Driving steady state control method and control system based on deep belief network

Abstract

The invention discloses a driving steady-state control method and system based on a deep belief network, and belongs to the technical field of steady-state control of an electric forklift driving system.The method comprises the steps that all data of an electric forklift are collected, and noise reduction processing is conducted on original data through a wavelet soft threshold method; carrying out orthogonal wavelet decomposition on the denoised data, carrying out convolution interlaced point sampling, and finally extracting time domain features and frequency domain features of the data; eliminating the influence of the value range and dimension difference of the time domain and frequency domain features on the deep belief network model through a minimum-maximum normalization method; and obtaining a model layer structure and parameters of steady state control based on the deep belief network according to the normalized time domain and frequency domain feature vectors. The method has the advantage that the dynamic performance and the steady-state precision of the starting system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of steady-state control systems for electric forklift drive systems, and specifically to a drive steady-state control method and control system based on a deep belief network. Background Art

[0002] Electric forklifts operate under complex conditions, with relatively frequent starting and braking. This requires the drive control system to adapt to these frequent dynamic changes, ensuring stable power output under various operating conditions. The impact of battery life must also be considered, as frequent starting and braking can damage the battery through overpower and overcurrent. Furthermore, electric forklifts operate under varying ground conditions and loads in various work scenarios, such as warehouses and factory floors. This requires the drive control system to be able to make real-time adjustments based on actual conditions to ensure the forklift's driving stability and safety.

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

[0004] There are common control algorithms available, such as vector control and fuzzy control, but these algorithms cannot solve the problems of nonlinearity and strong coupling in the electric forklift drive system in a timely manner. The dynamic performance and steady-state accuracy of the drive system need to be further improved. Summary of the Invention

[0005] The present invention aims to solve the technical problems mentioned in the above background technology and provide a driving steady-state control method based on a deep belief network to improve the dynamic performance and steady-state accuracy of the starting system.

[0006] The second aspect provides a control system and proposes a three-layer system architecture as the operating basis of the deep belief network-based drive steady-state control method of the first aspect.

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

[0008] S1. Collect various data of the electric forklift, including motor current, motor speed, motor torque, battery voltage, battery state of charge (SOC), battery temperature, battery health status (SOH), cargo weight, lifting height, hydraulic system pressure, etc.

[0009] S2, using wavelet soft threshold method to reduce noise of original data;

[0010] S3, decompose the noise-reduced data by orthogonal wavelet decomposition, and then use it with a low-pass filter Convolution, extract low-frequency components; and high-pass filter Convolution, extract high-frequency components; sample the convolution result at every other point to obtain low-frequency coefficients with half the length and high frequency coefficients ; For low frequency coefficients and high frequency coefficients Repeat convolution sampling to obtain deeper low-frequency and high-frequency components, and finally extract the time domain and frequency domain features of the data;

[0011] S4. Eliminate the impact of the value range and dimension differences of time domain and frequency domain features on the deep belief network model through the minimum-maximum normalization method;

[0012] S5. Determine the vector dimension of the input feature and the number of neurons in the visible layer based on the normalized time domain and frequency domain feature vectors and the number of hidden layer neurons , construct the first RBM layer, which is a two-layer neural network consisting of a visible layer v and a hidden layer h. There is no connection within the layer and full connection between layers;

[0013] S6. Pre-train the first RBM layer using the CD algorithm;

[0014] S7. After pre-training the first RBM layer, stack the RBM layers and use the hidden layer output of the first RBM layer as the visible layer input of the second RBM layer. Repeat the RBM layer construction in step S5 and the RBM layer pre-training in step S6 until all RBM layers are constructed. Then, use the CD algorithm to iterate the training data multiple times to continuously update the RBM weights and bias parameters to make the model better fit the data.

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

[0016] Further,

[0017] S2, using 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; if the absolute value is less than or equal to the threshold, the coefficient is set to zero, thereby ensuring that the smaller coefficients caused by noise are removed by the soft threshold processing method, as shown in the following formula:

[0019]

[0020]

[0021] Where: is the wavelet coefficient; is the threshold; is the wavelet coefficient The symbolic function of for and 0, whichever is greater; when Greater than threshold When and multiply by its sign; is the new wavelet coefficient; threshold The relationship between denoising effect and retaining signal characteristics can be balanced, and the threshold The selection of is determined using the Donoho and Johnstone empirical formula, as shown in the following formula:

[0022]

[0023] Where: is the noise standard deviation; is the signal length.

