is suitable for scenarios where the real-time operation data of the elevator is insufficient under traditional working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 This is a schematic diagram of the main process of Example 1 of the present invention.

[0076] Figure 2 Schematic diagram of the LSTM neural network in Example 1 of the present invention. DETAILED DESCRIPTION

[0077] In specific implementation, the feature cross-elevator entrapment time series prediction model construction method, prediction model, and prediction method of the present invention based on deep learning respectively contain all the technical features recorded in the corresponding parts of the previous invention content (including the features of the preferred solution), and the specific contents will not be repeated here.

[0078] The present invention will be described in further detail below with reference to the accompanying drawings and in combination with embodiments, but the present invention is not limited to the examples given.

[0079] Example 1

[0080] This embodiment is a method for constructing a feature-crossing elevator entrapment time series prediction model based on deep learning, and the basic process is as described above.

[0081] like Figure 1 As shown, the specific contents of this embodiment are as follows:

[0082] In the first step, the deep learning model uses the elevator model, speed limit, weight limit, location, manufacturing unit, weather, temperature, trapped record and fault record as input, and the dimension of the input tensor is [batch_size, seq_len, features], where batch_size is the number of time series input to the model at a single time, seq_len is the length of the time series, and features is the number of features at each time point.

[0083] Step 2: Input information f input Passed into a single fully connected layer and converted into a primary abstract feature f x :

[0084] f x=W FC ·f input +b FC I

[0085] In Formula I, W FC is the weight matrix of the fully connected layer, f input For input information, b FC is the bias of the fully connected layer, f x is the primary abstract feature; the features dimension of the fully connected layer output feature can be freely set, specifically: by setting the matrix W FC The number of columns f determines the primary abstract feature f x characteristic length.

[0086] The third step is to perform feature interaction on the primary abstract features themselves and obtain feature interaction results, which can introduce high-order interactions between features and make the model more powerful in predicting. The specific process is as follows:

[0087] The primary abstract feature f x Expand from the original dimension [batch_size, seq_len, features] to f x1 , whose dimensions are [batch_size, seq_len, 1, features], f x1 Copy and expand to f x2 , whose dimension is [batch_size, seq_len, features, features].

[0088] The feature interaction is f x1 and f x2 To do matrix multiplication, the formula is as follows:

[0089] f xx =f x1 ·f x2 II

[0090] In Formula II, f xx It is the result of feature interaction, and its dimension is [batch_size, seq_len, 1, features]. For the convenience of further calculation, its dimension is reduced to [batch_size, seq_len, features].

[0091] The fourth step is to pass the feature interaction results into the LSTM neural network (LSTM: long short-term memory) to obtain comprehensive temporal abstract features.

[0092] The LSTM model uses a forget gate, an input gate, and an output gate to maintain both long-term memory and short-term memory feature lines, making up for the problems of gradient vanishing and gradient exploding, and insufficient long-term memory capacity of the RNN model, enabling the recurrent neural network to truly and effectively utilize long-distance temporal information.

[0093] This embodiment adopts Figure 2 The LSTM neural network shown. Figure 2 In each LSTM unit A, each line marks the transfer of an entire vector from the output of one node to the input of another node. The circles represent bitwise operations, such as vector sums and vector dot products, while the rectangular boxes are learned neural network layers. The lines that come together represent connections between vectors, and the lines that separate represent content being copied and then distributed to different locations.

[0094] The key to LSTM is the cell state C t , is the horizontal line running through the top of the graph. The running channel of the cell state is similar to a conveyor belt. The cell state is transmitted throughout the chain, with only a small amount of linear interaction with other information. The information of the cell state is transmitted on the chain, and it is guaranteed that what is learned is a long-term memory with long-term effects. LSTM mainly works through three gate structures, namely the forget gate, input gate and output gate. Each gate structure will have an independent weight matrix to process the input information.

[0095] The forget gate determines which historical information is effective at the current time point and the weight of different information. The formula is shown in Formula III, where f t represents the forgetting weight, h t-1 represents the short-term memory output feature obtained at the previous time point (derived from the previous LSTM unit, the same below), x t Represents the input information at the current time point (i.e., the feature interaction result f xx , the same below), W f represents the weight matrix of the forget gate, b f represents the bias matrix of the forget gate, and σ represents the Sigmod activation function.

