A variable-ratio rack-and-pinion pair machining error detection method and system

Abstract

The application provides a variable transmission ratio rack-pinion pair machining error detection method and system, and relates to the field of transmission mechanism design. In view of the problem that the existing means is difficult to accurately detect the error of the variable transmission ratio rack-pinion pair, the kinetic characteristic data under the known multiple errors and multiple working conditions is acquired, a comprehensive virtual sample library is constructed, the parameters are reconfigured and optimized by using a two-dimensional comprehensive weight index, and a nonlinear mapping model between the error, the working condition and the response is established based on a neural network. The blank of the scarce fault sample is filled by using high-fidelity simulation data. Through the reverse calculation logic of the neural network model, the macro working condition and the response signal perceived by the sensor can be inversely decoupled into specific micro error types, so that the nonlinear time-varying interference of the variable transmission ratio structure is overcome without expensive destructive testing, and low-cost error detection of the variable transmission ratio rack-pinion pair error is realized.

Description

Technical Field

[0001] This invention relates to the field of transmission mechanism design, specifically to a method and system for detecting machining errors in a variable transmission ratio rack-gear sector pair. Background Technology

[0002] Variable ratio gear meshing mechanisms are widely used in precision mechanical transmissions due to their ability to balance low-speed flexibility and high-speed stability, with the variable ratio rack-sector gear pair inside automotive recirculating ball steering systems being the most typical example. Currently, the industry generally adopts a "offline static measurement" approach supplemented by "online threshold monitoring" for detecting the meshing accuracy and transmission error of such key components. Specifically, at the manufacturing end, static geometric verification of tooth profile deviation and cumulative tooth pitch error is mostly performed using coordinate measuring machines or gear measuring centers; at the vehicle service end, the input-output ratio of the overall transmission chain is mainly monitored through torque or angle sensors. Although the existing inspection system can eliminate explicit defects in the manufacturing process, it often lacks effective means to perceive the microscopic response changes of variable ratio structures caused by manufacturing errors during transmission, making it difficult to deeply explore the changes in the microscopic state of the meshing interface.

[0003] However, existing detection methods still fall short of meeting the comprehensive demands of modern industry for "precise identification of machining errors." Taking recirculating ball steering gears as an example, due to the nonlinear variation of its transmission ratio with the steering angle, the contact stress and meshing stiffness at different steering angles exhibit significant time-varying characteristics. Machining defects such as tooth profile errors and installation deviations are easily strongly coupled with complex operating loads, rendering traditional linear diagnostic models ineffective. Furthermore, intelligent detection schemes based on neural networks face challenges due to the scarcity of high-value samples. Obtaining experimental data covering all operating conditions and error types in real-world scenarios is costly and difficult, resulting in poor generalization ability of existing systems when faced with unseen machining errors. This prevents accurate reverse tracing from macroscopic response signals to microscopic machining error types. Therefore, without sufficient physical error samples, it is difficult to overcome the interference of strong nonlinearity in variable transmission ratio structures coupled with operating conditions, and it is difficult to meet the detection requirements for accurately identifying the types or forms of machining errors in variable transmission ratio rack-sector pairs based on the response. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by providing a method and system for detecting machining errors in rack-and-gear pairs with variable transmission ratios. Through the logic of dynamic simulation library construction, forward neural network construction, and inverse neural network detection, the invention achieves the detection of rack-and-gear pairs with variable transmission ratios, thereby reducing detection costs and maintenance difficulties.

[0005] The primary objective of this invention is to provide a method and system for detecting machining errors in a rack-and-gear pair with a variable transmission ratio, employing the following approach:

[0006] include: Acquire dynamic characteristic data of a rack-and-gear pair with known variable transmission ratio under multiple errors and operating conditions. The dynamic characteristic data includes error categories, operating condition data, and corresponding response data. A dynamic characteristic database is constructed by using a two-dimensional comprehensive weight index to classify the error type parameters and working condition parameters of the variable transmission ratio rack-sector pair. Based on the dynamic characteristic database, a neural network model of the error-operating condition-response relationship is constructed; By using an error-operating condition-response relationship neural network model, the operating condition data and response data sensed by the sensor are reversed to obtain the error types of the variable transmission ratio rack-sector pair.