[0024] Further,

[0025] S3, the noise-reduced data is filtered by a low-pass filter. , high-pass filter And the orthogonal condition formula is used for orthogonal wavelet decomposition. The orthogonal condition formula is as follows:

[0026] .

[0027] Further,

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

[0029] ,

[0030] Where: is the normalized value, is the original eigenvalue, 、 The maximum and minimum values ​​of this feature.

[0031] Further,

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

[0033] ;

[0034] Where: and are the number of neurons in the visible layer and the hidden layer respectively; Is the visible layer The bias of a neuron, is a hidden neuron, For visible neurons, The hidden layer The bias of a neuron, Is the visible layer neurons and hidden layers The weights between neurons;

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

[0036] ;

[0037] Where: is the partition function;

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

[0039] ;

[0040] Where: is the sigmoid function;

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

[0042] ;

[0043] Where: is the sigmoid function.

[0044] Further,

[0045] S6. Pre-train the first RBM layer using the CD algorithm;

[0046] S61, forward propagation: according to the input data Calculate the activation probability of the hidden layer , sampled ;

[0047] S62, Back Propagation: According to Calculate the activation probability of the visible layer and sample it , sampled ;

[0048] S63, repeat steps S61 and S62 in a loop, repeat k times, and obtain 、 ;

[0049] S64. Update parameters according to the following formula:

[0050]

[0051] Where: is the learning rate, represents the expectation on the training data, Represents expectations on the reconstructed data.

[0052] Further,

[0053] S8. Introduce the 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] Where: l is the number of layers, is the visible layer bias, , To hide the 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 objective,

[0058] ;

[0059] Where: N is the number of training samples, is the i-th input sample, For DBN based parameters The reconstructed output of

[0060] By setting a threshold , and update the PSO speed according to the following formula and location Information,

[0061] ; ;

[0062] Where: is the inertia weight, which usually decays linearly with iteration; is the learning factor; A random number between [0,1]; is the historical optimal position of the particle, and the corresponding formula is ; is the optimal position of all particles;

[0063] At the same time, in order to prevent parameter overflow, the particle position is restricted:

[0064] ;

[0065] Adjust the sensitivity of DBN parameters;

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

[0067] According to the second aspect, the present invention provides a technical solution as follows: a control system comprising a processor and a computer-readable storage medium electrically connected to each other, the computer-readable storage medium being suitable for storing a plurality of program codes, the program codes being suitable for being loaded and run by the processor to execute the driving steady-state control method based on a deep belief network provided in the first aspect;

[0068] Furthermore, it also includes a three-layer system architecture, which includes a data acquisition layer, a deep belief network (DBN) model layer and a controller layer that are connected in communication in sequence. The data acquisition layer collects data related to the electric forklift drive system, battery system and load system respectively. The data collected by the data acquisition layer is used to construct the DBN model layer. The controller layer is electrically connected to the processor, and constructs a DBN adaptive closed-loop control strategy based on control targets such as motor speed, torque, and energy consumption to predict the status of parameters such as motor speed and torque.

[0069] Furthermore, the drive system information includes motor current, motor speed, etc., the battery system information includes voltage, battery charge SOC, temperature, battery health state (SOH), etc., and the load system information includes cargo weight, lifting height, hydraulic system pressure, etc. At the same time, the wavelet noise reduction method is used for preliminary data cleaning;

[0070] The DBN model layer includes the bottom RBM (restricted Boltzmann machine), the middle RBM, the high-level RBM, and the output layer;

[0071] Use the underlying RBM to process the raw sensor data of the data acquisition layer and extract time domain / frequency domain features;

[0072] The middle-level RBM is used to process the coupling relationship between multiple source data such as fusion current and speed, and learn complex feature combinations;

[0073] High-level RBMs are used to monitor global steady-state characteristics such as motor operating modes and load mutation responses, while training is performed using the Contrastive Divergence (CD) algorithm.

[0074] The advantages of this invention over existing technologies include: a DBN adaptive closed-loop control strategy is constructed based on control targets such as motor speed, torque, and energy consumption, predicting the states of parameters such as motor speed and torque. This reduces speed and torque fluctuations during the driving process, startup, or sudden load changes of new energy forklifts, thereby enhancing operational smoothness and parameter robustness.

