Elevator Fault Diagnosis Method Based on Dynamic Step Size PSO Optimized Neural Network

Abstract

The present invention discloses an elevator fault diagnosis method based on optimizing a neural network with a dynamic step size PSO. After obtaining a plurality of detailed operation data through sensors, the present invention uploads the collected data to a cloud data platform for analysis based on Internet of Things technology. First, a normalization method is used to preprocess the original elevator data, with linear processing to ensure the effectiveness of the data and to optimize the accuracy, stability, and generality of subsequent algorithms for elevator fault diagnosis. Then, the ReliefF algorithm is used to screen a large number of feature parameters, leaving the eigenvalues with a relatively large correlation with the fault type to be detected, and a multi-layer perceptron neural network model is established. The improved dynamic step size PSO algorithm is used to optimize the neural network, determine key parameters such as weights and thresholds with sufficient accuracy, and achieve an accurate mapping from input eigenvalues to output faults. The present invention can improve the accuracy and efficiency of detecting elevator fault types.

Description

Technical Field

[0001] The invention belongs to the field of data analysis, and in particular relates to an elevator fault diagnosis method based on a dynamic step-size PSO optimized neural network. Background Art

[0002] In recent years, my country's elevator industry has developed rapidly. Elevators are used more and more widely in cities. They can be seen everywhere in shopping malls, apartments, hospitals and other places. This has also brought a series of safety problems, such as door lock circuit disconnection, deceleration fault, inverter fault, etc. As a complex electromechanical equipment with high dimension, strong nonlinearity and strong coupling, the elevator has suddenness, diversity, correlation, uncertainty and so on. Once a fault occurs, it is easy to involve multiple mechanical structures of the same level or secondary level. At present, many elevators have insufficient number of maintenance personnel or less experience. It is often difficult to quickly and accurately diagnose all types of elevator faults and corresponding fault parts simply by relying on simple instruments and equipment or the experience of maintenance personnel. This is a big hidden danger for maintenance efficiency and personnel safety. Therefore, elevator diagnosis is crucial to the maintenance of elevator safety. Due to the increase in fault types and the variability of fault parts, and the continuous shortage of maintenance personnel market, elevator diagnosis and post-diagnosis that rely too much on human subjectivity not only increase maintenance costs and affect passenger use, but also easily lead to other problems such as loss of other elevator parts. Therefore, how to analyze various operation data of elevators and perform efficient and accurate fault diagnosis of elevators is an issue that needs to be studied urgently.

[0003] Research in the field of machine learning has been going on for many years at home and abroad, and the technology for data mining and prediction has become more and more mature. At present, there are many intelligent diagnosis algorithms and they have been applied to elevator remote monitoring systems. For example, traditional fault tree, random forest method, neural network, support vector machine and other intelligent learning algorithms have also been applied to the field of fault diagnosis. In order to improve the accuracy of fault diagnosis, traditional fault diagnosis algorithms and optimization algorithms such as genetic algorithm and particle swarm algorithm have also been integrated into neural network. However, there are still problems such as local optimum and slow convergence speed. The diagnosis speed and accuracy of elevator faults need to be improved. Summary of the invention

[0004] In view of the problems that traditional algorithms may fall into local optimality and insufficient convergence speed, the present invention proposes an elevator fault diagnosis method based on dynamic step size PSO optimized neural network, aiming to propose a method suitable for the current situation where elevator faults are frequent and complex and the fault locations are changeable, and to improve the diagnosis speed and accuracy of some elevator faults.

[0005] The technical solution of the present invention is as follows:

[0006] The present invention comprises the following steps:

[0007] Step S1: Obtain the real-time data during the operation of the elevator, i.e., elevator operation data, through current sensors, voltage sensors, vibration sensors, force sensors, position sensors, and speed sensors. Use the normalization method to preprocess the elevator operation data as the data samples for subsequent algorithms.

[0008] Step S2: Denote each type of elevator operation data as a feature. Then all types of elevator operation data form a feature value set. Take four common fault types as the output layer, and use the reliefF method to screen the top ten feature value sets with the strongest correlation with the four fault types from the preprocessed elevator operation data as the input layer.

[0009] Step S3: Establish a BP neural network. The BP neural network consists of an input layer with ten nodes, an intermediate hidden layer with eight nodes, and an output layer with four nodes. Set the error target of the BP neural network to 0.001.

