A driving steady-state control method and control system based on deep belief networks

Abstract

This invention discloses a drive steady-state control method and control system based on deep belief networks, belonging to the field of steady-state control technology for electric forklift drive systems. The method involves collecting various data from the electric forklift and denoising the original data using wavelet soft thresholding. The denoised data is then subjected to orthogonal wavelet decomposition and convolutional sampling with intervals to extract time-domain and frequency-domain features. A minimum-maximum normalization method is used to eliminate the influence of differences in the value range and dimensions of the time-domain and frequency-domain features on the deep belief network model. Based on the normalized time-domain and frequency-domain feature vectors, the model layer structure and parameters for steady-state control based on the deep belief network are obtained. The advantages of this invention are improved dynamic performance and steady-state accuracy of the drive system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of steady-state control system of electric forklift driving system, in particular to a driving steady-state control method and control system based on deep belief network. BACKGROUND

[0002] The working conditions of electric forklifts are complex, and starting and braking are relatively frequent. This requires the driving control system to adapt to such frequent dynamic changes and ensure stable power output under different working conditions. In addition, the use of the battery life is also taken into account. Frequent starting and braking may cause over-power and over-current damage to the battery. In addition, electric forklifts need to run in different ground conditions and load situations in different working scenarios such as warehouses and factory workshops, and the driving control system needs to be adjusted in real time according to the actual situation to ensure the stability and safety of the forklift.

[0003] The development of current MOS tube, microprocessor and other device technologies has also greatly improved the computing power and control accuracy of the controller. Control algorithms based on hardware development can implement more complex control algorithms and functions.

[0004] The existing common control algorithms, such as vector control and fuzzy control, cannot timely solve the problems of nonlinearity and strong coupling in the driving system of electric forklifts, and the dynamic performance and steady-state accuracy of the driving system need to be further improved. SUMMARY

[0005] The present application aims to solve the technical problems mentioned in the background. A driving steady-state control method based on deep belief network is provided to improve the dynamic performance and steady-state accuracy of the starting system.

[0006] The second aspect provides a control system, which proposes a three-layer system architecture as the running basis of the driving steady-state control method based on deep belief network of the first aspect.

[0007] For the above purposes, the first aspect of the present application provides a technical solution: a driving steady-state control method based on deep belief network, comprising:

[0008] S1, collecting various data of the electric forklift, including collecting motor current, motor speed, motor torque, battery voltage, battery SOC, battery temperature, battery SOH, cargo weight, lifting height, hydraulic system pressure, etc.

[0009] S2, using wavelet soft threshold method to process the original data;

[0010] S3, decomposing the denoised data by orthogonal wavelet, and comparing it with the low-pass filter Convolution, extract low frequency component; with high-pass filter Convolution, extract high frequency component; on the convolution result point sampling, get half length low frequency coefficient And high frequency coefficient ; on low frequency coefficient And high frequency coefficient Repeat the convolution point sampling, get deeper low frequency and high frequency component, finally extract the time domain feature and frequency domain feature of data;

[0011] S4, eliminate the influence of the value range and dimension difference of time domain and frequency domain features on deep belief network model by minimum-maximum normalization method;

[0012] S5, according to the normalized time domain and frequency domain feature vectors, determine the vector dimension, visible layer neuron number And hidden layer neuron number , build the first RBM layer, the first RBM layer is a two-layer neural network, including visible layer v and hidden layer h, no connection within the layer, full connection between layers;

[0013] S6, pre-training the first RBM layer using CD algorithm;

[0014] S7, after pre-training the first RBM layer, stack the RBM layer, take the hidden layer output of the first RBM layer as the visible layer input of the second RBM layer, repeat the RBM layer construction of step S5 and the RBM layer pre-training of step S6 until all RBM layers are constructed; and through CD algorithm in multiple iterations of training data, constantly update the weight and bias parameters of RBM, so that the model can better fit the data;

[0015] S8, introduce particle swarm optimization (PSO) algorithm to optimize the weight parameters and bias parameters of deep belief network model; the optimal parameters Mapping back to each layer of deep belief network model, as the pre-training parameters of step S6, subsequent supervised fine-tuning can be carried out through back propagation (BP), so as to obtain the model layer structure and parameters of steady-state control based on deep belief network.

