Worm gear and worm meshing center distance dynamic automatic analysis method based on servo drive

Abstract

The invention provides a worm gear and worm meshing center distance dynamic automatic analysis method based on servo drive, and relates to the technical field of analysis. Using a neural network adaptive filter to process and correct the center distance data; establishing a kinetic model, calculating an optimal center distance value and obtaining a deviation sequence; predicting a change trend by adopting an attention mechanism enhanced bidirectional gating recurrent neural network; and outputting a compensation instruction to drive the servo motor to execute adjustment. According to the invention, real-time accurate measurement and intelligent dynamic compensation of the worm gear and worm meshing center distance are realized, and the precision and stability of a transmission system are improved.

Description

Technical Field

[0001] The present invention relates to analysis technologies, and in particular to a dynamic automatic analysis method for the meshing center distance of a worm and worm gear based on servo drive. Background Art

[0002] As an important part of mechanical transmissions, worm and worm gear drive systems are widely used in fields such as industrial robots, precision machine tools, and automated production lines. This drive system has advantages such as a large transmission ratio, a compact structure, and smooth transmission. However, the precise control of its meshing center distance directly affects the transmission accuracy, noise level, and service life of the system. In an actual working environment, due to factors such as temperature changes, load fluctuations, and wear, the meshing center distance of the worm and worm gear will change dynamically, resulting in a decline in the performance of the drive system.

[0003] Traditional methods for controlling the meshing center distance of worm and worm gear mainly rely on static adjustment and regular maintenance, and it is difficult to cope with dynamic changes during the working process. In recent years, with the development of intelligent manufacturing and Industry 4.0, the requirements for the accuracy and reliability of drive systems have been continuously increasing, and dynamic adjustment methods based on servo drive have gradually received attention.

[0004] Traditional center distance measurement methods lack real-time performance and accuracy. Usually, contact measurement or simple sensor monitoring is used, and it is impossible to effectively eliminate environmental interference and system noise. Especially in a working environment with large temperature fluctuations, the measurement results have obvious deviations, and it is difficult to provide a reliable basis for subsequent adjustments.

[0005] Existing center distance compensation strategies mostly adopt control methods based on fixed models, and have poor adaptability to system parameter changes and nonlinear characteristics. They cannot dynamically adjust the compensation strategy according to the system state, resulting in inconsistent compensation effects under different working conditions. Especially under extreme working conditions such as high speed and heavy load, the compensation accuracy decreases.

[0006] The existing technologies lack the ability to predict the change trend of the center distance. Most of them adopt passive reactive adjustments, and it is impossible to prospectively identify the change trend of system performance. It is difficult to implement preventive maintenance before the problem expands, resulting in insufficient system operation stability and difficult to guarantee long-term reliability. Summary of the Invention

[0007] Embodiments of the present invention provide a dynamic automatic analysis method for the meshing center distance of a worm and worm gear based on servo drive, which can solve the problems in the existing technologies.

[0008] In the first aspect of the embodiments of the present invention, a dynamic automatic analysis method for the meshing center distance of a worm and worm gear based on servo drive is provided, including: Collect the state parameters of the worm and worm gear drive system to generate an initial measurement dataset; use the initial measurement dataset to train a neural network adaptive filter. The neural network adaptive filter is based on the long short-term memory network architecture, establishes a non-linear mapping relationship with the center distance signal, and performs real-time correction on the center distance measurement result by introducing a temperature compensation coefficient, and outputs the corrected center distance data; According to the corrected center distance data, combined with the worm speed signal and the worm gear speed signal, establish a dynamic model of the worm and worm gear drive system, calculate the theoretical optimal center distance value, compare the theoretical optimal center distance value with the corrected center distance data, and obtain the center distance deviation sequence; Based on the center distance deviation sequence, use a bidirectional gated recurrent neural network enhanced by an attention mechanism to predict the center distance change trend by analyzing the time series characteristics in the historical compensation data, and continuously optimize the compensation strategy in combination with the online transfer learning method, and output the optimal compensation instruction; Convert the optimal compensation instruction into a displacement control signal of the servo motor, drive the servo motor to perform the center distance adjustment operation, and realize the dynamic compensation of the meshing center distance of the worm and worm gear.

[0009] Use the initial measurement dataset to train a neural network adaptive filter. The neural network adaptive filter is based on the long short-term memory network architecture, establishes a non-linear mapping relationship with the center distance signal, and performs real-time correction on the center distance measurement result by introducing a temperature compensation coefficient. The output of the corrected center distance data includes: Collect the real-time operation data of the worm and worm gear drive system, generate a multi-modal measurement dataset based on the real-time operation data; establish a high-fidelity mapping environment of the drive system based on the multi-modal measurement dataset, construct a non-linear mapping relationship between the center distance signal and the multi-modal measurement dataset in the high-fidelity mapping environment, and perform mapping processing on the initial center distance data through the non-linear mapping relationship to obtain the optimized center distance data; According to the optimized center distance data, calculate the temperature compensation coefficient by combining the temperature data in the multi-modal measurement dataset. The temperature compensation coefficient is obtained through an exponential decay function of the initial compensation coefficient and the temperature change amount. Perform weighted fusion on the temperature compensation coefficient and the optimized center distance data, and introduce a dynamic bias term for compensation to generate a compensated center distance measurement value; Input the compensated center distance measurement value into the high-fidelity mapping environment, calculate the compensation parameter based on the historical compensation data in the high-fidelity mapping environment, and perform optimized correction on the compensated center distance measurement value through the compensation parameter, and output the corrected center distance measurement data.

[0010] According to the corrected center distance data, combined with the worm gear rotation speed signal and the worm rotation speed signal, establish a dynamic model of the worm gear drive system, calculate the theoretical optimal center distance value, and compare the theoretical optimal center distance value with the corrected center distance data to obtain the center distance deviation sequence including: Collect the dynamic data set of the worm gear drive system, where the dynamic data set is used to characterize the real-time operating state of the drive system; construct a spiking neural network model based on the dynamic data set. The spiking neural network model simulates the characteristics of the drive system through the sodium ion channel conductance and the potassium ion channel conductance. Based on the sodium ion channel conductance and the potassium ion channel conductance, receive the dynamic data set and calculate the dynamic current, and perform a weighted combination of the dynamic current and the basic driving torque to generate a synaptic modulation torque; Substitute the synaptic modulation torque into the enhanced dynamic equation, and obtain the dynamic response characteristics of the drive system by solving the enhanced dynamic equation; perform time-domain encoding on the dynamic response characteristics through the spiking neural network model to generate a pulse sequence signal. The pulse sequence signal describes the time-varying characteristics of the dynamic characteristics through a time exponential decay function, and obtain the dynamic characteristic pulse encoding; Based on the dynamic characteristic pulse encoding, use the spiking neural network model to construct a time-varying kernel function. The time-varying kernel function is determined by the dynamically reconstructed synaptic structure. Perform spatio-temporal domain convolution operation on the dynamic characteristic pulse encoding through the time-varying kernel function to obtain the instantaneous state characteristics of the drive system, and calculate the theoretical optimal center distance based on the instantaneous state characteristics; Compare the theoretical optimal center distance with the corrected center distance data to obtain the center distance deviation data.

[0011] Substitute the synaptic modulation torque into the enhanced dynamic equation, and obtain the dynamic response characteristics of the drive system by solving the enhanced dynamic equation; perform time-domain encoding on the dynamic response characteristics through the spiking neural network model to generate a pulse sequence signal including: Based on the synaptic modulation torque, construct a field theory enhanced dynamic equation. The field theory enhanced dynamic equation introduces a field-matter coupling force term, and obtain the field state distribution of the drive system by solving the field theory enhanced dynamic equation. The field state distribution reflects the overall dynamic state of the drive system; Substitute the field state distribution into the field theory enhanced dynamic equation, dynamically update the coupling coefficient in the field-matter coupling force term according to the real-time operating state of the drive system, and adjust the parameters of the field theory enhanced dynamic equation in combination with the change law of the synaptic modulation torque. Solve the field theory enhanced dynamic equation again to obtain the coupled dynamic response of the drive system. The coupled dynamic response characterizes the non-linear dynamic behavior of the drive system; Apply a field state projection transformation to the coupled dynamic response, extract the characteristics of the system velocity component, acceleration component, and field state component, and generate a feature vector for field state characterization through a linear combination of characteristic mode functions. The feature vector contains multi-scale dynamic information of the transmission system; Input the feature vector of the field state characterization into a spiking neural network. Through the spiking neural network, perform time-domain encoding on the dynamic characteristics, and introduce a timestamp function and an exponential decay function to modulate the encoding process to generate a spiking sequence signal with time-varying characteristics.

[0012] Based on the center distance deviation sequence, use a bidirectional gated recurrent neural network enhanced by an attention mechanism to predict the center distance change trend by analyzing the temporal characteristics in historical compensation data, and continuously optimize the compensation strategy in combination with the online transfer learning method. The output optimal compensation instruction includes: Construct a bidirectional gated recurrent neural network, input the forward hidden layer state and backward hidden layer state of the bidirectional gated recurrent neural network into a swarm intelligence emergence layer. The swarm intelligence emergence layer introduces a swarm cooperation factor to enhance the hidden layer state and generate an enhanced hidden layer feature; Construct a self-evolving attention weight based on the enhanced hidden layer feature. The self-evolving attention weight calculates a fitness function through the ratio of the prediction error to the reference error threshold, and multiplies the fitness function by the attention base weight to obtain an adaptive attention weight; Establish an environmental pressure selection mechanism according to the adaptive attention weight. The environmental pressure selection mechanism calculates the transfer probability based on the adaptation distance from the source domain to the target domain, uses the transfer probability as the weight coefficient of the source domain loss function, and updates the transfer learning model parameters through gradient descent; Modulate the enhanced hidden layer feature based on the transfer learning model parameters to generate a dynamic feature vector. Input the dynamic feature vector and the adaptive attention weight into a compensation strategy generation module. The compensation strategy generation module introduces an evolutionary modulation factor, and the evolutionary modulation factor is calculated through the cumulative effect of multi-scale selection pressure indicators; Multiply the evolutionary modulation factor by the initial compensation instruction to obtain the optimal compensation instruction, and feedback the compensation effect corresponding to the optimal compensation instruction to the fitness function to update the self-evolving attention weight and realize the dynamic optimization of the compensation strategy.