[0010] Step S4: Initialize the parameters and structure of the particle swarm optimization algorithm. Use the error function of the BP neural network established in Step S3 as the fitness function. Set the conditions for exiting the particle swarm optimization algorithm, including restricting the value of the target fitness function and the maximum number of iterations. Take the weights and thresholds of the BP neural network as the positions of the particles. Randomly initialize the population size of the particle swarm and each position and speed, and limit the maximum number of iterations.

[0011] Step S5: Update the positions and speeds of the particles through the iterative formula of the improved particle swarm optimization algorithm. In the present invention, a particle movement step factor is added, making the movement of the particles related to the superiority or inferiority of the individual fitness in the group.

[0012] Step S6: Sort the particles after each iteration of the particle swarm algorithm according to their fitness. Consider the 1 / 3 particles with inferior fitness as the fireworks of the fireworks algorithm to generate sparks, and select the next generation of particles through the elite-roulette strategy. After iterative cycling, if the particle swarm meets the conditions, finally output the global optimal solution.

[0013] Step S7: Take the global optimal solution of the particle swarm as the initial weights and thresholds of the BP neural network to complete the optimization of the BP neural network. Train 70% of the preprocessed elevator operation data through the optimized BP neural network, and test 30% of the preprocessed elevator data through the optimized BP neural network.

[0014] Step S8: Train and test the same elevator data through an unoptimized standard BP neural network, and compare the training and test results with those in Step S7 to prove that the optimized BP neural network has a certain improvement in the accuracy of elevator fault diagnosis compared with the standard BP neural network.

[0015] Step S9: Apply the optimized BP neural network to elevator fault diagnosis.

[0016] Advantages of the present invention:

[0017] First, in order to optimize the convergence speed of the BP neural network, the classical swarm intelligence search algorithm, namely the particle swarm optimization algorithm (PSO), is added. The advantage of this algorithm is that it is simple and easy to implement, and has a relatively high global simultaneous search efficiency. However, both it and the BP neural network are prone to falling into the local optimum problem. In the present invention, first, the weights and thresholds of the BP neural network are used as the output results of the particle swarm optimization algorithm, and the error function of the BP neural network is used as the fitness function of the particle swarm optimization algorithm.

[0018] Second, the simple particle swarm optimization algorithm is still prone to falling into the local optimum situation. In order to reduce the probability of this situation, the parameter settings of the inertia weight and learning factor in the algorithm are adjusted to be dynamically non-linear, and a dynamic step size method is introduced to improve the PSO algorithm. The group and individual velocities, positions, and superiority and inferiority degrees of the particles are all considered in the iterative formula, increasing the possibility of the particle swarm escaping from the local optimum.

[0019] Third, in order to enable the particles to search for better solutions in all directions and avoid all being trapped in the local optimum, the inferior particles are diffused in all directions in the way of fireworks explosion to improve the global optimization ability. If the convergence updates the individual extreme value and global extreme value positions of the particles, updates the particle positions and velocities, otherwise continue to execute the PSO algorithm formula. Finally, judge whether the requirements or the maximum number of iterations are reached. If so, obtain the initial weights and thresholds of the BP neural network, otherwise use the particle swarm as the initial particle swarm to update and iterate again.

[0020] In summary, after the above improvements of the present invention, the probability of the BP neural network falling into the local optimum solution can be effectively reduced, the training speed is optimized, the training error is reduced, thereby improving the accuracy and efficiency of detecting elevator fault types. Brief Description of the Drawings

[0021] Figure 1 is the overall flowchart of the present invention;

[0022] Figure 2 is the schematic diagram of optimizing the BP neural network by the hybrid particle swarm algorithm used in the present invention;

[0023] Figure 3 is the schematic diagram of the BP neural network in the present invention;

[0024] Figure 4 is the error change comparison curve graph of the training method in the present invention and other training methods. Detailed Embodiment

[0025] The present invention proposes an elevator fault diagnosis method based on a neural network optimized by a dynamic step PSO. The present invention introduces a dynamic step movement mode into the particle swarm search method, incorporates the fitness quality of particles into the search ability, dynamically allocates the step size of each particle according to the quality of the fitness function value. The inferior particles search for solutions in the surrounding space with a larger step size, and the superior particles search for solutions in the surrounding space with a smaller step size, ensuring that all particles are not prone to premature aggregation. Combining the idea of fireworks explosion, some particles are diffused in random directions to search the surrounding solution space, increasing the probability of finding the global optimal solution, and comprehensively providing a more accurate and rapid fault diagnosis method.