[0007] Furthermore, the acquisition of dynamic characteristic data of a known variable transmission ratio rack-sector pair under multiple errors and operating conditions includes: A dynamic model of a rack-and-pinion pair with a variable transmission ratio is constructed. The dynamic model includes meshing mechanism, load transfer and error characteristic parameters. Based on the dynamic model, simulations were performed on the variable transmission ratio rack-sector pair under different working conditions with different error settings. The transmission ratio, meshing force, and meshing part data output from the simulation are collected as response data.

[0008] Furthermore, the multiple error categories include tooth profile error, tooth pitch error, gear sector shaft eccentricity error, and gear sector shaft perpendicularity error; The multi-condition data includes steering speed and load torque in scenarios such as low-speed maneuvering, urban roads, suburban roads, high-speed cruising, and emergency steering.

[0009] Furthermore, the weighting of error type parameters and operating condition parameters of the variable transmission ratio rack-sector pair using a dual-dimensional comprehensive weighting index includes: The parameters are assigned values ​​based on their causal importance in the meshing process to determine the priority weight of the mechanism. Based on the simulation data, the response change rate caused by a unit change in the statistical parameters is obtained. The higher the response change rate, the greater the weight. The response influence rate weight of each parameter is obtained. The comprehensive weight is calculated based on the mechanism priority weight and the response impact rate weight, and values ​​are assigned to the error type parameters and the operating condition parameters. A database of dynamic characteristics of variable transmission ratio rack-and-sector pair was established based on response data with comprehensive weights.

[0010] Furthermore, after normalizing the input parameters using comprehensive weights, the data features of high-weight parameters are enhanced by weighting them using comprehensive weights. The input parameters are divided into high-weight layer data and low-weight layer data based on the magnitude of the overall weight; When constructing the dynamic characteristics database, priority indexes are established for high-weight layer data so that the error-operating condition-response relationship neural network model can extract features preferentially.

[0011] Furthermore, the neural network model for constructing the error-operating condition-response relationship includes forward modeling and inverse modeling; Among them, forward modeling takes the weighted error type parameters and operating condition parameters as inputs and the response data as outputs, and trains the neural network to learn the forward mapping relationship between error and response; Reverse modeling takes response data and operating condition data as input and error types as output. It uses the dataset generated by the forward mapping relationship for training to establish a reverse inference model.

[0012] Furthermore, when training the forward model, a loss function containing a weight penalty term is used; the weight penalty term is used to constrain the prediction bias of the response data. When training the model, the comprehensive weights corresponding to the parameters are introduced as coefficients, so that parameters with high comprehensive weights are penalized more when they produce prediction bias, thereby improving the model's fitting accuracy to key parameters.

[0013] Furthermore, the error-condition-response neural network model includes an input layer, a hidden layer, and an output layer: The hidden layers are set as a multi-layer structure, where the first hidden layer is used to extract features related to high-weight parameters, and subsequent hidden layers are used to fit nonlinear mapping relationships; The output layer uses a probabilistic output function, and the output result is the probability value of various error types of the current variable transmission ratio rack-sector pair input.

[0014] Furthermore, the method for acquiring the operating condition data and response data sensed by the sensor is as follows: The system acquires operating condition data, including steering speed and load torque, through torque sensors, angle sensors, and displacement sensors. Response data including transmission ratio fluctuation, peak meshing force, and meshing parts are obtained through angle sensors, displacement sensors, and strain gauge sensors.