[0016] Further,

[0017] S2, use wavelet soft threshold method to denoise the original data;

[0018] For each wavelet coefficient, if its absolute value exceeds a preset threshold, the coefficient is retained but its amplitude is reduced to the threshold value; if the absolute value is less than or equal to the threshold, the coefficient is set to zero, thereby ensuring that the method of soft threshold processing removes smaller coefficients caused by noise, i.e., as shown in the following formula:

[0019]

[0020]

[0021] In the formula: is a wavelet coefficient; is a threshold value; is a wavelet coefficient is a sign function of the wavelet coefficient is a maximum of and 0; when is greater than the threshold value , the output is and multiplied by its sign; is a new wavelet coefficient; the threshold value can balance the relationship between the denoising effect and the preservation of signal characteristics, and the selection of the threshold value is determined by using the empirical formula of Donoho and Johnstone, as shown in the following formula:

[0022]

[0023] In the formula: is a noise standard deviation; is a signal length.

[0024] Further,

[0025] S3, the data after noise reduction is decomposed into orthogonal wavelets according to a low-pass filter , a high-pass filter and an orthogonal condition formula as follows:

[0026] .

[0027] Further,

[0028] S4, the influence of the value range and the dimension difference of the time domain and frequency domain characteristics on the deep belief network model is eliminated by a minimum-maximum normalization method, as shown in the following formula.

[0029] ,

[0030] In the formula: is a normalized value, is an original characteristic value, 、 The maximum and minimum values of the feature.

[0031] Further,

[0032] The energy function of the first RBM layer in S5 is shown in the following equation:

[0033] ;

[0034] In the equation: and are the number of neurons in the visible layer and hidden layer, respectively; is the bias of the th neuron in the visible layer, is the hidden neuron, is the visible neuron, is the bias of the th neuron in the hidden layer, is the weight between the th neuron in the visible layer and the th neuron in the hidden layer;

[0035] Based on the energy function, the joint probability distribution of the RBM is as follows:

[0036] ;

[0037] In the equation: is the partition function;

[0038] Given the visible layer v, the probability of the activation of the hidden layer neuron is as follows:

[0039] ;

[0040] In the equation: is the sigmoid function;

[0041] Given the visible layer h, the probability of the activation of the visible layer neuron is as follows:

[0042] ;

[0043] In the equation: is the sigmoid function.

[0044] Further,

[0045] S6, pre-training the first RBM layer using the CD algorithm;

[0046] S61, forward propagation: according to the input data Compute activation probabilities of hidden layers , sample to get ;

[0047] S62, Back propagation: according to Compute activation probabilities of visible layers, and sample to get , sample to get ;

[0048] S63, repeat steps S61 and S62 for k times, get , ;

[0049] S64, update parameters according to the following formula:

[0050]

[0051] In the formula: is the learning rate, represents the expectation on the training data, represents the expectation on the reconstructed data.

[0052] Further,

[0053] S8, introduce particle swarm optimization (PSO) algorithm to optimize the weight parameters and bias parameters of the deep belief network model;

[0054] Expand all weights and biases of the deep belief network into vectors Initialize the position and velocity of the particle swarm as follows:

[0055] ;

[0056] In the formula: l is the number of layers, is the bias of the visible layer, , is the hidden bias, is the weight matrix of the visible layer and the hidden layer;

[0057] Define the mean square reconstruction error (MSE) as the PSO optimization target,

[0058] ;

[0059] In the formula: N is the number of training samples, is the i-th input sample, is the reconstruction output of the DBN based on the parameter ;

[0060] By setting the threshold , update the PSO velocity and position information of the driving system,

[0061] ; ;