[0026] The specific technical solution of the present invention is as follows:

[0027] The present invention includes the following steps:

[0028] Step S1: Obtain the real-time data during the operation of the elevator, i.e., elevator operation data, through current sensors, voltage sensors, vibration sensors, force sensors, position sensors, and speed sensors, and preprocess the elevator operation data using a normalization method as the data samples for subsequent algorithms.

[0029] Step S2: Denote each type of elevator operation data as a feature, then all types of elevator operation data are the eigenvalue set. Take four common fault types as the output layer, and use the reliefF method to screen the eigenvalue set with the top ten correlation strengths with the four fault types from the preprocessed elevator operation data as the input layer.

[0030] Step S3: Establish a BP neural network. The BP neural network consists of an input layer containing ten nodes, an intermediate hidden layer containing eight nodes, and an output layer containing four nodes, and set the error target of the BP neural network to 0.001.

[0031] Step S4: Initialize the parameters and structure of the particle swarm optimization algorithm, and use the error function of the BP neural network established in Step S3 as the fitness function. Set the satisfaction conditions for exiting the particle swarm optimization algorithm, including restricting the target fitness function value and the maximum number of iterations. Take the weights and thresholds of the BP neural network as the positions of the particles. Randomly initialize the population size and each position and speed of the particle swarm, and restrict the maximum number of iterations.

[0032] Step S5: Update the positions and speeds of the particles through the iterative formula of the improved particle swarm optimization algorithm, and add the moving step size factor of the particles, so that the movement of the particles is related to the quality of the individual fitness in the group.

[0033] Step S6: Sort the particles after each iteration of the particle swarm algorithm, consider the 1 / 3 particles with inferior fitness as fireworks of the fireworks algorithm to generate sparks, and select the next generation of particles through the elite-roulette strategy. After iterative cycling, if the particle swarm meets the conditions, the global optimal solution is finally output.

[0034] Step S7: Use the global optimal solution of the particle swarm as the initial weights and thresholds of the BP neural network to complete the optimization of the BP neural network; train 70% of the preprocessed elevator operation data through the optimized BP neural network, and test 30% of the preprocessed elevator data through the optimized BP neural network.

[0035] Step S8: Train and test the same elevator data through the standard BP neural network without optimization, and compare the training and test results with those in Step S7 to prove that the optimized BP neural network has a certain improvement in the accuracy of elevator fault diagnosis compared to the standard BP neural network.

[0036] Step S9: Apply the optimized BP neural network to elevator fault diagnosis.

[0037] Furthermore, the said Step S1 includes the following steps:

[0038] Step S1.1: Traverse all elevator operation data to find the maximum and minimum values of each type of elevator operation data.

[0039] Step S1.2: Convert all elevator operation data into values within 0 to 1 according to the max-min method.

[0040] Furthermore, the said Step S2 includes the following steps:

[0041] Step S2.1: Randomly select a sample a from all the obtained elevator operation data, and find k nearest neighbor samples from the samples of the same type as sample a.

[0042] Step S2.2: Also take out k nearest neighbor samples from within all other sample groups of different classifications from sample a.

[0043] Step S2.3: Calculate the weights of each feature and take their average value.

[0044] Step S2.4: Sort the average values of the weights of all features and obtain the top ten feature values in terms of weight ranking.

[0045] Furthermore, the number of hidden layer nodes I in the said Step S3 is determined by the number of input layer nodes N and the number of output layer nodes M through the following formula, where round() is the rounding function.

[0046]

[0047] Furthermore, the maximum number of iterations of the step S4 is 100.

[0048] Furthermore, the step S7 reduces the error by the method of gradient descent until the error is less than the set target error value, and the training of the BP neural network ends.

[0049] For the technical solution of the present invention to be clearer, the following will be described in detail with reference to the accompanying drawings. As shown in this embodiment, the specific steps are as follows: Figure 1 As shown below, the specific steps are as follows:

[0050] Step S1.1: Obtain the real-time data during the operation of the elevator, that is, the elevator operation data, through a current sensor, a voltage sensor, a vibration sensor, a force sensor, a position sensor, and a speed sensor, and preprocess the elevator operation data using a normalization method as the data samples for subsequent algorithms.

[0051] Step S1.2: Perform normalization preprocessing on the data to ensure that all data is between [0, 1], that is, according to the formula:

[0052]

[0053] In the formula: χ max , χ min are respectively the minimum and maximum values of the collected data, and χ i is the actual value of the collected data.