[0013] Input the dynamic feature vector and the adaptive attention weight into a compensation strategy generation module. The compensation strategy generation module introduces an evolutionary modulation factor, and the evolutionary modulation factor is calculated through the cumulative effect of multi-scale selection pressure indicators, including: Combine the dynamic feature vector with the adaptive attention weights to generate multi-level features, input the multi-level features into a compensation strategy generation module, and introduce a dynamic response enhancement function through the compensation strategy generation module to perform enhancement processing on the multi-level features. The dynamic response enhancement function is adaptively adjusted based on the feature energy level difference to generate an initial compensation strategy, and the initial compensation strategy reflects the dynamic response characteristics of the system; Construct a multi-scale selection pressure index based on the initial compensation strategy. The multi-scale selection pressure index is calculated through the normalized distance between the initial compensation strategy and a reference strategy, and corresponding reference values are set for the reference strategy at different scales; Apply stability protection to the multi-scale selection pressure index, and perform structural enhancement on the multi-scale selection pressure index through a dynamic modulation field. The dynamic modulation field is dynamically adjusted based on the real-time changes of the system state, and continuous stability protection is provided through the closed-path integral of the dynamic modulation field to generate a steady-state enhanced selection pressure index, and the steady-state enhanced selection pressure index has anti-interference ability; Substitute the steady-state enhanced selection pressure index into the evolutionary modulation factor calculation formula, and perform weighted calculation on the steady-state enhanced selection pressure index through different-scale weight coefficients. The weight coefficients are adaptively adjusted as the scale changes, and the coordinated optimization of multi-scale features is achieved through the dynamic allocation of the different-scale weight coefficients to obtain an evolutionary modulation factor.

[0014] Providing continuous stability protection through the closed-path integral of the dynamic modulation field to generate a steady-state enhanced selection pressure index includes: The dynamic modulation field is dynamically adjusted based on the change trend of the real-time state of the system, and the stability requirements of the system are reflected through the change characteristics of the dynamic modulation field; Perform a closed-path integral operation on the dynamic modulation field. The closed-path integral operation is calculated along the closed trajectory of the system state space to generate a stability protection factor, and the stability protection factor is used to characterize the overall stability characteristics of the system; Fuse the stability protection factor with the selection pressure index, and apply a continuity constraint to the selection pressure index through the stability protection factor to generate a steady-state enhanced selection pressure index with anti-interference ability.

[0015] In a second aspect of the embodiments of the present invention, an electronic device is provided, including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0016] In a third aspect of the embodiments of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the foregoing method is implemented.

[0017] The beneficial effects of this application are as follows: The method for dynamically and automatically analyzing the center distance of worm and worm gear meshing based on servo drive provided by the present invention processes measurement data through a neural network adaptive filter and introduces a temperature compensation coefficient, improving the accuracy of center distance measurement and solving the problem of decreased accuracy of traditional measurement methods under temperature change conditions.

[0018] This method combines a dynamic model with measured data, establishes a comparison mechanism between the theoretical optimal center distance and the actual value, can quickly identify center distance deviations, and predicts the center distance change trend through a bidirectional gated recurrent neural network enhanced by an attention mechanism, realizing a more accurate compensation strategy formulation.

[0019] The present invention uses online transfer learning to continuously optimize the compensation strategy and performs real-time adjustment through a servo motor, realizing the adaptive dynamic compensation of the center distance of worm and worm gear meshing, effectively improving the operating stability and service life of the transmission system, reducing noise and vibration, and meeting the real-time control requirements of high-precision transmission occasions. Description of the Drawings

[0020] Figure 1 It is a schematic flow chart of the method for dynamically and automatically analyzing the center distance of worm and worm gear meshing based on servo drive according to the embodiments of the present invention; Figure 2 It is a bar chart for comparative analysis of the center distance measurement methods of the worm and worm gear transmission system according to the embodiments of the present invention; Figure 3 It is a schematic diagram for comparing the center distance measurement accuracy of the worm and worm gear transmission system according to the embodiments of the present invention; Figure 4 It is a flow chart of prediction and compensation strategy optimization of a bidirectional gated recurrent neural network enhanced by an attention mechanism according to the embodiments of the present invention; Figure 5 It is a bar chart for comparing the effects of multi-scale selection pressure indicators and evolutionary modulation factors according to the embodiments of the present invention. Detailed Embodiments

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0022] The technical solution of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0023] Figure 1 FIG. is a schematic flow chart of a dynamic automatic analysis method for the center distance of worm and worm gear meshing based on servo drive in an embodiment of the present invention. As Figure 1 shown, the method includes: Collect the state parameters of the worm and worm gear drive system to generate an initial measurement data set; use the initial measurement data set to train a neural network adaptive filter, which is based on a long short-term memory network architecture, establish a non-linear mapping relationship with the center distance signal, and perform real-time correction on the center distance measurement result by introducing a temperature compensation coefficient, and output the corrected center distance data; According to the corrected center distance data, combine the worm speed signal and the worm gear speed signal to establish a dynamic model of the worm and worm gear drive system, calculate the theoretical optimal center distance value, compare the theoretical optimal center distance value with the corrected center distance data, and obtain the center distance deviation sequence; Based on the center distance deviation sequence, use a bidirectional gated recurrent neural network enhanced by an attention mechanism to predict the center distance change trend by analyzing the temporal features in the historical compensation data, and continuously optimize the compensation strategy in combination with the online transfer learning method, and output the optimal compensation instruction; Convert the optimal compensation instruction into a displacement control signal of the servo motor, drive the servo motor to perform the center distance adjustment operation, and realize the dynamic compensation of the center distance of the worm and worm gear meshing.

[0024] In an alternative embodiment, using the initial measurement data set to train a neural network adaptive filter, which is based on a long short-term memory network architecture, establish a non-linear mapping relationship with the center distance signal, and perform real-time correction on the center distance measurement result by introducing a temperature compensation coefficient, and output the corrected center distance data includes: Collect the real-time operation data of the worm and worm gear drive system, generate a multi-modal measurement data set based on the real-time operation data; establish a high-fidelity mapping environment of the drive system based on the multi-modal measurement data set, construct a non-linear mapping relationship between the center distance signal and the multi-modal measurement data set in the high-fidelity mapping environment, and perform mapping processing on the initial center distance data through the non-linear mapping relationship to obtain the optimized center distance data; Calculate the temperature compensation coefficient according to the optimized center distance data in combination with the temperature data in the multi-modal measurement dataset. The temperature compensation coefficient is obtained through an exponential decay function of the initial compensation coefficient and the temperature change amount. Weightedly fuse the temperature compensation coefficient and the optimized center distance data, and introduce a dynamic bias term for compensation to generate a compensated center distance measurement value; Input the compensated center distance measurement value into the high-fidelity mapping environment, calculate the compensation parameter based on the historical compensation data in the high-fidelity mapping environment, and optimize and correct the compensated center distance measurement value through the compensation parameter to output the corrected center distance measurement data.

[0025] Collect the initial measurement dataset of the worm gear drive system. This dataset includes the original center distance signal, temperature signal, vibration signal, rotational speed signal, and load signal. When collecting, use a high-precision displacement sensor to measure the center distance change, with an accuracy of up to 0.001 mm; use a PT100 temperature sensor to monitor the system temperature, with a measurement range of -50°C to 150°C; use a three-axis acceleration sensor to collect vibration signals, with a sampling frequency set to 1 kHz; obtain the rotational speed signal through a Hall sensor, with a range covering 100 rpm to 3000 rpm; measure the load signal through a torque sensor, with a range of 0 - 100 N·m. All signals are synchronously collected through a data acquisition card, with a unified sampling rate of 10 kHz, forming a multi-modal initial measurement dataset.

[0026] Use the collected initial measurement dataset to train a neural network adaptive filter. This filter is designed based on the long short-term memory network architecture and includes an input layer, a hidden layer, and an output layer. The input layer receives multi-modal signal features, including the original center distance value, temperature value, vibration feature value, rotational speed value, and load value; the hidden layer consists of three layers of long short-term memory units, with each layer containing 128 neurons, and captures the long-term dependencies between signals through a gating mechanism; the output layer outputs the corrected center distance value. During the training process, use the mean square error as the loss function, optimize the network parameters through the stochastic gradient descent algorithm, set the learning rate to 0.001, the number of training epochs to 500, and the batch size to 64.

[0027] During the real-time operation phase, the system continuously collects the operation data of the worm and worm gear drive system to generate a multi-modal measurement dataset. In a typical example, the system collects 8 hours of continuous operation data, containing approximately 28,800 data points. Based on this data, the system constructs a high-fidelity mapping environment, which is a digital twin system that simulates the dynamic characteristics of the drive system under different working conditions. In this environment, a non-linear mapping relationship between the center distance signal and the multi-modal measurement data is established through a deep neural network. Specifically, a fully connected neural network structure is used, which includes 4 hidden layers, with 256, 128, 64, and 32 neurons in each layer respectively, and the ReLU activation function is adopted. This mapping relationship takes the original center distance data as input and outputs the optimized center distance data. The mapping process takes into account the influence of the system operation conditions and environmental factors on the center distance.

[0028] Based on the optimized center distance data, the system calculates the temperature compensation coefficient by combining the temperature data in the multi-modal measurement data. The temperature compensation coefficient is obtained through an exponential decay function of the initial compensation coefficient and the temperature change. In practical applications, the initial compensation coefficient is set to 1.0, and the temperature change is calculated by the difference between the current temperature and the standard temperature (25°C). When the temperature change is positive (temperature increase), the compensation coefficient shows a negative correlation change; when the temperature change is negative (temperature decrease), the compensation coefficient shows a positive correlation change. In a test scenario, when the system temperature rises from 25°C to 45°C, the temperature change is 20°C, and the calculated temperature compensation coefficient is 0.92; when the temperature drops from 25°C to 5°C, the temperature change is -20°C, and the calculated temperature compensation coefficient is 1.08.

[0029] The temperature compensation coefficient and the optimized center distance data are weighted and fused. The implementation method is to multiply the optimized center distance data by the temperature compensation coefficient. At the same time, a dynamic bias term is introduced for compensation, and this bias term is dynamically adjusted according to the system operation state. For example, during the system startup phase, the bias term is set to 0.005 mm; during the stable operation phase, the bias term is reduced to 0.002 mm; during the high-load state, the bias term is increased to 0.008 mm. Through this mechanism, the compensated center distance measurement value is generated.