[0015] A second objective of this invention is to provide a variable transmission ratio rack-and-pinion pair error detection system as described in the first objective, comprising: The data acquisition module is configured to acquire dynamic characteristic data of a known variable transmission ratio rack-sector pair under multiple errors and multiple working conditions. The dynamic characteristic data includes error categories, working condition data, and corresponding response data. The weighting module is configured to: use a two-dimensional comprehensive weighting index to weight the error type parameters and working condition parameters of the variable transmission ratio rack-sector pair, and build a dynamic characteristic database; The modeling module is configured to: construct a neural network model of the error-operating condition-response relationship based on the dynamic characteristic database; The detection module is configured to use an error-operating condition-response relationship neural network model to reverse-calculate the operating condition data and response data sensed by the sensor, and obtain the error types of the variable transmission ratio rack-sector pair.

[0016] Compared with the prior art, the advantages and positive effects of this invention are: To address the challenge of accurately detecting errors in variable ratio rack-and-sector gear pairs using existing methods, this study constructs a comprehensive virtual sample library by acquiring dynamic characteristic data under multiple known errors and operating conditions. A dual-dimensional comprehensive weighting index is used to reconstruct and optimize parameters, and a nonlinear mapping model between error, operating condition, and response is established based on a neural network. High-fidelity simulation data fills the gap in scarce fault samples, and the reverse calculation logic of the neural network model decouples the macroscopic operating conditions and response signals sensed by sensors into specific microscopic error types. This effectively overcomes the nonlinear time-varying interference of variable ratio structures without requiring expensive destructive testing, achieving low-cost, high-precision detection of errors in variable ratio rack-and-sector gear pairs.

[0017] By adopting a step-by-step strategy of "first establishing physical laws through forward modeling, and then achieving inversion detection through inverse modeling," the high-dimensional feature mapping relationship learned by the forward network is used to guide the training of the inverse network. This effectively solves the problem of non-unique solutions or non-convergence in direct inverse modeling. Furthermore, a loss function with weight penalty terms is introduced during the training process, which forces the model to impose a greater penalty on the prediction bias of high comprehensive weight parameters during iterative optimization. This is equivalent to introducing an attention mechanism at the algorithm level, ensuring that the model can prioritize the accuracy of identifying core error types when facing complex conditions with multi-source error coupling, thereby enhancing the robustness and reliability of the system in practical applications.

[0018] By constructing a two-dimensional comprehensive index system that includes mechanistic priority weights and response influence rate weights, the physical causal logic and data statistical laws in the rack-sector meshing process are combined. On the one hand, this avoids the pure data-driven model from getting stuck in local optima in the early stages of training due to a lack of physical constraints. On the other hand, sensitivity analysis quantifies the actual contribution of each parameter to the system response. By performing hierarchical processing and priority indexing of data based on comprehensive weights, the neural network can automatically focus on high-sensitivity and high-value feature parameters during training, effectively suppressing the interference of low-weight noise data. This improves the convergence speed and detection accuracy of the model when processing nonlinear signals with complex transmission ratio structures. Attached Figure Description

[0019] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0020] Figure 1 This is a flowchart illustrating the method for detecting machining errors in a variable transmission ratio rack-and-gear pair according to one or more embodiments of the present invention.

[0021] Figure 2 This is a schematic diagram of an error-condition-response neural network model in one or more embodiments of the present invention.

[0022] Figure 3 This is a schematic diagram illustrating the reverse calculation using an error-condition-response relationship neural network model in one or more embodiments of the present invention.

[0023] Figure 4 This is a schematic diagram illustrating the error type detection of a variable transmission ratio rack-sector pair in one or more embodiments of the present invention. Detailed Implementation

[0024] Example 1 In a typical embodiment of the present invention, such as Figures 1-4 As shown, a method for detecting machining errors in a variable transmission ratio rack-gear sector pair is presented, including: Acquire dynamic characteristic data of a rack-and-gear pair with known variable transmission ratio under multiple errors and operating conditions. The dynamic characteristic data includes error categories, operating condition data, and corresponding response data. A dynamic characteristic database is constructed by using a two-dimensional comprehensive weight index to classify the error type parameters and working condition parameters of the variable transmission ratio rack-sector pair. Based on the dynamic characteristic database, a neural network model of the error-operating condition-response relationship is constructed; By using an error-operating condition-response relationship neural network model, the operating condition data and response data sensed by the sensor are reversed to obtain the error types of the variable transmission ratio rack-sector pair.