[0054] Step S2: Use the reliefF feature selection method to screen various types of elevator data to determine the input layer:

[0055] Input: training set D, sampling times M, feature weight threshold δ, number of nearest neighbor samples k;

[0056] Output: feature weights T of each feature.

[0057] The specific steps of the algorithm are as follows:

[0058] (1) Set all feature weights to 0, and T is an empty set

[0059] (2) Calculate all samples according to the following formula:

[0060] ① Randomly select a sample R from D;

[0061] ② Find the k nearest neighbor samples H of R from the same class of R j (j = 1, 2,..., k), and find k nearest neighbors M from each different class sample set j (C);

[0062] (3) Loop N times with the following formula:

[0063]

[0064] (4) Sort the weighted average values of all features to obtain the top ten feature values in terms of weight.

[0065] Step S3.1: Set the error target of the BP neural network to 0.001.

[0066] Step S3.2: Determine the input layer of the BP neural network: Use the ten types of elevator operation data obtained in Step S2 as the input layer of the BP neural network.

[0067] Step S3.3: Determine the output layer of the BP neural network: Based on the elevator operation data obtained in Step S1, use the four common types of elevator faults as the output layer of the BP neural network: abnormal elevator speed, abnormal elevator position, motor overload, and brake contactor abnormality. Encode these four types of elevator faults as (1,0,0,0,0) T , (0,1,0,0,0) T , (0,0,1,0,0) T , (0,0,0,1,0) T , (0,0,0,0,1) T .

[0068] Step S3.4: The number of hidden layer nodes I of the BP neural network is determined by the number of input layer nodes N and the number of output layer nodes M through the following formula, where round() is the rounding function.

[0069]

[0070] The types of elevator operation data in the final input layer of the BP neural network are as follows in the table

[0071]

[0072]

[0073] The fault types in the output layer of the BP neural network are as follows in the table

[0074]

[0075] Steps S4 - S6 use the optimized particle swarm algorithm to optimize the initial weights and thresholds of the BP neural network. The overall summary flowchart is shown in Figure 2 , and the specific steps are as follows:

[0076] Step S4: Initialize the parameters and structure of the particle swarm optimization algorithm, and use the error function of the BP neural network established in step S3 as the fitness function. Set the satisfaction conditions for exiting the particle swarm optimization algorithm, including restricting the target fitness function value and the maximum number of iterations. Use the weights and thresholds of the BP neural network as the positions of the particles. Randomly initialize the population size and the positions and velocities of each particle, and limit the maximum number of iterations to 100.

[0077] Step S5.1: Initialize the population. Determine the population size R and dimension D according to the neural network structure, and the velocity v and position p of each particle. Use the error function of the BP neural network as the fitness value function to evaluate the quality of each particle: E(w) = y k (1 - y k )(y i - y k ).

[0078] Step S5.2: Since the particle swarm search method is related to the velocities and positions of the particles, update the particle velocities and positions according to the following formulas:

[0079] V i (k + 1) = ωV i (k) + step i (c 1 r 1 [p i - x i (k)] + c 2 r 2 [p g - x i (k)])

[0080] X i (k + 1) = X i (k) + V i (k + 1)

[0081] In the formula, k is the number of iterations, c 1 , c 2 are acceleration factors, r 1 , r 2 are random numbers between [0, 1], ω is the inertia coefficient. To prevent the particle movement range from being too large or the velocity from being too fast during the iteration process, appropriate thresholds need to be set for restriction. step i is the step coefficient, which functions to dynamically allocate the step size, allocate the minimum step size for the optimal ones, and the maximum step size for the worst particles.

[0082] In elevator fault diagnosis, it takes a certain amount of time for maintenance personnel to reach the elevator. The elevator fault diagnosis time is much smaller than the commuting time of maintenance personnel. Improving the accuracy of elevator fault diagnosis can greatly increase the maintenance rate of maintenance personnel and improve the maintenance efficiency. Therefore, compared with the standard PSO algorithm, this embodiment proposes a new particle search iteration formula, which can greatly increase the probability of finding the global optimal solution of the search algorithm while maintaining a high convergence speed. A step size coefficient related to the fitness level is added to the original iteration formula. The step size of particles with inferior fitness is longer, and the step size of particles with better fitness is shorter. The original PSO algorithm only adjusts the speed according to the iteration time of particles, and the particle search methods are too similar. When particles gather prematurely, there is no means to jump out of the local optimal region, while this solution can allocate step sizes according to the fitness level, providing a greater possibility for particles to jump out of the local optimal region and effectively reducing the probability of premature convergence.