[0075] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 1 is a schematic diagram of a framework of a driving steady-state control system according to an embodiment of the present invention;

[0077] Figure 2 1 is a schematic diagram of the overall flow of data collection and preprocessing according to an embodiment of the present invention;

[0078] Figure 3 This is a schematic diagram of the overall hierarchical construction of RBM in the DBN model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0079] The present invention is described in further detail below.

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

[0081] Combine Figure 1 、 Figure 2 As shown, this embodiment discloses a steady-state control method and control system based on a deep belief network (DBN) for a drive system of an electric forklift. The control system is generally divided into a data acquisition layer, a deep belief network (DBN) model layer, and a controller layer.

[0082] At the data acquisition layer, the original data of the electric forklift, such as motor current, motor speed, motor torque, battery voltage, battery charge SOC, battery temperature, battery health status SOH, cargo weight, lifting height, and hydraulic system pressure, are collected. The original data are denoised using the 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, thereby ensuring that the small coefficients caused by noise are removed by the soft threshold processing method, as shown in Formula 1:

[0083] (1)

[0084] (2)

[0085] Where: is the wavelet coefficient; is the threshold; is the wavelet coefficient The symbolic function of for and 0, whichever is greater; when Greater than threshold When and multiply by its sign; are the new wavelet coefficients.

[0086] Threshold The relationship between denoising effect and retaining signal characteristics can be balanced, and the threshold The selection of is determined using the Donoho and Johnstone empirical formula, as shown in Formula 3:

[0087] (3)

[0088] Where: is the noise standard deviation; is the signal length.

[0089] The noise-reduced data is filtered by a low-pass filter. , High-pass filter And the orthogonal condition formula (4) is used for orthogonal wavelet decomposition:

[0090] (4)

[0091] By low-pass filter Convolution, extract low-frequency components; and high-pass filter Convolution extracts high-frequency components. Sample the convolution result at every other point to obtain low-frequency coefficients with half the length. and high frequency coefficients ;

[0092] For low frequency coefficients and high frequency coefficients Repeated convolution sampling is performed to obtain deeper low-frequency and high-frequency components, and finally the time domain and frequency domain features of the data are extracted.

[0093] The example is as follows: Assume that the original discrete signal is: x = [4, 2, 6, 8,1], length N=5;

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

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

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

[0097]

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

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

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

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

[0102]

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

[0104] 3. Extract the time domain features of high-frequency components. Take cG1 as an example to calculate the mean and variance:

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

[0106] (2) Variance calculation ;

[0107] 4. Extract the 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 the low-frequency component cA1:

[0109] (2) Variance calculation

[0110] Thus, the time domain characteristics of high and low frequency components are obtained.

[0111] 5. Extract frequency domain features of high-frequency components and calculate high-frequency energy using cG1 as an example:

[0112] ;

[0113] 6. Extract the frequency domain features of low-frequency components and calculate the high-frequency energy using cA1 as an example:

[0114] ;

[0115] The influence of the value range and dimension difference of time domain and frequency domain features on the DBN model is eliminated by the minimum-maximum normalization method. The minimum-maximum normalization method is shown in formula (5).

[0116] (5)

[0117] Where: is the normalized value, is the original eigenvalue, 、 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 feature 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. RBM is a two-layer neural network, including a visible layer v and a hidden layer h. There is no connection within the layer, and the layers are fully connected. The RBM energy function is shown in formula (6).

[0119] (6)

[0120] Where: and are the number of neurons in the visible layer and the hidden layer respectively; Is the visible layer The bias of a neuron, is a hidden neuron, For visible neurons, The hidden layer The bias of a neuron, Is the visible layer neurons and hidden layers The weights between neurons.

[0121] Based on the energy function, the joint probability distribution of RBM is formula (7):

[0122] (7)

[0123] Where: is the partition function.

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

[0125] (8)

[0126] Where: is the sigmoid function.

[0127] Given the visible layer h, the visible layer neurons The probability of activation is formula (9):

[0128] (9)

[0129] Where: is the sigmoid function.