[0062] wherein: is an inertia weight, usually decaying linearly with iterations; is a learning factor; is a random number between 0 and 1; is the particle's historical best position, corresponding to the formula ; is the global best position of all particles;

[0063] At the same time, in order to prevent parameter overflow, the position of the particle is limited by the boundary:

[0064] ;

[0065] is the DBN parameter sensitivity adjustment;

[0066] The optimal parameters obtained by the PSO algorithm are mapped back to each layer of the deep belief network model as the pre-training parameters of step S6, and subsequent supervised fine-tuning can be performed through back propagation (BP), so as to obtain the model layer structure and parameters of the steady-state control based on the deep belief network.

[0067] According to the content of the second aspect, the technical scheme provided by the application is: a control system comprising a processor and a computer readable storage medium electrically connected to each other, the computer readable storage medium is adapted to store a plurality of program codes, the program codes are adapted to be loaded and run by the processor to execute the deep belief network based driving steady-state control method provided by the first aspect;

[0068] Further, it further comprises a three-layer system architecture, the three-layer system architecture comprises a data acquisition layer, a deep belief network (DBN) model layer and a controller layer which are sequentially communicatively connected, the data acquisition layer respectively acquires data related to an electric forklift driving system, a battery system and a load system, the data acquired by the data acquisition layer is used to construct the DBN model layer, the controller layer is electrically connected to the processor, a DBN adaptive closed-loop control strategy is constructed according to control targets such as motor speed, torque and energy consumption, and parameters such as motor speed and torque are predicted.

[0069] Further, the information of the driving system comprises motor current and motor speed, the information of the battery system comprises voltage, battery SOC, temperature, SOH and the like, and the information of the load system comprises cargo weight, lifting height, hydraulic system pressure and the like, and wavelet denoising method is used for preliminary data cleaning;

[0070] The DBN model layer comprises a bottom layer RBM (Restricted Boltzmann Machine), a middle layer RBM, a high layer RBM and an output layer;

[0071] The bottom layer RBM is used to process original sensor data of the data acquisition layer to extract time domain / frequency domain features.

[0072] The middle layer RBM is used to process coupling relationships of multi-source data such as fused currents and rotating speeds to learn complex feature combinations.

[0073] The high layer RBM is used to monitor global steady state characteristics such as motor working modes and load mutation responses, and is trained through a contrastive divergence (CD) algorithm.

[0074] Compared with the prior art, the application has the advantages that a DBN adaptive closed-loop control strategy is constructed according to motor rotating speeds, torques, energy consumptions and other control targets to predict parameter states such as motor rotating speeds and torques, so that rotating speed fluctuations, torque fluctuations and other problems during a new energy forklift driving process, at the time of starting or load mutation are reduced, and operation fluency and parameter robustness are enhanced.

[0075] Additional aspects and advantages of the application will be described in part below, will become apparent from the following description, or will be learned by practicing the application. BRIEF DESCRIPTION OF DRAWINGS

[0076] Figure 1 Fig. 1 is a drive steady state control system framework schematic diagram of an embodiment of the application;

[0077] Figure 2 Fig. 3 is a data acquisition and preprocessing overall flow schematic diagram of an embodiment of the application;

[0078] Figure 3 Fig. 5 is a DBN model layered building RBM overall schematic diagram of an embodiment of the application. DETAILED DESCRIPTION

[0079] The application will be further described in detail below.

[0080] In order to make the purpose, technical scheme and advantages of the embodiments of the application clearer, the technical scheme in the embodiments will be clearly and completely described below in combination with the drawings of the embodiments of the application, and the following embodiments are used to illustrate the application, but not to limit the scope of the application.

[0081] In combination with Figure 1 , Figure 2 As shown in the drawings, the embodiment discloses a deep belief network (DBN) applied to a drive system steady state control method and control system of an electric forklift, and the control system is divided into a data acquisition layer, a deep belief network (DBN) model layer and a controller layer.