[0025] Among them, obtaining dynamic characteristic data of rack-and-gear sector pairs with known transmission ratios under multiple errors and operating conditions includes: A dynamic model of a rack-and-pinion pair with a variable transmission ratio is constructed. The dynamic model includes meshing mechanism, load transfer and error characteristic parameters. Based on the dynamic model, simulations were performed on the variable transmission ratio rack-sector pair under different working conditions with different error settings. The transmission ratio, meshing force, and meshing part data output from the simulation are collected as response data.

[0026] Among them, the multiple error categories include tooth profile error, tooth pitch error, tooth sector shaft eccentricity error, and tooth sector shaft perpendicularity error; The multi-condition data includes steering speed and load torque in scenarios such as low-speed maneuvering, urban roads, suburban roads, high-speed cruising, and emergency steering.

[0027] The weighting of error type parameters and operating condition parameters of the variable transmission ratio rack-and-gear sector pair using a two-dimensional comprehensive weighting index includes: The parameters are assigned values ​​based on their causal importance in the meshing process to determine the priority weight of the mechanism. Based on the simulation data, the response change rate caused by a unit change in the statistical parameters is obtained. The higher the response change rate, the greater the weight. The response influence rate weight of each parameter is obtained. The comprehensive weight is calculated based on the mechanism priority weight and the response impact rate weight, and values ​​are assigned to the error type parameters and the operating condition parameters. A database of dynamic characteristics of variable transmission ratio rack-and-sector pair was established based on response data with comprehensive weights.

[0028] Furthermore, the input parameters are weighted and normalized using comprehensive weights to enhance the data characteristics of high-weight parameters; The input parameters are divided into high-weight layer data and low-weight layer data based on the magnitude of the overall weight; When constructing the dynamic characteristics database, a priority index is established for the data of the high-weight layer so that the error-operating condition-response relationship neural network model can extract features first.

[0029] Among them, the neural network model for constructing the error-operating condition-response relationship includes forward modeling and inverse modeling; Among them, forward modeling takes the weighted error type parameters and operating condition parameters as inputs and the response data as outputs, and trains the neural network to learn the forward mapping relationship between error and response; Reverse modeling takes response data and operating condition data as input and error types as output. It uses the dataset generated by the forward mapping relationship for training to establish a reverse inference model.

[0030] Furthermore, when training the forward model, a loss function with a weight penalty term is used. The weight penalty term is used to constrain the prediction bias of the response data. During model training, the comprehensive weights corresponding to the parameters are introduced as coefficients, so that parameters with high comprehensive weights are penalized more when they produce prediction bias, thereby improving the model's fitting accuracy to key parameters.

[0031] The error-condition-response neural network model includes an input layer, a hidden layer, and an output layer. The hidden layers are set as a multi-layer structure, where the first hidden layer is used to extract features related to high-weight parameters, and subsequent hidden layers are used to fit nonlinear mapping relationships; The output layer uses a probabilistic output function, and the output result is the probability value of various error types of the current variable transmission ratio rack-sector pair input.

[0032] The methods for acquiring the operating condition data and response data sensed by the sensors are as follows: The system acquires operating condition data, including steering speed and load torque, through torque sensors, angle sensors, and displacement sensors. Response data, including transmission ratio fluctuations, peak meshing force, and meshing points, are obtained through angle sensors, displacement sensors, and strain gauge sensors.

[0033] Specifically, in combination Figures 1-4 To address the problem of error detection in variable ratio rack-and-pinion pairs, this embodiment provides a method for detecting such errors. This method enables the detection of errors in the variable ratio rack-and-pinion pair under various steering conditions, such as... Figure 1 As shown, firstly, dynamic simulations with multiple errors and operating conditions are carried out based on the variable transmission ratio rack and pinion pair model. Dual weight analysis is conducted, a dynamic characteristic database is established, and an error-operating condition-response relationship neural network model is established and trained. The operating condition and response data sensed by the sensor are input into the model, and the reverse calculation capability of the model is used to realize the monitoring of the error of the variable transmission ratio rack and pinion pair.