[0030] Input the compensated center distance measurement value into the high-fidelity mapping environment, and the system calculates the compensation parameters based on the historical compensation data. The historical compensation data is stored in the system database and includes the compensation results and their corresponding working condition information within the past 24 hours. The sliding window technique is used for calculating the compensation parameters, with the window size set to 60 minutes and the step size to 10 minutes. Within each window, the system extracts the compensation trend features and calculates the average compensation value, standard deviation, and change rate. Based on these features, the system calculates the current compensation parameters through the weighted average method, with the weight distribution as follows: the weight of recent data (0 - 20 minutes) is 0.5, the weight of medium-term data (20 - 40 minutes) is 0.3, and the weight of long-term data (40 - 60 minutes) is 0.2. In practical applications, when the system detects that the compensation parameter exceeds the threshold (set to ±0.015 mm), the abnormal alarm mechanism is triggered to prompt the operator to check the system status.

[0031] Optimize and correct the compensated center distance measurement value through the compensation parameters, and output the finally corrected center distance measurement data. The correction process is to add the compensated center distance measurement value and the compensation parameter to obtain the final correction result. In actual tests, this method reduces the center distance measurement error from the original ±0.05 mm to ±0.008 mm, improving the measurement accuracy by approximately 84%, meeting the center distance control requirements of the high-precision worm and worm gear drive system.

[0032] Figure 2 This is the bar chart for the comparative analysis of the center distance measurement method of the worm and worm gear drive system in the embodiment of the present invention: This figure shows the performance comparison data of three different measurement methods (traditional measurement method, ordinary neural network filtering, and long short-term memory network filtering) under four different working conditions (normal temperature working condition, temperature fluctuation working condition, high load working condition, and composite working condition). Under the normal temperature working condition, the measurement accuracy of the long short-term memory network filtering reaches 87.6%, significantly better than 79.2% of the ordinary neural network filtering and 72.5% of the traditional measurement method. Under the temperature fluctuation working condition, the long short-term memory network performs best, reaching an accuracy of 82.3%, followed by the ordinary neural network at 74.5%, and the traditional method is the lowest at 63.8%. In the high load working condition, the performances of the three methods are: long short-term memory network 82.5%, ordinary neural network 78.9%, and traditional method 68.6%. Under the composite working condition, the long short-term memory network still maintains high performance, reaching 81.8%, while the ordinary neural network is 67.2%, and the traditional method drops to 58.4%. The data shows that under all working conditions, the long short-term memory network filtering method shows the best performance. Especially in the complex composite working condition, its advantage is more obvious, with an improvement of more than 20 percentage points compared to the traditional method, reflecting the strong adaptability and stability of this method in complex environments.

[0033] In an alternative embodiment, according to the corrected center distance data, combined with the worm gear rotation speed signal and the worm rotation speed signal, a dynamic model of the worm gear transmission system is established to calculate the theoretically optimal center distance value. Comparing the theoretically optimal center distance value with the corrected center distance data, the center distance deviation sequence obtained includes: Collect the dynamic data set of the worm gear transmission system, which is used to characterize the real-time operating state of the transmission system; Based on the dynamic data set, a spiking neural network model is constructed. The spiking neural network model simulates the characteristics of the transmission system through the sodium ion channel conductance and the potassium ion channel conductance. Based on the sodium ion channel conductance and the potassium ion channel conductance, the dynamic data set is received and the dynamic current is calculated. The dynamic current and the basic driving torque are weighted and combined to generate a synaptic modulation torque; Substitute the synaptic modulation torque into the enhanced dynamic equation, and obtain the dynamic response characteristics of the transmission system by solving the enhanced dynamic equation; Through the spiking neural network model, time-domain encoding is performed on the dynamic response characteristics to generate a pulse sequence signal. The pulse sequence signal describes the time-varying characteristics of the dynamic characteristics through a time exponential decay function, and a dynamic characteristic pulse encoding is obtained; Based on the dynamic characteristic pulse encoding, use the spiking neural network model to construct a time-varying kernel function. The time-varying kernel function is determined by a dynamically reconstructed synaptic structure. Through the time-varying kernel function, spatio-temporal domain convolution operation is performed on the dynamic characteristic pulse encoding to obtain the instantaneous state characteristics of the transmission system. Based on the instantaneous state characteristics, the theoretically optimal center distance is calculated; Compare the theoretically optimal center distance with the corrected center distance data to obtain the center distance deviation data.

[0034] Collect the dynamic data set of the worm gear transmission system, which is used to characterize the real-time operating state of the transmission system. The collected data includes the worm gear rotation speed signal ω1, the worm rotation speed signal ω2, and the input torque signal T in 、output torque signal T out, the bearing housing temperature signal θ1, the lubricating oil temperature signal θ2, and the vibration acceleration signal a. The rotational speed signal is collected by incremental encoders installed on the worm wheel shaft and the worm shaft, with a sampling frequency of 1000 Hz; the torque signal is collected by torque sensors installed on the input shaft and the output shaft, with a measuring range of 0 - 200 N·m and an accuracy of 0.5% F.S.; the temperature signal is collected by temperature sensors installed on the bearing housing and the lubricating oil tank, with a measuring range of -20°C to 120°C and an accuracy of ±0.5°C; the vibration signal is collected by an acceleration sensor installed on the bearing housing, with a frequency range of 1 Hz - 10 kHz and a sensitivity of 100 mV / g. All signals are converted into digital signals by a 16-bit AD converter and high-frequency noise is filtered out by a low-pass filter (cutoff frequency 400 Hz).

[0035] Based on the collected kinetic data set, a pulse neural network model is constructed. This model uses the Hodgkin-Huxley neuron model to simulate the characteristics of the transmission system through the sodium ion channel conductance and the potassium ion channel conductance. The membrane potential kinetic equation is expressed as the membrane capacitance multiplied by the derivative of the membrane potential with respect to time equals the total ionic current plus the external input current. Among them, the membrane capacitance C m takes a value of 1 microfarad per square centimeter; the total ionic current includes the sodium ion current I Na , the potassium ion current I K and the leakage current I L ; the external input current I ext is determined by the kinetic data.

[0036] The sodium ion current I Na equals the maximum sodium ion channel conductance g Na multiplied by the cube of the sodium ion channel activation variable m, multiplied by the sodium ion channel inactivation variable h, and multiplied by the difference between the membrane potential V and the sodium ion equilibrium potential E Na . Among them, the maximum sodium ion channel conductance g Na is 120 millisiemens per square centimeter, and the sodium ion equilibrium potential E Na is 50 millivolts. The kinetic equation of the sodium ion channel activation variable m is that the derivative of m with respect to time equals (1 - m) multiplied by α m minus m multiplied by β m , where α m and β m are rate constants related to the membrane potential. The kinetic equation of the sodium ion channel inactivation variable h is similar, that is, the derivative of h with respect to time equals (1 - h) multiplied by α h minus h multiplied by β h .

[0037] The potassium ion current I K equals the maximum potassium ion channel conductance g KMultiplied by the fourth power of the potassium ion channel activation variable n and the difference between the membrane potential V and the potassium ion equilibrium potential E K where the maximum potassium ion channel conductance g K is 36 millisiemens per square centimeter and the potassium ion equilibrium potential E K is -77 millivolts. The kinetic equation of the potassium ion channel activation variable n is that the derivative of n with respect to time is equal to (1 - n) multiplied by α n minus n multiplied by β n where α n and β n are rate constants related to the membrane potential.

[0038] The leakage current I L is equal to the leakage current conductance g L multiplied by the difference between the membrane potential V and the leakage current equilibrium potential E L where the leakage current conductance g L is 0.3 millisiemens per square centimeter and the leakage current equilibrium potential E L is -54.4 millivolts.

[0039] The external input current I ext is linearly mapped from the kinetic data, and the mapping relationship is that I ext is equal to the rotational speed signal mapping coefficient k ω multiplied by the worm wheel rotational speed signal ω1 plus k' ω multiplied by the worm rotational speed signal ω2 plus the torque signal mapping coefficient k T multiplied by the input torque signal T in plus k' T multiplied by the output torque signal T out plus the temperature signal mapping coefficient k θ multiplied by the bearing housing temperature signal θ1 plus k' θ multiplied by the lubricating oil temperature signal θ2 plus the vibration acceleration signal mapping coefficient k a multiplied by the vibration acceleration signal a. Among them, each mapping coefficient is obtained through training and optimization, and the typical values are k ω = 0.05 μA·min / r, k' ω = 0.03 μA·min / r, k T = 0.08 μA / N·m, k' T = 0.12 μA / N·m, k θ = 0.02 μA / ℃, k' θ = 0.015 μA / ℃, k a = 0.01 μA·s 2 / m.

[0040] After the kinetic data is input into the neural network, the kinetic current I is calculated based on the sodium ion channel conductance and the potassium ion channel conductancedyn The formula for the kinetic current is I dyn equals g Na multiplied by the cube of m, multiplied by h, multiplied by (V - E Na ) plus g K multiplied by the fourth power of n, multiplied by (V - E K ) plus g L multiplied by (V - E L ). Under normal operating conditions, the typical values of the kinetic current are a peak sodium ion current of approximately -12 μA / cm², a peak potassium ion current of approximately 7 μA / cm², and a leakage current of approximately 0.5 μA / cm².

[0041] The calculated kinetic current is combined with the basic driving torque through weighted combination to generate the synaptic modulation torque T syn The formula for the synaptic modulation torque is T syn equals the basic driving torque T base plus the weighting coefficient λ multiplied by the kinetic current I dyn The basic driving torque T base is determined according to the design parameters and operating conditions of the transmission system. For a worm gear reducer with a power of 5.5 kW, T base takes a value of 35 N·m. The weighting coefficient λ is optimized through a genetic algorithm. For different operating conditions, the value range of λ is from 0.15 to 0.85 N·m·cm² / μA, and the typical value is 0.35 N·m·cm² / μA. Under typical operating conditions, the variation range of the synaptic modulation torque is ±15% of the basic driving torque, that is, from 29.75 to 40.25 N·m.

[0042] The synaptic modulation torque is substituted into the enhanced kinetic equation, and the dynamic response characteristics of the transmission system are obtained by solving this equation. The enhanced kinetic equation is expressed as the mass matrix M multiplied by the second derivative of the generalized coordinate q with respect to time, plus the damping matrix C multiplied by the first derivative of the generalized coordinate q with respect to time, plus the stiffness matrix K multiplied by the generalized coordinate q equals the external excitation vector F. Among them, the generalized coordinate q is a four-dimensional vector, including the rotational angular displacement θ1 of the worm, the rotational angular displacement θ2 of the worm gear, the axial displacement x1 of the worm, and the radial displacement x2 of the worm gear.