[0082] In the data acquisition layer, raw data such as motor current, motor speed, motor torque, battery voltage, battery SOC, battery temperature, battery health status SOH, cargo weight, lifting height, and hydraulic system pressure of the electric forklift are collected. The raw data is then denoised using a wavelet soft thresholding method: for each wavelet coefficient, if its absolute value exceeds a preset threshold, the coefficient is retained but its amplitude is reduced to the threshold value; if the absolute value is less than or equal to the threshold, the coefficient is set to zero. This ensures that smaller coefficients caused by noise are removed using soft thresholding, as shown in Formula 1.

[0083] (1)

[0084] (2)

[0085] In the formula: These are wavelet coefficients; For threshold; wavelet coefficients The sign function; for The larger of 0 and 0; when Greater than the threshold When, the output is And multiply by its sign; These are the new wavelet coefficients.

[0086] threshold It can balance the relationship between denoising effect and preservation of signal features, threshold The selection of is determined using the empirical formula of Donoho and Johnstone, as shown in Formula 3:

[0087] (3)

[0088] In the formula: The standard deviation of noise; This is the signal length.

[0089] The noise-reduced data is then processed using a low-pass filter. High-pass filter Orthogonal wavelet decomposition is performed using the orthogonality condition formula (4):

[0090] (4)

[0091] By using a low-pass filter Convolution, extracting low-frequency components; combined with a high-pass filter. Convolution is used to extract high-frequency components. The convolution result is sampled at intervals to obtain low-frequency coefficients of half their length. and high frequency coefficient ;

[0092] For low-frequency coefficients and high frequency coefficients By repeatedly performing convolutional sampling at intervals, deeper low-frequency and high-frequency components are obtained, and finally, the time-domain and frequency-domain features of the data are extracted.

[0093] Example: Assume the original discrete signal is: x = [4, 2, 6, 8, 1], with a length N = 5;

[0094] 1. Low-pass filtering and intermittent sampling to extract low-frequency components:

[0095] (1) Set the low-pass filter as follows: ;

[0096] (2) Convolution calculation: Perform linear convolution between the low-pass filter and the signal x:

[0097]

[0098] (3) Take the even-numbered index position (starting from 0) of the convolution result to obtain the low-frequency components: ;

[0099] 2. High-pass filtering and interval sampling to extract high-frequency components:

[0100] (1) Set the high-pass filter as follows: ;

[0101] (2) Convolution calculation: Perform linear convolution between the high-pass filter and the signal x:

[0102]

[0103] (3) Take the even-numbered index position (starting from 0) of the convolution result to obtain the high-frequency components: ;

[0104] 3. High-frequency component temporal feature extraction: Taking cG1 as an example, the mean and variance are calculated respectively.

[0105] (1) Calculation of the mean value of low-frequency component cG1: ;

[0106] (2) Variance calculation ;

[0107] 4. Extraction of time-domain features of low-frequency components, taking cA1 as an example to calculate the mean and variance.

[0108] (1) Calculation of the mean value of low-frequency component cA1:

[0109] (2) Variance calculation

[0110] This yielded the time-domain characteristics of the high and low frequency components.

[0111] 5. High-frequency component frequency domain feature extraction, taking cG1 as an example to calculate high-frequency energy:

[0112] ;

[0113] 6. Extraction of low-frequency components in the frequency domain, taking cA1 as an example to calculate high-frequency energy:

[0114] ;

[0115] The influence of the differences in the value range and dimensions of time-domain and frequency-domain features on the DBN model is eliminated by the min-max normalization method, as shown in equation (5).

[0116] (5)

[0117] In the formula: The normalized value, These are the original eigenvalues. , These are the maximum and minimum values ​​of the feature, respectively.

[0118] RBM layer construction: In the DBN model training and construction subsystem, the dimension of the input features and the number of hidden layer neurons are determined based on the normalized time-domain and frequency-domain feature vectors. and the number of neurons in the visible layer Construct the first RBM layer. The RBM is a two-layer neural network containing a visible layer v and a hidden layer h. There are no connections within the layer, but full connections between the layers. The RBM energy function is shown in equation (6).