[0096] f t =σ(W f ·[h t-1 ,x t ]+b f ) III

[0097] The input gate determines which information at the current time point is effective and the weight of different information. The formula is shown in Formula IV and Formula V, where i t represents the input weight, h t-1 Represents the short-term memory output feature obtained at the previous time point, x tRepresents the input information at the current time point, W i represents the weight matrix of the input gate, b i represents the bias matrix of the input gate, σ represents the Sigmod activation function, is the candidate feature, W C Represents the weight matrix of candidate features, b C Represents the bias matrix of the candidate feature, and tanh is the tanh activation function.

[0098] For each LSTM unit, the input gate consists of two parts. The first part combines short-term memory and current input, and outputs a value between 0 and 1 through the Sigmod activation function to generate input weights and determine which features need to input new information. The second part combines short-term memory and current input through the tanh activation function to generate a candidate feature, and each value of the candidate feature is between -1 and 1.

[0099] i t =σ(W i ·[h t-1 ,x t ]+b i )IV

[0100]

[0101] Combining the long-term memory processed by the forget gate and the short-term memory processed by the input gate, as well as the current input, the LSTM neural network will update the long-term memory. The formula is shown in Formula VI, where C t-1 Represents the long-term memory obtained at the previous point in time.

[0102]

[0103] The output gate mixes historical information with current information to jointly determine the current output. The formula is shown in Formula VII and Formula VIII, where o t represents the output weight, h t-1 Represents the short-term memory output feature obtained at the previous time point, x t Represents the input information at the current time point, W o represents the weight matrix of the output gate, b o represents the bias matrix of the output gate, σ represents the Sigmod activation function, and h t is the short-term memory output feature at the current time point, C t represents long-term memory and tanh is the tanh activation function.

[0104] o t =σ(W o ·[h t-1 ,x t ]+bo ) VII

[0105] h t =o t *tanh(C t ) VIII

[0106] After the feature interaction results are extracted by LSTM, the short-term memory output features extracted by each LSTM unit are output in time sequence, and the short-term memory output features at the last time point are taken as the comprehensive time series abstract feature s t .

[0107] Step 5: Abstract the comprehensive temporal features s t The data is passed to two fully connected layers and converted into the predicted probability of being trapped. The calculation method is as follows:

[0108] f 1 =W 1 ·s t +b 1 IX

[0109]

[0110] In Formula IX, W 1 is the weight matrix of the first fully connected layer, b 1 is the bias matrix of the first fully connected layer; in formula X, W 2 is the weight matrix of the second fully connected layer, b 2 is the bias matrix of the second fully connected layer, To predict the probability of being trapped.

[0111] Step 6: Pass the model-predicted probability of people being trapped and the actual situation of people being trapped into the cross-entropy loss function, calculate the prediction loss L, and update the neural network parameters through back propagation to train the model; the actual value of people being trapped comes from the historical information data of the elevator.

[0112] Note: Updating neural network parameters by back-propagation is a prior art method in this field, for example: Rumelhart, D., Hinton, G. & Williams, R. Learning representations by back-propagating errors. Nature 323, 533–536 (1986). https: / / doi.org / 10.1038 / 323533a0 ), so the specific content will not be described here.

[0113] According to the analysis of the data source, elevator entrapment incidents are sparse in time series. The number of records of no-people-trapped elevators in the time series is at least 20 times that of records of people-trapped elevators. Inputting the entire time series of people-trapped elevators and the time series of no-people-trapped elevators into the deep learning model for training will cause the model to ignore the input features and directly predict the probability of people being trapped as zero. Therefore, it is necessary to add weight balancing and add different training weights for different types of samples so that the model can combine positive and negative samples to obtain a more balanced model.

[0114] In this embodiment, the weight balancing is implemented as follows: when calculating the loss between the predicted loss value and the actual value, different types of samples are multiplied by different weight coefficients. The weight coefficient is calculated as follows:

[0115]

[0116]

[0117] Among them, α true is the weight coefficient of the positive sample, α false is the weight coefficient of negative samples, n true is the number of positive samples, n false is the number of negative samples.