[0043] The mass matrix M is a diagonal matrix, and the diagonal elements are the moment of inertia J1 of the worm, the moment of inertia J2 of the worm gear, the mass m1 of the worm, and the mass m2 of the worm gear respectively. For a worm gear reducer with a power of 5.5 kW, J1 takes a value of 0.012 kg·m², J2 takes a value of 0.185 kg·m², m1 takes a value of 1.8 kg, and m2 takes a value of 4.5 kg.

[0044] The damping matrix C includes bearing friction damping and meshing damping, and is expressed as a fourth-order square matrix. The main diagonal elements are the worm bearing damping coefficient c1, the worm gear bearing damping coefficient c2, the worm axial damping coefficient c3, and the worm gear radial damping coefficient c4, with values of 0.25 N·m·s / radian, 0.38 N·m·s / radian, 3.5 N·s / m, and 3.8 N·s / m respectively. The off-diagonal elements represent coupling damping, and generally take values from 10% to 30% of the diagonal elements.

[0045] The stiffness matrix K includes shaft stiffness and meshing stiffness, and is expressed as a fourth-order square matrix. The main diagonal elements are the worm shaft torsional stiffness k1, the worm gear shaft torsional stiffness k2, the worm axial stiffness k3, and the worm gear radial stiffness k4, with values of 1.2×10 4 N·m / radian, 1.8×10 4 N·m / radian, 8.5×10 4 N / m, 9.2×10 4 N / m. The off-diagonal elements represent coupling stiffness, which is determined by the gear meshing relationship. In particular, k 12 (worm-worm gear coupling stiffness) takes a value of 2.3×10 5 N / m.

[0046] The external excitation vector F is a four-dimensional vector, including the worm input torque T in , the worm gear output torque T out , the worm axial force F ax , and the worm gear radial force F rad . Among them, T in is equal to the synaptic modulation torque T syn , T out is the load torque, and F ax and F rad are determined by the meshing force and are related to the input torque, meshing angle, friction coefficient, etc.

[0047] The solution of the enhanced dynamic equation uses the fourth-order Runge-Kutta method, with a time step of 0.001 seconds and a total calculation time of 10 seconds. The solution process includes the following steps: (1) Initialize the generalized coordinates q and the generalized velocity v; (2) Calculate the system matrices M, C, K, and the external excitation vector F at the current moment; (3) Convert the second-order differential equation into a first-order differential equation system, that is, the derivative of q with respect to time is equal to v, and the derivative of v with respect to time is equal to the inverse matrix of M multiplied by (F minus C multiplied by v minus K multiplied by q); (4) Use the fourth-order Runge-Kutta method to solve the first-order differential equation system to obtain the generalized coordinates q and the generalized velocity v at the next moment; (5) Repeat steps (2)-(4) until the calculation of the entire time domain is completed.

[0048] The dynamic response characteristics are encoded in the time domain through a spiking neural network model to generate a spiking sequence signal. The method of time-domain encoding is as follows: when the membrane potential V of a neuron exceeds the threshold V th (with a value of -55 mV), the neuron emits a spike and then enters the refractory period, during which the membrane potential is reset to the resting potential V rest (with a value of -70 mV); when the membrane potential is below the threshold, the neuron does not emit a spike. The spike emission time is recorded as the timestamp t spike , with a precision of milliseconds.

[0049] The time exponential decay function E(t) is introduced to describe the time-varying characteristics of the dynamic characteristics. The expression is that E(t) is equal to exp[-(t - t spike ) / τ], where t is the current time, t spike is the time of the most recent spike emission, and τ is the time constant with a value of 25 milliseconds. In this way, spikes that occurred recently have higher weights, while the weights of spikes that occurred earlier gradually decay, effectively capturing the time-varying nature of the dynamic characteristics.

[0050] For a worm gear drive system, the characteristics of the spiking sequence signal are as follows: under normal operating conditions (input speed 1440 r / min, load torque 35 N·m), the average spike emission rate is 85 times per second, the mean of the spike intervals is 11.8 milliseconds, and the standard deviation is 3.2 milliseconds; under light load conditions (input speed 1440 r / min, load torque 17.5 N·m), the average spike emission rate decreases to 65 times per second, and the mean of the spike intervals increases to 15.4 milliseconds; under heavy load conditions (input speed 1440 r / min, load torque 52.5 N·m), the average spike emission rate increases to 110 times per second, and the mean of the spike intervals decreases to 9.1 milliseconds.

[0051] Based on the pulse coding of dynamic characteristics, a time-varying kernel function K(t, t') is constructed using a spiking neural network model. The time-varying kernel function is determined by the dynamically reconfigured synaptic structure, where the synaptic weight w ij is adjusted according to the spike timing characteristics. The adjustment of the synaptic weight adopts the spike-timing-dependent plasticity rule, and the expression is that Δw ij is equal to A + multiplied by exp[-(t j - t i ) / τ + (when t j > t i ) or A - multiplied by exp[(t j - t i ) / τ - (when t j < t i ). Among them, ti is the pulse time of the presynaptic neuron, t j is the pulse time of the postsynaptic neuron, A + is the long-term potentiation coefficient (with a value of 0.1), A - is the long-term depression coefficient (with a value of -0.12), τ + is the long-term potentiation time constant (with a value of 20 milliseconds), τ - is the long-term depression time constant (with a value of 60 milliseconds).

[0052] The expression of the time-varying kernel function K(t, t') is that K(t, t') equals the weighted sum of all synaptic weights w ij with the weights being the activity states of the corresponding neurons at times t and t'. For the worm gear drive system, the typical characteristics of the time-varying kernel function are: the maximum value of the kernel function appears at t = t', with a value of approximately 0.85; as |t - t'| increases, the kernel function value decays rapidly, and at |t - t'| = 50 milliseconds, it decays to about 10% of the maximum value.

[0053] Through the spatio-temporal domain convolution operation of the time-varying kernel function on the dynamic characteristic pulse coding, the instantaneous state characteristic s(t) of the drive system is obtained. The expression of the spatio-temporal domain convolution operation is that s(t) equals the integral of K(t, t') multiplied by the pulse coding function Spike(t') from negative infinity to t with respect to t'. In actual calculations, the integral is discretized into a summation form, with a time step of 1 millisecond and an integral window length of 500 milliseconds. The instantaneous state characteristic s(t) is a 12-dimensional vector, including characteristics such as the speed ratio, torque fluctuation, and temperature change of the system.

[0054] Calculate the theoretical optimal center distance a opt . The calculation process uses a multi-parameter optimization method, and the optimization objective function is J = w1·ε 2 + w2·(1 - η) 2 , where ε is the transmission error, η is the transmission efficiency, and w1 and w2 are weight coefficients (with values of 0.6 and 0.4 respectively). The calculation formula for the transmission error ε is ε = (i - i0) / i0, where i is the actual transmission ratio and i0 is the theoretical transmission ratio (equal to the ratio of the number of worm spiral coils to the number of worm gear teeth). The calculation formula for the transmission efficiency η is η = T out ·ω2 / (T in ·ω1), where T out is the output torque, ω2 is the angular velocity of the worm gear, T in is the input torque, and ω1 is the angular velocity of the worm.

[0055] The relationship between the theoretical optimal center distance a opt and the instantaneous state characteristic s(t) is a opt=a0·[1 + Σ(γ i ·s i (t))], where a0 is the nominal center distance (assumed to be 100 mm), and γ i is the optimization coefficient, which is calculated by the gradient descent method. During the optimization process, the constraint condition is that the change range of the theoretical optimal center distance does not exceed ±3% of the nominal center distance, that is, 97 to 103 mm. Under typical working conditions, the numerical range of the optimization coefficient γ i is ±0.005, and the change range of the theoretical optimal center distance is 99.2 to 101.5 mm.

[0056] Compare the theoretical optimal center distance a opt with the corrected center distance data a cor to obtain the center distance deviation data Δa. The deviation calculation formula is Δa = a opt - a cor . The center distance deviation data is used to evaluate the operating state of the current transmission system. Excessive deviation indicates that there may be problems with the transmission system and adjustments or maintenance are required. For a normally operating worm and worm gear transmission system, the mean value of the center distance deviation should be close to zero, and the standard deviation should not exceed 0.5 mm; if the mean value of the deviation exceeds ±1 mm or the standard deviation exceeds 1 mm, it indicates that there may be wear, improper installation, or other faults in the transmission system.

[0057] In the actual application of a certain type of worm and worm gear reducer (model ZK40, rated power 5.5 kW, transmission ratio i0 = 40), after adjusting the center distance using this method, the transmission efficiency is increased from the original 82.5% to 87.3%, the transmission error is reduced from the original ±1.2% to ±0.5%, the noise level is reduced from the original 75 dB to 68 dB, and the service life is extended from the original 12,000 hours to 18,000 hours, significantly improving the reliability and economy of the equipment.

[0058] Figure 3 Schematic diagram for comparing the center distance measurement accuracy of the worm and worm gear transmission system according to the embodiment of the present invention: The figure shows the performance change trends of three different technical solutions (this technical solution, traditional kinetic model, and basic neural network model) within 100 minutes of running time. This technical solution demonstrates optimal performance and stability, with its accuracy always remaining above 94%, steadily increasing from the initial 95.2% to 97.5% at 100 minutes, showing an overall slow upward trend. The basic neural network model ranks second, with an initial accuracy of 89.5%. Although there are minor fluctuations during the running process, it generally shows an upward trend and finally reaches 92.3% at 100 minutes. The performance of the traditional kinetic model is relatively the weakest, with an initial accuracy of 85.4%. It has large fluctuations during the running process. Although it also shows a weak upward trend, it only reaches 86.8% at 100 minutes. It can be seen from the trend graph that this technical solution not only has the highest accuracy but also shows stability and reliability. Especially under long-term running conditions, its performance advantage is more obvious, and the performance gap with the other two solutions gradually expands, fully demonstrating the superiority of this solution in practical applications.