[0119] (6)

[0120] In the formula: and These represent the number of neurons in the visible layer and the hidden layer, respectively. It is the visible layer. Bias of each neuron To hide neurons, Visible neurons It is the hidden layer. Bias of each neuron It is the visible layer. The first neuron and the hidden layer The weights between neurons.

[0121] Based on the energy function, the joint probability distribution of RBMs is given by equation (7):

[0122] (7)

[0123] In the formula: It is a partition function.

[0124] Given the visible layer v, the hidden layer neurons The probability of activation is given by equation (8):

[0125] (8)

[0126] In the formula: This is the sigmoid function.

[0127] Given a visible layer h, the visible layer neurons The probability of activation is given by equation (9):

[0128] (9)

[0129] In the formula: This is the sigmoid function.

[0130] RBM layer pre-training: The first RBM layer is trained using the CD algorithm. The specific steps are as follows:

[0131] Step 1, Forward Propagation: Based on the input data Calculate the activation probability of the hidden layer Sampling ;

[0132] Step 2, Backpropagation: According to Calculate the activation probability of the visible layer and sample it. Sampling ;

[0133] Step 3: Repeat steps 1 and 2 k times to obtain... , ;

[0134] Step 4, update the parameters according to equation (10):

[0135] (10)

[0136] In the formula: For learning rate, This represents the expectation on the training data. This indicates the expectation regarding the reconstruction of data.

[0137] After pre-training the RBM layers, the RBM layers are stacked, with the output of the hidden layer of the first RBM layer used as the input of the visible layer of the second RBM layer. This process of building and pre-training RBM layers is repeated until all RBM layers are built. The CD algorithm iterates multiple times on the training data, continuously updating the weights and bias parameters of the RBMs to enable the model to better fit the data.

[0138] Fine-tuning the DBN model: Particle swarm optimization (PSO) algorithm is introduced to optimize the weight and bias parameters of the DBN model, expanding all weights and biases of the DBN into vectors. Initialize the position and velocity of the particle swarm as shown in equation (11):

[0139] (11)

[0140] In the formula: l is the number of layers, For visible layer bias, , To hide the bias, This is the weight matrix for the visible and hidden layers;

[0141] The mean squared reconstruction error (MSE) is defined as the PSO optimization objective.

[0142] (12)

[0143] In the formula: N is the number of training samples, For the i-th input sample, DBN based on parameters The reconstructed output.

[0144] By setting a threshold And according to formula (13), the PSO velocity and position information are further obtained.

[0145] (13)

[0146] In the formula: The inertial weight typically decays linearly with iteration; For learning factors; A random number between [0, 1]; For the optimal position in the particle's history, the corresponding equation is (14). ; This is the optimal position for all particles.

[0147] At the same time, to prevent parameter overflow, the particle positions are subject to boundary constraints:

[0148] (14)

[0149] Adjust the sensitivity of DBN parameters.

[0150] The optimal parameters were obtained through the PSO algorithm. The parameters are mapped back to the layers of the DBN and used as pre-training parameters. They can then be fine-tuned in a supervised manner through backpropagation (BP) to obtain the model layer structure and parameters of the DBN steady-state control system.

[0151] In the steady-state control subsystem module, various state parameters of the electric forklift are monitored and fed back in real time. Control parameters are mapped, such as motor current, motor speed, battery voltage, and battery SOC. These parameters are then input into the trained DBN steady-state control model. The model calculates the corresponding control output based on the current input state and sends the control signal to the actuators of the drive system, such as the motor controller, to adjust the operating state of the drive system and enhance the steady-state regulation of specific parameters of the control system.

[0152] Adaptive Adjustment: Since forklift operating conditions may change, such as variations in load or road surface, the DBN steady-state control model needs to have adaptive adjustment capabilities. This can be achieved by periodically updating training data and retraining the model, or by employing online learning methods, allowing the model to adjust parameters in real time based on new data to adapt to different operating conditions and maintain steady-state operation of the drive system.