[0034] Specifically, it includes the following steps: Step 1: Obtain dynamic characteristic data of variable transmission ratio gear sector-rack pair under multiple errors and working conditions.

[0035] This step involves constructing a dynamic model and performing multi-dimensional operating condition simulations to obtain a high-quality simulation dataset covering the entire combination of error, operating conditions, and response, ensuring data integrity and reliability. Specifically, this includes: 1.1 Constructing a high-precision dynamic model of a variable transmission ratio gear sector-rack pair Based on MATLAB software, the coordinates of the variable transmission ratio tooth profile points are calculated according to a given transmission ratio and a known standard tooth profile. Based on the modeling software SOLIDWORKS, the coordinates of the tooth profile points of the variable transmission ratio are imported and meshed into tooth profile lines to establish a three-dimensional model of the variable transmission ratio tooth sector-rack pair. Based on the multibody dynamics software RecurDyn, a refined dynamic model was constructed that includes meshing mechanism, load transfer, and error characteristics. The model parameters include geometric parameters, error parameters, and operating condition parameters.

[0036] Geometric parameters include module, number of teeth, pressure angle, and tooth width; error parameters include tooth profile error, tooth pitch error, eccentricity error of the gear sector shaft, and perpendicularity error of the gear sector shaft; operating parameters include steering speed and load torque.

[0037] 1.2 Generate multi-dimensional, comprehensive operating condition sequences Based on actual automotive usage scenarios, the steering gear operating conditions are divided into 5 typical scenarios, covering all scenarios including daily use, extreme conditions, and environmental differences: low-speed maneuvering, urban road conditions, suburban road conditions, high-speed cruising, and emergency steering conditions. Each condition corresponds to a certain steering speed and load torque. For each typical working condition, the core dynamic parameters of load torque and steering speed are sampled uniformly with gradients, and several working condition points are set. 1.3 Execution Error - Working Condition Simulation and Data Preprocessing Dynamic simulations were performed on all combinations of four types of errors and five working conditions. For each combination, the transmission ratio, normal meshing force, and meshing location were collected as response data. To ensure data quality, the collected response data is preprocessed: outliers are removed to ensure that the data deviation is within a reasonable range; all input parameters (error types, steering speed, load torque) and response data are normalized to eliminate the impact of dimensional differences on training.

[0038] Step 2: Determine parameter priorities based on weight analysis and construct a weighted dynamic characteristic database. This step uses simulation data as the data source to conduct weight analysis, quantifying the influence of each parameter on the response and assigning weights to ensure that the neural network training focuses on high-weight parameters. Specifically, this includes: 2.1 The impact of quantization parameters on priority 2.1.1 Determine the candidate set of parameters Input parameters include error parameters: tooth profile error, tooth pitch error, sector shaft eccentricity error, and sector shaft perpendicularity error; and operating condition parameters: steering speed and load torque. Response parameters include: transmission ratio fluctuation, peak meshing force, meshing force fluctuation, and meshing location.

[0039] 2.1.2 Define the dual-dimensional comprehensive weight calculation index Based on the importance of the causal relationship of parameters in the meshing process, error parameters are the cause of the response, operating condition parameters are the adjustment factors, and response parameters are the output target, thus determining the mechanism priority weight W1. Based on the simulation data, the response change rate caused by the unit change of the statistical parameters is calculated. The higher the change rate, the greater the weight, thus avoiding the subjectivity of the mechanism assignment: for each candidate parameter, 10% of its value range is taken as the unit change amount; other parameters are fixed, and the average change rate of the response parameter is calculated when the parameter changes; the change rate is normalized to obtain the response influence rate weight W2 of each parameter.