[0059] In an alternative embodiment, substitute the synaptic modulation torque into the enhanced kinetic equation, and obtain the dynamic response characteristics of the transmission system by solving the enhanced kinetic equation; perform time-domain encoding on the dynamic response characteristics through the spiking neural network model to generate a pulse sequence signal including: Construct a field theory enhanced kinetic equation based on the synaptic modulation torque. The field theory enhanced kinetic equation introduces a field-matter coupling force term, and obtain the field state distribution of the transmission system by solving the field theory enhanced kinetic equation. The field state distribution reflects the overall dynamic state of the transmission system; Substitute the field state distribution into the field theory enhanced kinetic equation, dynamically update the coupling coefficient in the field-matter coupling force term according to the real-time operating state of the transmission system, and adjust the parameters of the field theory enhanced kinetic equation in combination with the change law of the synaptic modulation torque. Solve the field theory enhanced kinetic equation again to obtain the coupled dynamic response of the transmission system. The coupled dynamic response characterizes the nonlinear dynamic behavior of the transmission system; Apply a field state projection transformation to the coupled dynamic response, extract the characteristics of the system speed component, acceleration component, and field state component, and generate a feature vector characterized by the field state through a linear combination of feature mode functions. The feature vector contains the multi-scale dynamic information of the transmission system; Input the feature vector characterized by the field state into the spiking neural network, perform time-domain encoding on the dynamic characteristics through the spiking neural network, and introduce a timestamp function and an exponential decay function to modulate the encoding process to generate a pulse sequence signal with time-varying characteristics.

[0060] The synaptic modulation torque refers to the dynamic torque generated by modulating the synaptic structure of a spiking neural network model. For a gear transmission system, the calculation method of the synaptic modulation torque is as follows: Multiply the synaptic weight matrix in the spiking neural network by the neuron activation state vector, and then map it to the torque domain through a transfer function. Specifically, if a network composed of 16 neurons is used, the dimension of the synaptic weight matrix W is 16×16, the dimension of the neuron activation state vector A is 16×1, and the synaptic modulation torque T syn is calculated as T syn =k·(W·A)+T base , where k is the mapping coefficient with a value of 0.15, and T base is the basic driving torque with a value of 25 N·m. The membrane potential threshold of the neuron is set to -55 mV, and the resting potential is -70 mV. Through this calculation method, the synaptic modulation torque can achieve a dynamic modulation range of ±20% based on the basic driving torque, that is, 20~30 N·m.

[0061] When constructing the field theory enhanced dynamics equation based on the synaptic modulation torque, a field-matter coupling force term is introduced. The traditional dynamics equation mainly considers rigid body motion, elastic deformation, and damping effects, and its expression form is , where M is the mass matrix, C is the damping matrix, K is the stiffness matrix, and F(t) is the external excitation force. The field theory enhanced dynamics equation adds a field-matter coupling term on this basis, and its expression form is , where α is the coupling coefficient, is the field potential function, represents the field potential gradient. The field potential function satisfies the field equation , where D is the field diffusion coefficient with a value of 0.25, S(x,t) is the source term, which is linearly related to the synaptic modulation torque T syn , and the relationship is S(x,t)=β·T syn , where β is the mapping coefficient with a value of 0.04.

[0062] The improved finite difference time domain method is used to solve the field theory enhanced dynamics equation. The specific steps are as follows: First, the computational domain is discretized into a 200×200 uniform grid with a grid spacing of Δx = Δy = 0.01 m; then the time domain is discretized into several time steps with a time step size of Δt = 0.001 s; then the central difference scheme is used to discretize the spatial partial derivative term, ; the forward difference scheme is used to discretize the time partial derivative term, ; finally, the iterative format , is obtained. The boundary conditions are set as periodic boundaries, that is, , N . The initial conditions are set as a uniform field distribution , .

[0063] During the solution process, to improve the computational efficiency, GPU parallel computing technology is adopted. The implementation method of parallel computing is as follows: a 200×200 grid is divided into 10×10 computing blocks, and each block contains 20×20 grid points; each GPU thread is responsible for calculating one grid point; shared memory is used to store the data within the block to reduce the number of global memory accesses; the double-buffering technology is adopted to avoid waiting caused by data dependencies. Using an NVIDIA RTX 3080 graphics card, the calculation time for a single iteration is about 0.25 milliseconds, and the total calculation time for a 10-second simulation time (10,000 steps) is about 2.5 seconds.

[0064] The field state distribution of the transmission system is obtained by solving the field theory enhanced dynamics equation. The field state distribution is represented in the form of a two-dimensional scalar field φ(x,y), and the numerical range is between [0,1]. In the gear transmission system, the field state distribution reflects the energy / stress distribution state of the system, and the high-value region corresponds to the high-stress or high-energy region. The typical characteristics of the field state distribution are as follows: the field state value in the gear meshing region is high, about 0.85; the field state value in the gear body region is medium, about 0.6; the field state value in the region far from the gear is low, about 0.1 - 0.3. The field state distribution is stored in matrix form, and the matrix size is 200×200, corresponding to the discrete calculation grid.

[0065] Substitute the field state distribution into the field theory enhanced dynamics equation, and dynamically update the coupling coefficient α in the field-matter coupling force term according to the real-time operating state of the transmission system. The update strategy of the coupling coefficient is based on the gradient information of the field state distribution, and the update formula is , where α min =0.2, α max =0.8, γ = 0.15 is the gradient threshold, is the modulus of the field state gradient, and the calculation method is . When the field state gradient is large, the coupling coefficient approaches α max , enhancing the coupling effect between the field and matter; when the field state gradient is small, the coupling coefficient approaches α min , weakening the coupling effect to reduce the computational complexity.

[0066] Adjust the parameters of the field theory enhanced dynamics equation in combination with the change law of the synaptic modulation torque. The relationship between the source term intensity S(x,t) and the synaptic modulation torque T syn is , where S base is the basic source term intensity, with a value of 1.0, and δ is the modulation coefficient, with a value of 0.8. When the synaptic modulation torque increases, the source term intensity increases accordingly; when the synaptic modulation torque decreases, the source term intensity decreases accordingly. This parameter adjustment mechanism enables the field theory enhanced dynamics equation to adapt to the dynamic process of the transmission system load change.

[0067] Solve the field theory enhanced dynamics equation again to obtain the coupled dynamics response of the transmission system. The coupled dynamics response includes the displacement vector x(t), velocity vector , acceleration vector , and field state distribution , a total of four parts. The displacement vector is calculated using the improved Newmark-β method. The specific steps are as follows: predict the displacement , predict the velocity , calculate the acceleration , correct the displacement , correct the velocity . Among them, β = 0.25 and γ = 0.5 are integration parameters. Under standard working conditions, the amplitude range of the displacement response is ±0.15 mm, the amplitude range of the velocity response is ±0.5 m / s, the amplitude range of the acceleration response is ±15 m / s 2 , and the amplitude range of the field state response is ±0.25.

[0068] Apply the field state projection transformation to the coupled dynamics response to extract the characteristics of the system velocity component, acceleration component, and field state component. The field state projection transformation uses the singular value decomposition technique to decompose the original field state distribution matrix into , where U and V are orthogonal matrices, Σ is a diagonal matrix, and the diagonal elements are singular values σ i . Sort the singular values from largest to smallest, and select the first 12 largest singular values σ1, σ2... σ 12 and their corresponding left singular vectors u1, u2... u 12 and right singular vectors v1, v2... v 12 . These vectors form the characteristic mode functions.

[0069] Generate the characteristic vector of the field state representation through the linear combination of the characteristic mode functions. The calculation formula of the characteristic vector f is f = [f1, f2... f 12 T , where represents the matrix inner product operation. The linear combination coefficients are determined by the least squares method to minimize the reconstruction error. For the gear transmission system, under normal working conditions, the typical values of the characteristic vector are [0.85, 0.63, 0.42, 0.38, 0.25, 0.22, 0.18, 0.15, 0.12, 0.08, 0.05, 0.03]. The first component represents the contribution of the main mode and has the largest value; the subsequent components decrease in turn, representing the contributions of the secondary modes.

[0070] ​Input the eigenvector of the field state characterization into the spiking neural network, and perform time-domain encoding of the dynamic characteristics through the spiking neural network. The spiking neural network adopts the Hodgkin-Huxley neuron model, and the dynamic equation of the membrane potential V is , where C m is the membrane capacitance, with a value of ; g Na , g K , g L are the conductances of the sodium ion channel, potassium ion channel, and leakage current channel respectively, with values of 120 mS / cm 2 , 36 mS / cm 2 , 0.3 mS / cm 2 ; E Na , E K , E L are the corresponding equilibrium potentials, with values of 50 mV, -77 mV, -54.4 mV respectively; m, h, n are gating variables, satisfying and other differential equations; I ext is the external input current, which is linearly related to the eigenvector f, and the relationship is I ext = ρ·f, where ρ is the mapping coefficient vector.

[0071] Introduce the timestamp function and exponential decay function to modulate the encoding process. The timestamp function T(t) records the time since the start of the simulation, with a precision of 0.1 ms; the exponential decay function , where τ is the time constant, with a value of 25 ms, and t0 is the reference time. The modulated pulse intensity S(t) is calculated as , where S0 is the base pulse intensity, and t spike is the time of the most recent pulse emission. This time-varying modulation mechanism can effectively capture the temporal variation law of the dynamic characteristics of the drive system.

[0072] The simulation of the spiking neural network uses the fourth-order Runge-Kutta method with an adaptive time step. The initial time step is set to 0.01 ms, the maximum time step is limited to 0.1 ms, and the minimum time step is limited to 0.001 ms. The adaptive adjustment of the time step is based on error estimation. If the estimated error exceeds the tolerance error (set to 10 -6 ), the time step is reduced; if the estimated error is much smaller than the tolerance error, the time step is increased. The pulse emission determination condition is that the membrane potential V exceeds the threshold , and after emitting a pulse, the neuron enters the refractory period, with a duration of 2 ms.

[0073] A pulse sequence signal with time-varying characteristics is generated by the above method. In practical applications, for a gear transmission system, the average firing rate of the pulse sequence is 85 Hz, the mean of the pulse interval is 11.8 ms, and the standard deviation is 3.2 ms. The pulse sequence is represented by 0-1 binary coding, where "1" indicates that the neuron fires a pulse at that moment, and "0" indicates no pulse. Specifically, the time axis is discretized into time slots of 1 ms. If a pulse is fired in a certain time slot, the corresponding position is marked as "1", otherwise it is marked as "0", thus forming a binary pulse sequence coding. This coding method can effectively compress the data volume while retaining the key information of the dynamic characteristics.