[0118] The sample balanced cross entropy loss function is used, as shown in the following formula:

[0119]

[0120] in, is the actual situation of someone being trapped in the elevator, and takes the value 0 or 1; To predict the probability of being trapped. This can reduce the impact of sample imbalance on model prediction and enable the model to adapt to situations where the number of samples of different types differs greatly.

[0121] This completes the construction of the prediction model.

[0122] Example 2

[0123] This embodiment is a feature-crossing elevator entrapment time series prediction method based on deep learning, and the basic process is as described above.

[0124] This embodiment is the formal deployment stage of Embodiment 1. The specific contents of this embodiment are as follows:

[0125] Using Example 1, the input prediction sequence is filtered using multiple trained models, and the recognized probability of being trapped by multiple models is taken as the final prediction result. The filtering methods include taking the average value and taking the minimum value. That is, the predicted probability of being trapped by multiple models is filtered to obtain the comprehensive probability of being trapped as the prediction result.

[0126] The multiple models used in this embodiment have the same structure and use the same training data. They are models saved at different stages of the same training process, so the weight parameters of these models are different. Due to the generalization requirements of deep learning models, each model has its own emphasis on feature extraction of time series. The model obtained by training in the next stage is a fine-tuning of the previous model, and its focus on feature extraction will change. By combining the focus of different models and connecting these models in series, a comprehensive probability of being trapped can be obtained.

[0127] In this embodiment, the collected 151,128 elevators in use are used as the data source, and their data in July, August and September 2021 are used as the test data set. After the probability of elevator entrapment is extracted by the above method, the prediction performance of the model is evaluated from multiple angles, and the recognition performance comparison table is shown in Table 1. It can be seen that the elevator entrapment prediction model can achieve better prediction performance under the premise of lacking real-time operation data of elevator components. Note: When the difference between the predicted probability of elevator entrapment and the actual entrapment situation is less than the threshold, the model prediction is judged to be correct.

[0128] In July, August and September 2021, the model's prediction accuracy reached 98.14%, 98.47% and 98.54% respectively, but for actual trapped incidents, the model's prediction accuracy was 11.75%, 10.29% and 8.60% respectively.

[0129] This means that the model of this embodiment can be used as an effective aid for elevator maintenance in practical applications, and can screen out a number of elevators with a high probability of entrapment. After focused maintenance, the probability of elevator entrapment incidents can be reduced. At the same time, in actual elevator maintenance, the level of detail of manually recorded elevator information under a large number of traditional working conditions is consistent with the level of detail of the information used in this model. Therefore, the model of this embodiment has high engineering feasibility and practicality.

[0130] Table 1

[0131] month Accuracy Accuracy Number of hits Prediction number The real number of trapped people Deviation threshold 2021-07 98.14％ 11.75％ 225 1915 1353 0.001 2021-08 98.47％ 10.29％ 155 1507 1113 0.001 2021-09 98.54％ 8.60％ 149 1732 775 0.001

[0132] In addition to the above embodiments, the present invention may also have other implementation modes. Any technical solution formed by equivalent replacement or equivalent transformation falls within the protection scope required by the present invention.