[0083] This embodiment adopts a non-linear inertia weight that dynamically decreases according to time, that is

[0084] ω(j) = ω min +(ω max -ω min )×exp(-k 1 (j / j max ) 2 )

[0085] K 1 is the control factor, generally with a value range of (3.0 - 4.0), ω max 、ω min 、j、j max are the maximum weight value, minimum weight value, current iteration number, and maximum iteration number.

[0086] The step size coefficient formula is as follows:

[0087]

[0088] top(f(x i )) = a k2

[0089] step 0 is the coefficient that determines the accuracy of particle optimization; a determines the width of the step size, and the test range is between 1.0 - 1.6; K 2 represents the particle number, and top(f(x i )) is the sorting of particles from smallest to largest according to the function value.

[0090] In this way, more attention is paid to global search in the initial stage of the algorithm, and rapid convergence occurs in the later stage, optimizing the accuracy and speed of the algorithm to find the optimal solution.

[0091] Step S5.3: Calculate the fitness value of the particles after each iteration and sort the new particle swarm pop 1 Sorting

[0092] Step S6: Consider the inferior particles (the worst 1 / 4 particles in terms of fitness value) in the new particle swarm as fireworks in the fireworks algorithm. The fireworks algorithm uses an explosion of a certain amplitude to spread the central particles out, with characteristics such as randomness, locality, explosiveness, implicit parallelism, diversity, and instantaneity. Combining it with the PSO algorithm can utilize the inferior particles to enhance the ability to search for the global optimal solution. After generating the fireworks, use the elitist strategy to select the next generation of sparks

[0093] Formula for the number of sparks generated by the fireworks algorithm

[0094]

[0095] S i is the number of sparks generated by the i-th firework; m is a constant used to predict the total number of sparks generated; Y max represents the maximum fitness value in the current population; f(x i ) is the fitness value of the individual x i ; ε is a very small constant to avoid division by zero

[0096] For each firework, set the following formula for the limit of the number of sparks generated

[0097]

[0098] a < b < 1

[0099] Round() is a function for rounding to the nearest integer; a and b are given constants

[0100] The calculation of the explosion amplitude range is as follows

[0101]

[0102] is a constant representing the maximum explosion amplitude; Y min represents the minimum fitness value in the current population

[0103] Particle displacement operation formula

[0104] ΔX i (k) = X i (k) + rand(0, A i )

[0105] rand(0, A i ) means generating a uniform random number within the amplitude A i

[0106] ​Step S6.1: Update the final position and velocity of the particle at the current iteration, and update the individual and global extreme values.

[0107] Step S6.2: Determine whether the maximum number of iterations has been reached or the fitness condition has been met. If so, output the optimal particle, i.e., the optimal initial weights and thresholds of the BP neural network; otherwise, return to Step S5.2.

[0108] Finally, use the elevator operation data and the samples corresponding to elevator faults as the training set of the BP neural network.

[0109] The schematic diagram of BP neural network training in Step S7 is shown in Figure 3 , and the BP neural network training is mainly divided into two processes: the forward transmission subprocess of the working signal and the backward transmission subprocess of the error signal. The task of the forward transmission of the working signal is to implement the output of the current node according to the output values of all nodes in the upper layer, the weights between the current node and all nodes in the upper layer, the threshold of the current node, and the activation function. The task of the backward transmission of the error signal is to repeatedly correct the weights and thresholds by the method of gradient descent to minimize the error function value, that is, to reduce the error between the actual output and the expected output of the system by changing the connection weights between neurons.

[0110] The specific algorithm steps of the BP neural network are as follows: Initialize the weights w and thresholds θ, that is, the connection weights w from the input value unit to the hidden layer unit ij , the connection weights v from the hidden layer to the output layer ij , the threshold θ of the hidden layer j , the threshold θ of the output layer unit k .

[0111] Let the training sample of the input layer be the vector x i , and the expected output vector be

[0112] Then the output of the hidden layer is

[0113] The output value of the output layer is

[0114] The error between the actual output value and the expected value is

[0115] Backpropagate the error to the hidden layer, then the error of the hidden layer is

[0116] Iteratively correct the weights from the hidden layer to the output layer as

[0117] The threshold is

[0118] Figure 4 This is the comparison of the training performance between the optimized BP neural network and the traditional network in this embodiment. Under the elevator fault diagnosis error rate that meets the requirements, the BP neural network in this embodiment has a faster convergence speed and better performance.