[0040] 2.1.3 Calculate the overall weight To ensure logical correctness, the mechanism priority weight W1 is defined as 60%, and the response impact rate weight W2 is defined as 40%, which aligns with the actual patterns in the simulation data. Therefore, the comprehensive weight formula is as follows: W = 0.6 × W1 + 0.4 × W2 2.2 Constructing a weighted dynamic characteristic database Based on the weight analysis results, the parameters are preprocessed with weights and stored in a structured manner to ensure that the neural network prioritizes fitting parameters with higher weights during training, as detailed below: 2.2.1 Weighted Preprocessing of Core Parameters The input parameters are weighted and normalized by multiplying the parameter values ​​by the overall weights to enhance the influence of high-weight parameters. The formula is as follows: X=X norm ×W Among them, X norm is the normalized value of the parameter, and W is the overall weight of the parameter.

[0041] 2.2.2 Data storage optimization The core input parameters are divided into high-weight layers (W≥0.25) and low-weight layers (W<0.25) according to their comprehensive weights. They are stored in the database using a hierarchical index. During neural network training, the high-weight layer data is read first to improve fitting efficiency. Based on the established error-operating condition-response steering gear dynamic characteristic database, 200,000 sets of data were generated for training the neural network, and 20,000 sets of data were generated for verification. Both the training set and the verification set data were obtained from simulation data. Step 3: Train the forward neural network and establish the error-operating condition-response mapping relationship. Based on a weighted dynamic characteristic database, a forward neural network is trained using weighted error parameters and operating condition parameters as inputs to enhance the fitting priority of high-weight parameters, as detailed below: 3.1 Forward Neural Network Design Input layer: Number of nodes = Number of weighted input parameters = 6 (tooth profile error, tooth pitch error, sector shaft eccentricity error, sector shaft perpendicularity error, steering speed, load torque), the input data are the weighted preprocessed parameter values; Hidden layers: There are 2 hidden layers. The first layer has 35 nodes and focuses on extracting features of high-weight parameters. The second layer has 45 nodes and fits the nonlinear mapping of weight parameters to response. Features of high-weight parameters are passed first. Output layer: Number of nodes = Number of output parameters = 4 (transmission ratio fluctuation value, peak meshing force, meshing force fluctuation value, meshing part); Activation functions: The hidden layers use the ReLU function, and the output layer uses a linear activation function.

[0042] The established forward neural network, such as Figure 2 As shown.

[0043] 3.2 Training Parameter Settings The Adam optimizer is used with an initial learning rate of 1×10⁻⁶. -4 ; Normalized mean square error (MSE) is an indicator used to quantify the accuracy of model monitoring. The formula for calculating MSE is as follows:

[0044] In the formula, and These are the monitored value and the actual response value of the positive relational neural network, respectively. m is the total number of iterations, and i is the current step number.

[0045] The loss function introduces a weighted prediction bias penalty term on top of the mean squared error (MSE). This term imposes a greater penalty on the response output bias corresponding to parameters with high overall weights, thereby enhancing the fitting accuracy of key parameters. The specific formula is as follows:

[0046] in, The combined weight of the j-th input parameter. To respond to the actual value of the parameter, To compensate for the predicted values ​​of the response parameters, a penalty term is used to constrain the prediction bias of the response corresponding to high-weight parameters, which significantly improves the model's fitting accuracy to key parameters. To avoid model overfitting, set the learning rate to [value]. The Adam optimizer was selected, and the number of training epochs was set to 1000. The convergence of the positive relation neural network training process was determined by the mean squared error value after 1000 iterations.