[0130] RBM layer pre-training: Use the CD algorithm to train the first RBM layer. The specific steps are as follows:

[0131] Step 1, forward propagation: according to the input data Calculate the activation probability of the hidden layer , sampled ;

[0132] Step 2, back propagation: According to Calculate the activation probability of the visible layer and sample it , sampled ;

[0133] Step 3: Repeat steps 1 and 2 k times to get 、 ;

[0134] Step 4: Update the parameters according to formula (10):

[0135] (10)

[0136] Where: is the learning rate, represents the expectation on the training data, Represents expectations on the reconstructed data.

[0137] After pre-training the RBM layers, stack the RBM layers, using the hidden layer output of the first RBM layer as the visible layer input of the second RBM layer. Repeat the RBM layer construction and RBM layer pre-training steps until all RBM layers are built. The CD algorithm iterates through the training data multiple times, continuously updating the RBM weights and bias parameters to ensure the model better fits the data.

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

[0139] (11)

[0140] Where: l is the number of layers, is the visible layer bias, , To hide the bias, is the weight matrix of the visible layer and the hidden layer;

[0141] Define the mean square reconstruction error (MSE) as the PSO optimization objective,

[0142] (12)

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

[0144] By setting a threshold , and according to formula (13), the PSO speed and position information is:

[0145] (13)

[0146] Where: is the inertia weight, which usually decays linearly with iteration; is the learning factor; A random number between [0,1]; is the historical optimal position of the particle, corresponding to formula (14) ; is the optimal position of all particles.

[0147] At the same time, in order to prevent parameter overflow, the particle position is restricted:

[0148] (14)

[0149] It is the sensitivity adjustment of DBN parameters.

[0150] The optimal parameters obtained by PSO algorithm optimization Mapped back to each layer of DBN as pre-training parameters, they can be subsequently fine-tuned in a supervised manner through back propagation (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, and control parameter mapping is performed, such as motor current, motor speed, battery voltage, battery state of charge (SOC), and other related parameters. 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 actuator 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: Because forklift operating conditions may change, such as changes in load or road conditions, the DBN steady-state control model needs to be adaptive. This can be achieved by regularly updating training data and retraining the model, or by using online learning methods, allowing the model to adjust parameters in real time based on new data to adapt to changing operating conditions and maintain steady-state operation of the drive system.

[0153] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the claims and their equivalents.

Claims

1. A driving steady-state control method based on a deep belief network, characterized in that: include: S1. Collect various data of the electric forklift, including motor current, motor speed, motor torque, battery voltage, battery state of charge (SOC), battery temperature, battery health status (SOH), cargo weight, lifting height, hydraulic system pressure, etc. S2, using wavelet soft threshold method to reduce noise of original data; S3, decompose the noise-reduced data by orthogonal wavelet decomposition, and then use it with a low-pass filter Convolution, extract low-frequency components; and high-pass filter Convolution, extract high-frequency components; sample the convolution result at every other point to obtain low-frequency coefficients with half the length and high frequency coefficients ; For low frequency coefficients and high frequency coefficients Repeat convolution sampling to obtain deeper low-frequency and high-frequency components, and finally extract the time domain and frequency domain features of the data; S4. Eliminate the impact of the value range and dimension differences of time domain and frequency domain features on the deep belief network model through the minimum-maximum normalization method; S5. Determine the vector dimension of the input feature and the number of neurons in the visible layer based on the normalized time domain and frequency domain feature vectors and the number of hidden layer neurons , construct the first RBM layer, which is a two-layer neural network consisting of a visible layer v and a hidden layer h. There is no connection within the layer and full connection between layers; S6. Pre-train the first RBM layer using the CD algorithm; S7. After pre-training the first RBM layer, stack the RBM layers and use the hidden layer output of the first RBM layer as the visible layer input of the second RBM layer. Repeat the RBM layer construction in step S5 and the RBM layer pre-training in step S6 until all RBM layers are constructed. Then, use the CD algorithm to iterate the training data multiple times to continuously update the RBM weights and bias parameters to make the model better fit the data. S8. Introduce the particle swarm optimization (PSO) algorithm to optimize the weight parameters and bias parameters of the deep belief network model; the optimal parameters obtained by the PSO algorithm are Mapping back to each layer of the deep belief network model, as the pre-training parameters of step S6, can be subsequently supervised fine-tuned through back propagation (BP) to obtain the model layer structure and parameters of 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, using wavelet soft threshold method to denoise the original data; For each wavelet coefficient, if its absolute value exceeds a preset threshold, the coefficient is retained but its amplitude is reduced to the threshold; if the absolute value is less than or equal to the threshold, the coefficient is set to zero, thereby ensuring that the smaller coefficients caused by noise are removed by the soft threshold processing method, as shown in the following formula: ； ； Where: is the wavelet coefficient; is the threshold; is the wavelet coefficient The symbolic function of for and 0, whichever is greater; when Greater than threshold When and multiply by its sign; is the new wavelet coefficient; threshold The relationship between denoising effect and retaining signal characteristics can be balanced, and the threshold The selection of is determined using the Donoho and Johnstone empirical formula, as shown in the following formula: ； Where: 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 noise-reduced data is filtered by a low-pass filter. , high-pass filter And the orthogonal condition formula is used for orthogonal wavelet decomposition. 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 features on the deep belief network model is eliminated by the minimum-maximum normalization method. The minimum-maximum normalization method is shown in the following formula: ， Where: is the normalized value, is the original eigenvalue, 、 The maximum and minimum values ​​of this feature.