[0119] Step S8: Train and test the same elevator data through the unoptimized standard BP neural network, and compare the training and test results with those in Step S7 to prove that the optimized BP neural network has a certain improvement in the accuracy of elevator fault diagnosis compared with the standard BP neural network.

[0120] The partial results of the final elevator fault diagnosis are shown in the following table

[0121]

[0122]

[0123] It can be seen from the results that the optimized BP neural network described in this embodiment has good effects in elevator fault diagnosis.

[0124] Step S9: Apply the optimized BP neural network to elevator fault diagnosis.

Claims

1. An elevator fault diagnosis method based on a neural network optimized by a dynamic step size PSO, characterized in that this method comprises the following steps: Step S1: Obtain real-time data during elevator operation, i.e., elevator operation data, through a current sensor, a voltage sensor, a vibration sensor, a force sensor, a position sensor, and a speed sensor, and preprocess the elevator operation data using a normalization method; Step S2: Denote each type of elevator operation data as a kind of feature, then all types of elevator operation data are a feature value set; Take four fault types as the output layer, and use the reliefF method to screen the top ten feature value sets with the strongest correlation with the four fault types from the preprocessed elevator operation data as the input layer; Step S3: Establish a BP neural network, where the BP neural network consists of an input layer containing ten nodes, an intermediate hidden layer containing eight nodes, and an output layer containing four nodes, and set the error target of the BP neural network; Step S4: Initialize the parameters and structure of the particle swarm optimization algorithm, and use the error function of the BP neural network established in Step S3 as the fitness function; Set the satisfaction conditions for exiting the particle swarm optimization algorithm, including restricting the value of the target fitness function and the maximum number of iterations; Take the weights and thresholds of the BP neural network as the positions of the particles; Randomly initialize the population size of the particle swarm and each position and speed, and restrict the maximum number of iterations; Step S5: Update the positions and speeds of the particles through the iterative formula of the improved particle swarm optimization algorithm, where the improvement lies in: adding a moving step size factor of the particles, making the movement of the particles related to the superiority and inferiority of the individual fitness in the group; Step S6: Sort the particles after each iteration of the particle swarm algorithm according to fitness, regard the 1 / 3 particles with inferior fitness as fireworks of the fireworks algorithm to generate sparks, and select the next generation of particles through the elite-roulette strategy; If the particle swarm meets the conditions after cyclic iteration, finally output the global optimal solution; Step S7: Take the global optimal solution of the particle swarm as the initial weights and thresholds of the BP neural network to complete the optimization of the BP neural network; Train 70% of the preprocessed elevator operation data through the optimized BP neural network, and test 30% of the preprocessed elevator data through the optimized BP neural network; Step S8: Train and test the same elevator data through a standard BP neural network without optimization, and compare the training and test results with the training and test results in Step S7 to prove that the optimized BP neural network has a certain improvement in the accuracy of elevator fault diagnosis compared with the standard BP neural network; Step S9: Apply the optimized BP neural network to elevator fault diagnosis; The said Step S2 comprises the following steps: Step S2.1: Randomly select a sample a from all the obtained elevator operation data, and find k nearest neighbor samples from the samples of the same kind as sample a; Step S2.2: Also take out k nearest neighbor samples from within all other sample groups with different classifications from sample a respectively; Step S2.3: Calculate the weight of each feature and take its average value; Step S2.4: Sort the average weights of all features to obtain the top ten eigenvalues of the weights.

2. The elevator fault diagnosis method based on a neural network optimized by dynamic step PSO according to claim 1, wherein, the step S1 includes the following steps: Step S1.1: Traverse all elevator operation data to find the maximum and minimum values of each type of elevator operation data; Step S1.2: Convert all elevator operation data into values within 0 to 1 according to the max-min method in proportion.

3. The elevator fault diagnosis method based on a neural network optimized by dynamic step PSO according to claim 1, wherein, the number of hidden layer nodes I in the step S3 is determined by the number of input layer nodes N and the number of output layer nodes M through the following formula: I = round( ) + 4 where round() is the rounding function.

4. The elevator fault diagnosis method based on a neural network optimized by dynamic step PSO according to claim 1, wherein, the maximum number of iterations in the step S4 is 100.

5. The elevator fault diagnosis method based on a neural network optimized by dynamic step PSO according to claim 1, wherein, the step S7 reduces the error by the method of gradient descent until the error is less than the set target error value, and the training of the BP neural network ends.

Images