[0074] In the fault diagnosis application of a certain type of gearbox, the pulse sequence signal obtained by using this method is used to train a fault classifier. Compared with the traditional spectrum analysis method, the diagnostic accuracy rate is increased from 86.5% to 94.3%, and the early warning time is extended from an average of 18 hours to 32 hours, significantly improving the reliability and service life of the equipment. In particular, the recognition ability for minor faults and compound faults is greatly improved, reducing the missed alarm rate and false alarm rate.

[0075] In an alternative implementation, based on the center distance deviation sequence, a bidirectional gated recurrent neural network enhanced by an attention mechanism predicts the center distance change trend by analyzing the temporal features in the historical compensation data, and continuously optimizes the compensation strategy in combination with the online transfer learning method. The output optimal compensation instructions include: Construct a bidirectional gated recurrent neural network, input the forward hidden layer state and the backward hidden layer state of the bidirectional gated recurrent neural network into the swarm intelligence emergence layer. The swarm intelligence emergence layer introduces a swarm cooperation factor to enhance the hidden layer state and generate enhanced hidden layer features; Construct a self-evolving attention weight based on the enhanced hidden layer features. The self-evolving attention weight calculates the fitness function through the ratio of the prediction error to the reference error threshold, and multiplies the fitness function by the attention base weight to obtain the adaptive attention weight; Establish an environmental pressure selection mechanism according to the adaptive attention weight. The environmental pressure selection mechanism calculates the transfer probability based on the adaptation distance from the source domain to the target domain, uses the transfer probability as the weight coefficient of the source domain loss function, and updates the transfer learning model parameters through gradient descent; Modulate the enhanced hidden layer features based on the transfer learning model parameters to generate a dynamic feature vector, input the dynamic feature vector and the adaptive attention weight into the compensation strategy generation module. The compensation strategy generation module introduces an evolutionary modulation factor, and the evolutionary modulation factor is calculated through the cumulative effect of the multi-scale selection pressure index; Multiply the evolutionary modulation factor by the initial compensation instruction to obtain an optimal compensation instruction, and feedback the compensation effect corresponding to the optimal compensation instruction to the fitness function to update the self-evolving attention weight, thereby realizing the dynamic optimization of the compensation strategy.

[0076] As Figure 4 shown, the method further includes: The input of the bidirectional gated recurrent neural network is the center distance deviation sequence, which contains 120 groups of center distance deviation data of the worm and worm gear transmission system collected continuously. Each group of data includes a timestamp, an actual center distance value, a theoretical optimal center distance value, and a temperature value. The bidirectional gated recurrent neural network is set with 64 hidden layer neurons, uses the tanh activation function, has 200 training epochs, a learning rate of 0.001, and a batch size of 32.

[0077] When constructing the bidirectional gated recurrent neural network, the forward hidden layer state and the backward hidden layer state are input into the swarm intelligence emergence layer after being processed. The implementation method of the swarm intelligence emergence layer is as follows: The swarm size is set to 10, and each agent is represented as a sub-network with independent parameters. The introduced swarm cooperation factor is set to 0.75, which is used to measure the information interaction degree between agents. During the hidden layer state enhancement process, each agent receives the hidden layer state information of adjacent agents and performs weighted fusion with its own hidden layer state, and the weight coefficient is the swarm cooperation factor. After 5 iterations of calculation, enhanced hidden layer features are generated, and the feature dimension is 128.

[0078] The implementation method of constructing the self-evolving attention weight based on the enhanced hidden layer features is as follows: The prediction error is calculated using the root mean square error, and the reference error threshold is set to 0.05 mm. The ratio of the two is used as the input of the fitness function. The fitness function adopts an exponential decay form. When the ratio is less than 1, the fitness value is close to 1; when the ratio is greater than 1, the fitness value decays exponentially as the ratio increases.

[0079] The attention base weight is calculated through the softmax function, and its dimension is the same as the sequence length, and the initial value is set to a uniform distribution. Multiply the output value of the fitness function by the attention base weight to obtain the adaptive attention weight. In practical applications, when the center distance deviation prediction error is 0.03 mm and the reference error threshold is 0.05 mm, the ratio is 0.6, and the calculated fitness function value is 0.92. Multiply this value by the base weight to obtain the adaptive attention weight.

[0080] The implementation method of the environmental pressure selection mechanism is as follows: The source domain data selects the data of the worm and worm gear drive system collected in the laboratory, including 5000 groups of samples; the target domain data selects the data collected under actual working conditions, including 1000 groups of samples. The adaptation distance is calculated based on the distance between the source domain and target domain data distributions. The statistical characteristic difference measurement method is adopted. When the data distributions of the source domain and target domain are close, the adaptation distance is small; otherwise, it is large. The migration probability is obtained by mapping the adaptation distance through an exponential function. When the adaptation distance is 0.3, the calculated migration probability is 0.74. The migration probability is used as the weight coefficient of the source domain loss function, and the parameters of the transfer learning model are updated by the gradient descent method. The learning rate is set to 0.0005, and the number of iterations is 150 times.

[0081] The implementation method of enhancing the hidden layer features based on the modulation of the transfer learning model parameters is as follows: The parameters output by the transfer learning model are converted into a modulation coefficient matrix, and the dimension is the same as that of the hidden layer features. The value range of the modulation coefficient is from 0.8 to 1.2. The modulation process adopts element-wise multiplication operation, and the corresponding elements of the modulation coefficient matrix and the hidden layer features are multiplied to generate a dynamic feature vector. In practical applications, when the value of the hidden layer feature element is 0.65 and the corresponding modulation coefficient is 1.15, the value of the modulated feature element is 0.7475.

[0082] The dynamic feature vector and the adaptive attention weight are input into the compensation strategy generation module, and an evolutionary modulation factor is introduced. The calculation of the evolutionary modulation factor is based on the cumulative effect of the multi-scale selection pressure index, including the selection pressure indexes at three scales: short-term (10 time steps), medium-term (30 time steps), and long-term (50 time steps). The selection pressure index at each scale is calculated by the center distance compensation effect score, and the score range is from 0 to 1. Weights of 0.2, 0.3, and 0.5 are respectively assigned for weighted summation to obtain the evolutionary modulation factor. In practical applications, when the short-term, medium-term, and long-term selection pressure indexes are 0.85, 0.78, and 0.92 respectively, the calculated evolutionary modulation factor is 0.861.

[0083] The initial compensation instruction is generated by the basic compensation network, which is a three-layer fully connected neural network. The input is the dynamic feature vector, and the output is the compensation amount. The number of hidden layer nodes is 32, and the ReLU activation function is adopted. The evolutionary modulation factor is multiplied by the initial compensation instruction to obtain the optimal compensation instruction. When the initial compensation instruction is 0.08 mm and the evolutionary modulation factor is 0.861, the optimal compensation instruction is 0.06888 mm.

[0084] The compensation effect corresponding to the optimal compensation instruction is fed back to the fitness function, the error between the actual center distance and the target center distance is calculated, and the self-evolving attention weight is updated. Through this feedback mechanism, the dynamic optimization of the compensation strategy is realized. After 10 iterations of optimization, the center distance control accuracy is improved from the initial ±0.05 mm to ±0.02 mm, meeting the requirements of the high-precision transmission system.

[0085] In the application of a certain worm and worm gear reducer production line, this method is used to conduct dynamic compensation tests on 100 prototypes. Before compensation, the average deviation of the center distance is 0.062 mm, and after compensation, the average deviation drops to 0.018 mm, with an improvement rate of 70.97%. At the same time, the transmission noise is reduced by an average of 5.2 dB, the transmission efficiency is increased by 2.8%, and the service life is expected to be extended by more than 25%.

[0086] Through the above implementation methods, high-precision dynamic compensation of the center distance of the worm and worm gear meshing is achieved, effectively improving the performance and reliability of the transmission system. This method can adaptively adjust the compensation strategy according to changes in the operating environment, and has strong versatility and robustness.

[0087] In an alternative embodiment, the dynamic feature vector and the adaptive attention weight are input into the compensation strategy generation module, and the compensation strategy generation module introduces an evolutionary modulation factor. The evolutionary modulation factor is calculated through the cumulative effect of the multi-scale selection pressure index, including: The dynamic feature vector and the adaptive attention weight are combined to generate multi-level features, and the multi-level features are input into the compensation strategy generation module. The compensation strategy generation module introduces a dynamic response enhancement function to enhance the multi-level features. The dynamic response enhancement function is adaptively adjusted based on the feature energy level difference to generate an initial compensation strategy, and the initial compensation strategy reflects the dynamic response characteristics of the system; Based on the initial compensation strategy, a multi-scale selection pressure index is constructed. The multi-scale selection pressure index is calculated through the normalized distance between the initial compensation strategy and the reference strategy, and the reference strategy is set with corresponding reference values at different scales; Stability protection is applied to the multi-scale selection pressure index, and the multi-scale selection pressure index is structurally enhanced through a dynamic modulation field. The dynamic modulation field is dynamically adjusted based on the real-time changes of the system state, and continuous stability protection is provided through the closed-path integral of the dynamic modulation field to generate a steady-state enhanced selection pressure index, and the steady-state enhanced selection pressure index has anti-interference ability; Substitute the steady-state enhanced selection pressure index into the evolutionary modulation factor calculation formula, and perform weighted calculation on the steady-state enhanced selection pressure index through different scale weight coefficients. The weight coefficients are adaptively adjusted according to the scale change, and the cooperative optimization of multi-scale features is achieved through the dynamic allocation of the different scale weight coefficients, so as to obtain the evolutionary modulation factor.

[0088] Obtain the dynamic feature vector and the adaptive attention weight. These parameters contain the real-time description of the system operation state. The dynamic feature vector is usually obtained by preprocessing and feature extraction of the signals collected by multiple sensors. For example, when the system detects that the environmental parameters are temperature 25°C, humidity 65%, and atmospheric pressure 101 kPa, a feature vector [0.325, 0.421... 0.651] with a dimension of 128 can be extracted and formed. The adaptive attention weight reflects the importance of each feature and is generated by evaluating the correlation between the feature and the system performance. For example, the weight vector [0.08, 0.12... 0.05] can be obtained.