Claims

1. A method for constructing a feature-crossing elevator trapped person time series prediction model based on deep learning, characterized in that: The following steps are involved: Step 1: The deep learning model organizes the historical information data of the elevator into time series for input according to time points; the dimension of the input tensor is [batch_size, seq_len, features], where batch_size is the number of time series input into the deep learning model at a time, seq_len is the length of the time series, and features is the number of features contained in the time series at each time point; The second step is to pass the input information into a single fully connected layer and convert it into primary abstract features; the conversion is performed as follows: I In Formula I, is the weight matrix of the fully connected layer, To input information, is the bias of the fully connected layer, It is a primary abstract feature; The third step is to perform feature interaction on the primary abstract features themselves to obtain feature interaction results. The specific process of feature interaction is as follows: First, the primary abstract features Expanded from the original dimension [batch_size, seq_len, features] to ,and The dimension of is [batch_size, seq_len, 1, features]; Copy and expand ,and The dimension is [batch_size, seq_len, features, features]; After that, press the and Do matrix multiplication to interact features: II In Formula II, is the result of feature interaction, and The dimension is [batch_size, seq_len, 1, features]; After that, first The dimension is reduced to [batch_size, seq_len, features] for subsequent calculations; Step 4: Pass the feature interaction results into the LSTM neural network and obtain comprehensive temporal abstract features ; Step 5: Abstract the features of the comprehensive time series Pass it into two fully connected layers and convert it into the predicted probability of being trapped; convert it as follows: IX X In Formula IX, is the weight matrix of the first fully connected layer, is the bias matrix of the first fully connected layer; in formula X, is the weight matrix of the second fully connected layer, is the bias matrix of the second fully connected layer, To predict the probability of people being trapped; Step 6: Pass the predicted probability of trapped people and the actual value of trapped people into the cross entropy loss function, calculate the prediction loss, and update the LSTM neural network parameters through back propagation to train the model; the actual value of trapped people comes from the historical information data of the elevator; wherein, the sample balanced cross entropy loss function is used, as shown in the following formula: XI XII XIII In Formula XI and Formula XII, is the weight coefficient of the positive sample, is the weight coefficient of negative samples, is the number of positive samples, is the number of negative samples; In Formula XIII, is the actual situation of someone being trapped in the elevator, and takes the value 0 or 1; To predict the probability of people being trapped; To predict losses; The prediction model construction is completed.

2. According to the method for constructing a feature-crossing elevator entrapment time series prediction model based on deep learning in claim 1, it is characterized in that, in the first step, the historical information data is selected from or includes the elevator model, speed limit, weight limit, location, manufacturing unit, weather, temperature, entrapment record, and fault record.

3. The method for constructing a feature-crossing elevator entrapment time series prediction model based on deep learning according to claim 1 is characterized in that: By setting the matrix The number of columns f determines the primary abstract features characteristic length.

4. The method for constructing a feature-crossing elevator trapped person time series prediction model based on deep learning according to claim 1, characterized in that, in the fourth step, the LSTM neural network is composed of a plurality of LSTM units; each LSTM unit is composed of a forget gate, an input gate, and an output gate; First, the forget gate is processed as follows to obtain the forget weight : III In formula III, Represents the short-term memory output feature obtained at the previous time point, Represents the feature interaction result at the current time point , represents the weight matrix of the forget gate, represents the bias matrix of the forget gate, Represents the Sigmod activation function; Secondly, the input gate is processed as follows and the input weight is obtained and candidate features : IV V In Formula IV, represents the weight matrix of the input gate, represents the bias matrix of the input gate, Represents the Sigmod activation function; In Formula V, represents the weight matrix of candidate features, represents the bias matrix of the candidate feature, tanh is the tanh activation function; After that, combined with the forget weight of the forget gate , the input weight of the input gate and candidate features , the current LSTM unit updates the long-term memory according to the formula : WE In Formula VI, Represents the long-term memory obtained at the previous point in time; Next, the output gate is processed as follows and obtains the short-term memory output feature of the current LSTM unit : VII VIII In Formula VII, represents the output weight, represents the weight matrix of the output gate, represents the bias matrix of the output gate, Represents the Sigmod activation function; In Formula VIII, is long-term memory, and tanh is the tanh activation function.

5. The method for constructing a feature-crossing elevator entrapment time series prediction model based on deep learning according to claim 4 is characterized in that: After extracting the incoming feature interaction results, the LSTM neural network outputs the short-term memory output features extracted by each LSTM unit in time sequence, and takes the short-term memory output features at the last time point as the comprehensive time series abstract features. .

6. A feature-crossing elevator trapped person time series prediction method based on deep learning, characterized in that: The following steps are involved: S1. Using the deep learning-based feature crossover elevator entrapment time series prediction model construction method according to any one of claims 1 to 5, by using the same data source and saving the models at different stages in the same training process, a prediction model with different weight parameters is obtained; S2. Collect the original data of the target elevator, input them into the multiple prediction models of S1 respectively and obtain multiple predicted entrapment probabilities, filter them and obtain the comprehensive entrapment probability as the prediction result; the filtering process is selected from or includes taking the average value and taking the minimum value; the prediction method ends.

Images