[0047] Step 4: Train the inverse neural network to detect error types. Based on the error-condition-response mapping of the forward network, an inverse neural network is trained, taking measurable weighted parameters as input and outputting the error type, as detailed below: 4.1 Inverse Neural Network Structure Design Input layer: Number of nodes = Number of measurable weighted parameters = 6 (transmission ratio fluctuation value, peak meshing force, meshing force fluctuation value, meshing part, steering speed, load torque), input data are preprocessed with the same weight; Hidden layers: There are 2 hidden layers. The first layer has 45 nodes and extracts high-dimensional features of error-condition-response, prioritizing the capture of features related to high-weight error parameters. The second layer has 55 nodes and enhances the error type classification ability, prioritizing the matching of high-weight error features. Output layer: Number of nodes = Number of error types = 4, output is the probability of each error type; Activation functions: The hidden layer uses the ReLU function, and the output layer uses the Softmax function.

[0048] The established inverse neural network structure is as follows: Figure 3 As shown.

[0049] 4.2 Inverse Neural Network Training The input samples are the transmission ratio fluctuation value, peak meshing force, meshing force fluctuation value, meshing part, steering speed, and load torque in the database, and the error type is used as the label. Training is stopped after 2000 training rounds and the validation set accuracy is ≥97% (error type accuracy ≥97%).

[0050] Step 5: Use the structure-condition-response relationship neural network to reverse-calculate the error types from the condition-response data sensed by the sensor.

[0051] This embodiment develops the reverse calculation capability of the error-operating condition-response relationship neural network proxy model. It takes response data and operating condition data as input, extracts high-dimensional features through the neural network, outputs structural data, and realizes the detection of error types.

[0052] Specifically, such as Figure 4 As shown, torque sensor, angle sensor, and displacement sensor acquire working condition information, while angle sensor, displacement sensor, and strain gauge sensor acquire response information, which are then input into the trained error-working condition-response neural network.

[0053] The error-condition-response neural network establishes a nonlinear mapping relationship between input and output variables.

[0054] The error-condition-response relationship neural network includes a forward relationship neural network and a backward detection neural network. First, the forward relationship neural network is used to establish the positive relationship between error, condition, and response, that is, the response of various errors under different conditions is known.

[0055] Then, based on the positive relationship between error, operating condition, and response, reverse calculation is performed in the reverse detection neural network. Specifically, the positive relationship neural network takes error parameters such as tooth profile error, tooth pitch error, gear sector shaft eccentricity error, and gear sector shaft perpendicularity error, as well as operating condition parameters such as steering speed and load torque, as inputs. It extracts high-dimensional features through the neural network and outputs transmission ratio fluctuation value, peak meshing force, meshing force fluctuation value, and response parameters of the meshing part.

[0056] The reverse monitoring neural network takes transmission ratio fluctuation, peak meshing force, meshing force fluctuation, meshing part response parameters, and time-varying external working condition load as inputs, extracts high-dimensional features through the neural network, and outputs error categories.

[0057] Example 2 In another typical embodiment of the present invention, such as Figures 1-4 As shown, a variable transmission ratio rack-and-pinion pair error detection system is presented, comprising: The data acquisition module is configured to acquire dynamic characteristic data of a known variable transmission ratio rack-sector pair under multiple errors and multiple working conditions. The dynamic characteristic data includes error categories, working condition data, and corresponding response data. The weighting module is configured to: use a two-dimensional comprehensive weighting index to weight the error type parameters and working condition parameters of the variable transmission ratio rack-sector pair, and build a dynamic characteristic database; The modeling module is configured to: construct a neural network model of the error-operating condition-response relationship based on the dynamic characteristic database; The detection module is configured to use an error-operating condition-response relationship neural network model to reverse-calculate the operating condition data and response data sensed by the sensor, and obtain the error types of the variable transmission ratio rack-sector pair.

[0058] The working method of the variable transmission ratio rack-gear sector pair error detection system is the same as that of the variable transmission ratio rack-gear sector pair machining error detection method in Example 1, and will not be repeated here.