[0089] The system performs a fusion operation on the dynamic feature vector and the adaptive attention weight to generate multi-level features. The specific implementation method is through weighted combination, multiplying each element of the feature vector by the corresponding attention weight to obtain the first-level fusion feature. Subsequently, the system uses a non-linear mapping function to transform the first-level fusion feature to construct a three-layer feature hierarchy. The first layer retains the original information, the second layer captures patterns with medium complexity, and the third layer extracts high-level abstract features. For example, for the input feature [0.325×0.08, 0.421×0.12... 0.651×0.05], three-layer features can be obtained: the first layer [0.026, 0.051... 0.033], the second layer [0.038, 0.062... 0.045], and the third layer [0.047, 0.071... 0.056].

[0090] After generating the multi-level features, the system inputs them into the compensation strategy generation module. This module introduces a dynamic response enhancement function to enhance the multi-level features. The dynamic response enhancement function is designed to be adaptively adjusted based on the feature energy level difference, amplifying the feature changes with information value and suppressing noise interference. The feature energy level difference is calculated as the mean square error between adjacent feature levels. For example, the energy level difference between the first layer and the second layer is 0.0153, and the energy level difference between the second layer and the third layer is 0.0112.

[0091] Based on these differences, the system constructs a response curve to perform a non-linear transformation on the input features. For example, when the energy level difference is greater than the threshold of 0.01, the enhancement coefficient is set to 1.35; when the energy level difference is between 0.005 and 0.01, the enhancement coefficient is 1.15; when the energy level difference is less than 0.005, the enhancement coefficient is 0.95. After applying these enhancement coefficients, the system obtains enhanced multi-level features and generates an initial compensation strategy vector [0.052, 0.078... 0.067] through weighted combination.

[0092] Based on the initial compensation strategy, the system constructs a multi-scale selection pressure index. The system pre-sets reference strategies at three scales: the micro-scale reference value [0.060, 0.085... 0.072], the meso-scale reference value [0.055, 0.080... 0.070], and the macro-scale reference value [0.050, 0.075... 0.065]. The system calculates the normalized distances between the initial compensation strategy and the reference strategies at each scale: the micro-scale distance is 0.183, the meso-scale distance is 0.135, and the macro-scale distance is 0.092. These distance values form the multi-scale selection pressure index [0.183, 0.135, 0.092].

[0093] To improve the system stability, it is necessary to apply stability protection to the multi-scale selection pressure index. A dynamic modulation field is introduced to enhance the structure of the selection pressure index. The dynamic modulation field is adjusted based on the real-time changes of the system state and provides continuous stability protection through closed-path integration. Specifically, when implementing, the change rate of environmental parameters and the system response delay are monitored to construct a modulation field intensity matrix. When the change rate of environmental parameters is 0.5 °C per minute and the system response delay is 200 milliseconds, the modulation field intensity is set to 1.2; when the change rate of environmental parameters is 0.3 °C per minute and the system response delay is 150 milliseconds, the modulation field intensity is 1.1; when the change rate of environmental parameters is 0.1 °C per minute and the system response delay is 100 milliseconds, the modulation field intensity is 1.0. After applying the modulation field, the original selection pressure index [0.183, 0.135, 0.092] is converted into a steady-state enhanced selection pressure index [0.220, 0.149, 0.092].

[0094] The system substitutes the selection pressure index with enhanced steady state into the calculation process of the evolutionary modulation factor. The system sets weight coefficients at different scales according to the current system state: the weight at the micro scale is 0.25, the weight at the meso scale is 0.35, and the weight at the macro scale is 0.40. These weight coefficients are dynamically adjusted with the system state. For example, when the system is in the startup phase, the weight at the micro scale is increased to 0.40; when the system is in the stable operation phase, the weight at the macro scale is increased to 0.50. Through weighted calculation of the weight coefficients and the selection pressure index with enhanced steady state: 0.25×0.220 + 0.35×0.149 + 0.40×0.092 = 0.140, the evolutionary modulation factor is obtained. This modulation factor reflects the adaptability of the system in a multi-scale environment and is used for fine-tuning of subsequent compensation strategies to enhance the adaptability of the system to complex environments.

[0095] In summary, through the fusion processing of dynamic feature vectors and adaptive attention weights, combined with dynamic response enhancement, construction of multi-scale selection pressure indicators, stability protection, and calculation of evolutionary modulation factors, the system can generate compensation strategies with high adaptability and stability, effectively coping with complex and changeable operating environments.

[0096] Figure 5 This is a bar chart comparing the effects of the multi-scale selection pressure index and the evolutionary modulation factor in the embodiments of the present invention: This figure shows the comparison results of the basic method, the single-scale modulation method, and the multi-scale evolutionary modulation method in four key performance indicators. In terms of anti-interference ability, the multi-scale evolutionary modulation method reaches 118.5%, significantly superior to 100.2% of the single-scale modulation method and 90.5% of the basic method. In the steady-state response accuracy test, the multi-scale evolutionary modulation method performs best, reaching 121.2%, while the single-scale modulation method is 103.8% and the basic method is only 82.4%. In the dynamic tracking efficiency evaluation, the multi-scale evolutionary modulation method continues to maintain the leading advantage, reaching 124.8%, the single-scale modulation method is 109.6%, and the basic method is 92.7%. In the system stability protection index, the performance of the multi-scale evolutionary modulation method is more prominent, reaching a high level of 138.2%, far exceeding 115.5% of the single-scale modulation method and 99.8% of the basic method. The data shows that the multi-scale evolutionary modulation method shows significant advantages in all evaluation indicators, especially in system stability protection, where its performance improvement is the most obvious, with an increase of nearly 40 percentage points compared to the basic method, fully demonstrating the comprehensive advantages of this method in complex system control.

[0097] In an alternative embodiment, continuous stability protection is provided through the closed-path integration of the dynamic modulation field, and the generation of the selection pressure index with enhanced steady state includes: The dynamic modulation field is dynamically adjusted based on the changing trend of the real-time state of the system, and the stability requirements of the system are reflected through the changing characteristics of the dynamic modulation field; Perform a closed-path integration operation on the dynamic modulation field. The closed-path integration operation is calculated along a closed trajectory in the system state space to generate a stability protection factor, which is used to characterize the overall stability characteristics of the system; Fuse the stability protection factor with the selection pressure index, and impose a continuity constraint on the selection pressure index through the stability protection factor to generate a steady-state enhanced selection pressure index with anti-interference ability.

[0098] Establish a dynamic modulation field, which is dynamically adjusted based on the changing trend of the real-time state of the system. In specific implementation, key parameters of the system are collected as state vectors, including multi-dimensional indicators such as resource utilization rate, response time, throughput, and error rate. Perform time-series analysis on the state vectors at each time point t, calculate the change rate between adjacent sampling points, and form a change trend matrix. According to the change trend matrix, construct a dynamic modulation field function, which maps the system state space to the modulation field intensity space. When the system fluctuates violently, the modulation field intensity increases accordingly; when the system tends to be stable, the modulation field intensity decreases accordingly.

[0099] For example, in the cloud computing resource scheduling scenario, if the system CPU utilization rate rapidly rises from 75% to 95% and the memory utilization rate rises from 60% to 85%, the calculated change rates are 20% and 25% respectively. The dynamic modulation field will determine that this change rate exceeds the safe range according to the preset threshold, and thus generate a higher field intensity value in the corresponding state area, such as the field intensity rising from 0.2 to 0.8.

[0100] Perform a closed-path integration operation, which is calculated along a closed trajectory in the system state space to generate a stability protection factor. In actual operation, select a representative closed path in the system state space, which covers typical operating state points of the system. Discretize the state space into finite sampling points, measure the dynamic modulation field intensity at each sampling point, and then use numerical integration methods to accumulate the product of these field intensity values and the path micro-elements along the closed path to obtain the closed-path integration result.

[0101] During specific implementation, a discrete point sampling method can be adopted. 128 sampling points are evenly selected in the state space to form a closed path. The modulation field intensity at each sampling point is recorded, and the trapezoidal integration method is used to accumulate and calculate the path integral value. For example, for a certain sampling, the modulation field intensity values at each point on the closed path are [0.35, 0.42, 0.56, 0.78, 0.65, 0.47, 0.39, 0.32], and the path element length is uniformly set to 0.25. Then the calculated path integral value is 0.985. This value is the original value of the stability protection factor. Through normalization processing, the final stability protection factor is 0.657.

[0102] After the stability protection factor is calculated, it is fused with the selection pressure index to generate a steady-state enhanced selection pressure index with anti-interference ability. The selection pressure index is a key quantity guiding the evolution direction of the system and is usually composed of factors such as system efficiency, resource consumption, and response time. The fusion process adopts a weighted combination method, adding the stability protection factor as a continuity constraint term to the original selection pressure index.

[0103] During specific implementation, let the original selection pressure index be SPO and the stability protection factor be SPF. Then the calculation process of the steady-state enhanced selection pressure index ESPO is as follows: First, determine the basic weight coefficient α and the stability weight coefficient β, and these two coefficients satisfy the constraint condition of α + β = 1; then calculate ESPO = α×SPO + β×SPF×adjustment function. The adjustment function is adaptively adjusted according to the current stable state of the system. When the system is in a high-stability region, the adjustment function value is relatively small, such as 0.2; when the system is in a low-stability region, the adjustment function value is relatively large, such as 0.9.

[0104] Taking network traffic control as an example, the original selection pressure index SPO is 0.72, the calculated stability protection factor SPF is 0.657, the system is currently in a medium-stability state, the adjustment function value is 0.5, the weight coefficient α = 0.6, and β = 0.4. Then the steady-state enhanced selection pressure index ESPO = 0.6×0.72 + 0.4×0.657×0.5 = 0.5634.

[0105] The steady-state enhanced selection pressure index generated by this method has the following characteristics: In the stable region of the system, it maintains the dominant role of the original selection pressure and promotes system optimization; in the unstable region of the system, the constraint role of the stability protection factor is enhanced to prevent the system from making decisions that lead to instability; overall, it shows the characteristic of smooth transition and avoids decision jitter.

[0106] In practical applications, this method can be implemented in scenarios such as resource scheduling, load balancing, and network traffic control. For example, in a distributed database system, when it is detected that the query load of a certain node has increased sharply from 100 times per second to 450 times per second, the original selection pressure index suggests immediately transferring 50% of the load to other nodes. However, the enhanced selection pressure index constrained by the stability protection factor suggests a progressive transfer of the load in three steps, with each transfer being 16.7%, avoiding system jitter caused during the load migration process.