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

6. The driving steady-state control method based on deep belief network according to claim 1, characterized in that: S6. Pre-train the first RBM layer using the CD algorithm; S61, forward propagation: according to the input data Calculate the activation probability of the hidden layer , sampled ; S62, Back Propagation: According to Calculate the activation probability of the visible layer and sample it , sampled ; S63, repeat steps S61 and S62 in a loop, repeat k times, and obtain 、 ; S64. Update parameters according to the following formula: ； Where: is the learning rate, represents the expectation on the training data, Represents expectations on the reconstructed data.

7. The driving steady-state control method based on deep belief network according to claim 1, characterized in that: S8. Introduce the particle swarm optimization (PSO) algorithm to optimize the weight parameters and bias parameters of the deep belief network model; Expand all weights and biases of the deep belief network into vectors , initialize the position and velocity of the particle swarm as follows: ； Where: l is the number of layers, is the visible layer bias, , To hide the bias, is the weight matrix of the visible layer and the hidden layer; Define the mean square reconstruction error (MSE) as the PSO optimization objective, ； Where: N is the number of training samples, is the i-th input sample, For DBN based parameters The reconstructed output of By setting a threshold , and update the PSO speed according to the following formula and location Information, ； ； Where: is the inertia weight, which usually decays linearly with iteration; is the learning factor; A random number between [0,1]; is the historical optimal position of the particle, and the corresponding formula is ; is the optimal position of all particles; At the same time, in order to prevent parameter overflow, the particle position is restricted: ； Adjust the sensitivity of DBN parameters; The optimal parameters obtained by PSO algorithm optimization Mapping back to each layer of the deep belief network model, as the pre-training parameters of step S6, can be subsequently supervised fine-tuned through back propagation (BP) to obtain the model layer structure and parameters of steady-state control based on the deep belief network.

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

9. The control system according to claim 8, characterized in that: It also includes a three-layer system architecture, which includes a data acquisition layer, a deep belief network (DBN) model layer and a controller layer that are communicatively connected in sequence. The data acquisition layer respectively collects data related to the electric forklift drive system, battery system and load system. The data collected by the data acquisition layer is used to construct the DBN model layer. The controller layer is electrically connected to the processor, and constructs a DBN adaptive closed-loop control strategy according to control targets such as motor speed, torque, and energy consumption to predict the states of parameters such as motor speed and torque.

10. The control system according to claim 9, characterized in that: The drive system information includes motor current, motor speed, etc.; the battery system information includes voltage, battery state of charge (SOC), temperature, battery state of health (SOH), etc.; the load system information includes cargo weight, lifting height, hydraulic system pressure, etc.; and preliminary data cleaning is performed using wavelet noise reduction method; The DBN model layer includes the bottom RBM (restricted Boltzmann machine), the middle RBM, the high-level RBM and the output layer; Use the underlying RBM to process the raw sensor data of the data acquisition layer and extract time domain / frequency domain features; The middle-level RBM is used to process the coupling relationship between multiple source data such as fusion current and speed, and learn complex feature combinations; High-level RBMs are used to monitor global steady-state characteristics such as motor operating modes and load mutation responses, while training is performed using the Contrastive Divergence (CD) algorithm.

Images