[0153] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for drive steady state control based on deep belief networks, characterized in that, Comprise: S1, collecting various data of the electric forklift, including collecting motor current, motor speed, motor torque, battery voltage, battery power SOC, battery temperature, battery health state SOH, cargo weight, lifting height, hydraulic system pressure of the electric forklift; S2, the original data is denoised by using wavelet soft threshold method; S3, the data after noise reduction is decomposed by orthogonal wavelet, and is convolved with low-pass filter to extract low-frequency component; and is convolved with high-pass filter to extract high-frequency component; and the convolution result is sampled at intervals to obtain low-frequency coefficients and high-frequency coefficients with half length ; the low-frequency coefficients and the high-frequency coefficients are repeatedly convolved and sampled at intervals to obtain deeper low-frequency component and high-frequency component, and finally the time-domain feature and frequency-domain feature of data are extracted;​ S4, the influence of the value range and dimension difference of time domain and frequency domain characteristics on the deep belief network model is eliminated by using minimum-maximum normalization method; S5, determining the vector dimension of the input feature, the number of visible layer neurons and the number of hidden layer neurons according to the normalized time domain and frequency domain feature vectors and hidden layer neurons , constructing the first RBM layer, the first RBM layer being a two-layer neural network, containing a visible layer v and a hidden layer h, no connection within the layer, and full connection between the layers; S6, the first RBM layer is pre-trained using CD algorithm; S7, after pre-training the first RBM layer, the RBM layer stacking is carried out, the hidden layer output of the first RBM layer is taken as the visible layer input of the second RBM layer, and the RBM layer construction of step S5 and the RBM layer pre-training of step S6 are repeated until all RBM layers are constructed; And through the CD algorithm, the weight and bias parameters of the RBM are updated multiple times in the training data iteration, so that the model can better fit the data; S8, a particle swarm optimization algorithm is introduced to optimize the weight parameters and bias parameters of the deep belief network model; the optimal parameters obtained by the particle swarm optimization algorithm mapping back to the layers of the deep belief network model as the pre-training parameters of step S6, and subsequent supervised fine-tuning can be performed through back propagation, thereby obtaining the model layer structure and parameters of the steady-state control based on the deep belief network.

2. The driving steady state control method based on deep belief network according to claim 1, characterized in that: S2, the original data is denoised by using wavelet soft threshold method; For each wavelet coefficient, if its absolute value exceeds a preset threshold, the coefficient is retained but its amplitude is reduced to the threshold value; if the absolute value is less than or equal to the threshold, the coefficient is set to zero, so as to remove the small coefficients caused by noise by using the soft threshold processing method, that is, as shown in the following formula: ； ； In the formula: These are wavelet coefficients; For threshold; Wavelet coefficients The sign function; for The larger of 0 and 0; when Greater than the threshold When, the output is And multiply by its sign; For the new wavelet coefficients; threshold It can balance the relationship between denoising effect and preservation of signal features, threshold The selection of is determined using the empirical formula of Donoho and Johnstone, as shown in the following formula: ； In the formulae: is the noise standard deviation; is the signal length.

3. The driving steady state control method based on deep belief network according to claim 1, characterized in that: S3, the data after noise reduction is decomposed by orthogonal wavelet according to low-pass filter , high-pass filter and orthogonal condition formula, the orthogonal condition formula is as follows: 。 4. The driving steady state control method based on deep belief network according to claim 1, characterized in that: S4, the influence of the value range and dimension difference of time domain and frequency domain characteristics on the deep belief network model is eliminated by using minimum-maximum normalization method, and the minimum-maximum normalization method is as shown in the following formula: ； In the formula: is a normalized value, is an original characteristic value, , the maximum and minimum values of the characteristic.