[0107] By providing continuous stability protection through the closed-path integration of the dynamic modulation field, the system can maintain stable operation while optimizing performance, reduce system fluctuations caused by sudden changes, and improve the overall robustness and reliability of the system.

[0108] In a second aspect of the embodiments of the present invention, there is provided an electronic device, including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0109] In a third aspect of the embodiments of the present invention, there is provided a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0110] The present invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium, on which computer-readable program instructions for executing various aspects of the present invention are uploaded.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A dynamic automatic analysis method for the center distance of worm and worm gear meshing based on servo drive, characterized in that, Including: Collect the state parameters of the worm and worm gear drive system to generate an initial measurement data set; Use the initial measurement data set to train a neural network adaptive filter. The neural network adaptive filter is based on the long short-term memory network architecture, establishes a non-linear mapping relationship with the center distance signal, and performs real-time correction on the center distance measurement result by introducing a temperature compensation coefficient, and outputs the corrected center distance data; According to the corrected center distance data, combined with the worm speed signal and the worm speed signal, establish a dynamic model of the worm and worm gear drive system, calculate the theoretical optimal center distance value, compare the theoretical optimal center distance value with the corrected center distance data, and obtain the center distance deviation sequence; Based on the center distance deviation sequence, use a bidirectional gated recurrent neural network enhanced by an attention mechanism to predict the center distance change trend by analyzing the temporal characteristics in the historical compensation data, and continuously optimize the compensation strategy in combination with the online transfer learning method, and output the optimal compensation instruction; Convert the optimal compensation instruction into a displacement control signal of the servo motor, drive the servo motor to perform the center distance adjustment operation, and realize the dynamic compensation of the meshing center distance of the worm and worm gear.

2. The method according to claim 1, wherein Use the initial measurement data set to train a neural network adaptive filter. The neural network adaptive filter is based on the long short-term memory network architecture, establishes a non-linear mapping relationship with the center distance signal, and performs real-time correction on the center distance measurement result by introducing a temperature compensation coefficient, and outputs the corrected center distance data including: Collect the real-time operation data of the worm and worm gear drive system, generate a multi-modal measurement data set based on the real-time operation data; establish a high-fidelity mapping environment of the drive system based on the multi-modal measurement data set, construct a non-linear mapping relationship between the center distance signal and the multi-modal measurement data set in the high-fidelity mapping environment, and perform mapping processing on the initial center distance data through the non-linear mapping relationship to obtain the optimized center distance data; According to the optimized center distance data, calculate the temperature compensation coefficient by combining the temperature data in the multi-modal measurement data set. The temperature compensation coefficient is obtained through an exponential decay function of the initial compensation coefficient and the temperature change amount. Perform weighted fusion on the temperature compensation coefficient and the optimized center distance data, and introduce a dynamic bias term for compensation to generate a compensated center distance measurement value; Input the compensated center distance measurement value into the high-fidelity mapping environment, calculate the compensation parameter based on the historical compensation data in the high-fidelity mapping environment, and perform optimized correction on the compensated center distance measurement value through the compensation parameter, and output the corrected center distance measurement data.

3. The method according to claim 1, wherein According to the corrected center distance data, combined with the worm speed signal and the worm speed signal, establish a dynamic model of the worm and worm gear drive system, calculate the theoretical optimal center distance value, compare the theoretical optimal center distance value with the corrected center distance data, and obtain the center distance deviation sequence including: Collect the dynamic data set of the worm and worm gear drive system, where the dynamic data set is used to characterize the real-time operating state of the drive system; construct a spiking neural network model based on the dynamic data set, where the spiking neural network model simulates the characteristics of the drive system through the sodium ion channel conductance and the potassium ion channel conductance, receives the dynamic data set based on the sodium ion channel conductance and the potassium ion channel conductance, and calculates the dynamic current, and generates a synaptic modulation torque by weighted combination of the dynamic current and the basic driving torque; Substitute the synaptic modulation torque into the enhanced dynamic equation, and obtain the dynamic response characteristics of the drive system by solving the enhanced dynamic equation; perform time-domain encoding on the dynamic response characteristics through the spiking neural network model to generate a pulse sequence signal, where the pulse sequence signal describes the time-varying characteristics of the dynamic characteristics through a time exponential decay function, and obtain the dynamic characteristic pulse encoding; Based on the dynamic characteristic pulse encoding, use the spiking neural network model to construct a time-varying kernel function, where the time-varying kernel function is determined by the dynamically reconstructed synaptic structure, perform spatio-temporal domain convolution operation on the dynamic characteristic pulse encoding through the time-varying kernel function, obtain the instantaneous state characteristics of the drive system, and calculate the theoretical optimal center distance based on the instantaneous state characteristics; Compare the theoretical optimal center distance with the corrected center moment data to obtain the center distance deviation data.

4. The method according to claim 3, wherein, Substitute the synaptic modulation torque into the enhanced dynamic equation, and obtain the dynamic response characteristics of the drive system by solving the enhanced dynamic equation; Perform time-domain encoding on the dynamic response characteristics through the spiking neural network model to generate a pulse sequence signal including: Construct a field theory enhanced dynamic equation based on the synaptic modulation torque, where the field theory enhanced dynamic equation introduces a field-matter coupling force term, and obtain the field state distribution of the drive system by solving the field theory enhanced dynamic equation, where the field state distribution reflects the overall dynamic state of the drive system; Substitute the field state distribution into the field theory enhanced dynamic equation, dynamically update the coupling coefficient in the field-matter coupling force term according to the real-time operating state of the drive system, and adjust the parameters of the field theory enhanced dynamic equation in combination with the change law of the synaptic modulation torque, and solve the field theory enhanced dynamic equation again to obtain the coupled dynamic response of the drive system, where the coupled dynamic response characterizes the nonlinear dynamic behavior of the drive system; Apply a field state projection transformation to the coupled dynamic response, extract the characteristics of the system velocity component, acceleration component, and field state component, and generate a feature vector characterized by the field state through a linear combination of the feature mode functions, where the feature vector contains the multi-scale dynamic information of the drive system; Input the feature vector characterized by the field state into a spiking neural network, perform time-domain encoding on the dynamic characteristics through the spiking neural network, introduce a timestamp function and an exponential decay function to modulate the encoding process, and generate a pulse sequence signal with time-varying characteristics.

5. The method according to claim 1, wherein Based on the center distance deviation sequence, a bidirectional gated recurrent neural network enhanced by an attention mechanism predicts the center distance change trend by analyzing the temporal features in historical compensation data, and continuously optimizes the compensation strategy in combination with the online transfer learning method. The output optimal compensation instruction includes: Construct a bidirectional gated recurrent neural network, input the forward hidden layer state and the backward hidden layer state of the bidirectional gated recurrent neural network into the swarm intelligence emergence layer, and the swarm intelligence emergence layer introduces a swarm cooperation factor to enhance the hidden layer state and generate enhanced hidden layer features; Construct a self-evolving attention weight based on the enhanced hidden layer features. The self-evolving attention weight calculates the fitness function through the ratio of the prediction error to the reference error threshold, and multiplies the fitness function by the attention base weight to obtain the adaptive attention weight; Establish an environmental pressure selection mechanism according to the adaptive attention weight. The environmental pressure selection mechanism calculates the transfer probability based on the adaptation distance from the source domain to the target domain, uses the transfer probability as the weight coefficient of the source domain loss function, and updates the transfer learning model parameters through gradient descent; Modulate the enhanced hidden layer features based on the transfer learning model parameters to generate a dynamic feature vector. Input the dynamic feature vector and the adaptive attention weight into the compensation strategy generation module. The compensation strategy generation module introduces an evolutionary modulation factor, and the evolutionary modulation factor is calculated through the cumulative effect of the multi-scale selection pressure index; Multiply the evolutionary modulation factor by the initial compensation instruction to obtain the optimal compensation instruction, and feedback the compensation effect corresponding to the optimal compensation instruction to the fitness function to update the self-evolving attention weight and realize the dynamic optimization of the compensation strategy.

6. The method according to claim 5, wherein Input the dynamic feature vector and the adaptive attention weight into the compensation strategy generation module. The compensation strategy generation module introduces an evolutionary modulation factor, and the evolutionary modulation factor is calculated through the cumulative effect of the multi-scale selection pressure index, including: Combine the dynamic feature vector and the adaptive attention weight to generate multi-level features, input the multi-level features into the compensation strategy generation module, and the compensation strategy generation module introduces a dynamic response enhancement function to enhance the multi-level features. The dynamic response enhancement function is adaptively adjusted based on the feature energy level difference to generate an initial compensation strategy, and the initial compensation strategy reflects the dynamic response characteristics of the system; Construct a multi-scale selection pressure index based on the initial compensation strategy. The multi-scale selection pressure index is calculated through the normalized distance between the initial compensation strategy and the reference strategy, and the reference strategy sets corresponding reference values at different scales; Apply stability protection to the multi-scale selection pressure index, enhance the structure of the multi-scale selection pressure index through a dynamic modulation field. The dynamic modulation field is dynamically adjusted based on the real-time change of the system state, and provides continuous stability protection through the closed path integral of the dynamic modulation field to generate a steady-state enhanced selection pressure index, and the steady-state enhanced selection pressure index has anti-interference ability; Substitute the steady-state enhanced selection pressure index into the evolutionary modulation factor calculation formula, and perform weighted calculation on the steady-state enhanced selection pressure index through different scale weight coefficients. The weight coefficients are adaptively adjusted according to the scale change, and the collaborative optimization of multi-scale features is achieved through the dynamic allocation of the different scale weight coefficients to obtain the evolutionary modulation factor.

7. The method according to claim 1, characterized in that, Provide continuous stability protection through the closed-path integral of the dynamic modulation field. The generation of the steady-state enhanced selection pressure index includes: The dynamic modulation field is dynamically adjusted based on the change trend of the real-time state of the system, and the stability requirement of the system is reflected through the change characteristics of the dynamic modulation field; Perform a closed-path integral operation on the dynamic modulation field. The closed-path integral operation is calculated along the closed trajectory of the system state space to generate a stability protection factor, and the stability protection factor is used to characterize the overall stability characteristics of the system; Fuse the stability protection factor with the selection pressure index, and impose a continuity constraint on the selection pressure index through the stability protection factor to generate a steady-state enhanced selection pressure index with anti-interference ability.

8. An electronic device, characterized in that, Including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, the method according to any one of claims 1 to 7 is implemented.

Images