5. The deep belief network-based drive steady control method according to claim 1, characterized in that: The energy function of the first RBM layer in S5 is as shown in the following formula: ； where: and are the number of neurons of the visible and hidden layers, respectively; is the bias of the th neuron of the visible layer, is the hidden neuron, is the visible neuron, is the bias of the th neuron of the hidden layer, is the weight between the th neuron of the visible layer and the th neuron of the hidden layer; Based on the energy function, the joint probability distribution of the RBM is as follows: ； wherein: is the partition function; Given a visible layer v, the hidden layer neuron The probability of activation is given by: ； wherein: sigmoid is a sigmoid function; Given a visible layer h, the visible layer neuron The probability of activation is given by: ； In the formula: sigmoid is a sigmoid function.

6. The driving steady state control method based on deep belief network according to claim 1, characterized in that: S6, the first RBM layer is pre-trained using CD algorithm; S61, forward propagation: according to input data computing activation probabilities of hidden layers , sampling to obtain ; S62, back propagation: according to calculate the activation probability of the visible layer, and sample to obtain , sample to obtain ; S63, repeating steps S61 and S62 cyclically, k times, to obtain , ; S64, the parameters are updated according to the following formula: ； ； ； In the formula: is the learning rate, denotes the expectation on the training data, denotes the expectation on the reconstruction data.

7. The driving steady state control method based on deep belief network according to claim 1, characterized in that: S8, the particle swarm optimization algorithm is introduced to optimize the weight parameters and bias parameters of the deep belief network model; unfolding all weights and biases of the deep belief network into vectors initialize the position and velocity of the particle swarm as follows: ； wherein: l is the number of layers, is the bias vector for the lth layer of visible layers, is the bias vector for the lth layer of hidden layers, is the weight matrix for the lth layer of visible and hidden layers; The mean square reconstruction error is defined as the particle swarm optimization target, ； where N is the number of training samples, is the i-th input sample, is the reconstruction output of the DBN based on parameters . By setting a threshold value and updating the particle swarm optimization velocity and position information according to the following formula, ； ； where: is the inertia weight, which decreases linearly with iteration; is the learning factor; is a random number between [0, 1]; is the particle's history best position, corresponding to the equation , is the global best position of all particles; At the same time, in order to prevent parameter overflow, the position of the particle is limited within the boundary: ； a minimum boundary for DBN parameter sensitivity adjustment, a maximum boundary for DBN parameter sensitivity adjustment; The optimal parameters are obtained by a particle swarm optimization algorithm The mapping back to the layers of the deep belief network model as pre-training parameters of step S6, subsequent supervised fine-tuning can be performed by back propagation, thereby obtaining the model layer structure and parameters of the steady-state control based on the deep belief network.

8. A control system characterized by: Comprise a processor and a computer readable storage medium electrically connected to each other, the computer readable storage medium is suitable for storing a plurality of program codes, the program codes are suitable for being loaded and run by the processor to execute the driving steady state control method based on deep belief network of any one of claims 1-7.

9. The control system of claim 8, wherein: Also include a three-layer system architecture, the three-layer system architecture includes data acquisition layer, deep belief network model layer and controller layer which are sequentially connected in communication, the data acquisition layer respectively collects electric forklift driving system, battery system, load system related data, the data collected by the data acquisition layer is used to construct the deep belief network model layer, the controller layer is electrically connected with the processor, a deep belief network adaptive closed-loop control strategy is constructed according to motor speed, torque and energy consumption control target to predict motor parameter state.

10. The control system of claim 9, wherein: The information of the driving system includes motor current and motor speed, the information of the battery system includes voltage, battery power SOC, temperature and battery health state SOH, the information of the load system includes cargo weight, lifting height and hydraulic system pressure, and wavelet denoising method is used for preliminary data cleaning; The deep belief network model layer includes bottom layer RBM, middle layer RBM, high layer RBM and output layer; The bottom layer RBM is used to process original sensor data of the data acquisition layer to extract time domain / frequency domain features; The middle layer RBM is used to process coupling relationship of fused current and speed multi-source data to learn complex feature combination; The high layer RBM is used to monitor motor working mode and load mutation response global steady state characteristics, and is trained through CD algorithm.

Images