[0059] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting machining errors in a variable transmission ratio rack-gear sector pair, characterized in that, include: Acquire dynamic characteristic data of a known variable transmission ratio rack-sector pair under multiple errors and operating conditions. The dynamic characteristic data includes error categories, operating condition data, and corresponding response data. A dynamic characteristic database is constructed by using a two-dimensional comprehensive weight index to classify the error type parameters and working condition parameters of the variable transmission ratio rack-sector pair. Based on a dynamic characteristic database, a neural network model of the error-operating condition-response relationship is constructed. By using an error-operating condition-response relationship neural network model, the operating condition data and response data sensed by the sensor are reversed to obtain the error types of the variable transmission ratio rack-sector pair; Based on the importance of the causal relationship of parameters in the meshing process, error parameters are the cause of the response, operating condition parameters are the adjustment factors, and response parameters are the output target, thus determining the mechanism priority weight W1. Based on the simulation data, the response change rate caused by a unit change in the parameters is statistically analyzed. The higher the change rate, the greater the weight, thus obtaining the response influence rate weight W2 of each parameter. The mechanism priority weight W1 accounts for 60%, and the response influence rate weight W2 accounts for 40%. The comprehensive weight formula is: W = 0.6 × W1 + 0.4 × W2. After normalizing the input parameters using comprehensive weights, the data features of high-weight parameters are enhanced by weighting with comprehensive weights. The input parameters are divided into high-weight layer data and low-weight layer data according to the magnitude of the comprehensive weights. When constructing the dynamic characteristic database, a priority index is established for high-weight layer data so that the error-operating condition-response relationship neural network model can extract features first. The neural network model for constructing the error-operating condition-response relationship includes forward modeling and inverse modeling; Among them, forward modeling takes the weighted error type parameters and operating condition parameters as inputs and the response data as outputs, and trains the neural network to learn the forward mapping relationship between error and response; Reverse modeling takes response data and operating condition data as input and error types as output. It uses the dataset generated by the forward mapping relationship for training to establish a reverse inference model. When training the forward model, a loss function with a weight penalty term is used. The weight penalty term is used to constrain the prediction bias of the response data. During model training, the comprehensive weights corresponding to the parameters are introduced as coefficients, so that parameters with high comprehensive weights are penalized more when they produce prediction bias, thereby improving the model's fitting accuracy to key parameters.

2. The method for detecting machining errors of a variable transmission ratio rack-sector gear pair as described in claim 1, characterized in that, The acquisition of dynamic characteristic data of a known variable transmission ratio rack-sector pair under multiple errors and operating conditions includes: A dynamic model of a rack-and-pinion pair with a variable transmission ratio is constructed. The dynamic model includes meshing mechanism, load transfer and error characteristic parameters. Based on the dynamic model, simulations were performed on the variable transmission ratio rack-sector pair under different working conditions with different error settings. The transmission ratio, meshing force, and meshing part data output from the simulation are collected as response data.

3. The method for detecting machining errors of a variable transmission ratio rack-sector gear pair as described in claim 2, characterized in that, Multiple error categories include tooth profile error, tooth pitch error, gear sector shaft eccentricity error, and gear sector shaft perpendicularity error; Multi-condition data includes steering speed and load torque in scenarios such as low-speed maneuvering, urban roads, suburban roads, high-speed cruising, and emergency steering.

4. The method for detecting machining errors of a variable transmission ratio rack-sector gear pair as described in claim 1, characterized in that, The method of using a dual-dimensional comprehensive weighting index to weight the error type parameters and operating condition parameters of the variable transmission ratio rack-sector pair includes: The parameters are assigned values ​​based on their causal importance in the meshing process to determine the priority weight of the mechanism. Based on the simulation data, the response change rate caused by a unit change in the statistical parameters is obtained. The higher the response change rate, the greater the weight. The response influence rate weight of each parameter is obtained. The comprehensive weight is calculated based on the mechanism priority weight and the response impact rate weight, and values ​​are assigned to the error type parameters and the operating condition parameters. A database of dynamic characteristics of variable transmission ratio rack-and-sector pair was established based on response data with comprehensive weights.

5. The method for detecting machining errors of a variable transmission ratio rack-sector gear pair as described in claim 1, characterized in that, The error-condition-response neural network model includes an input layer, a hidden layer, and an output layer: The hidden layers are set as a multi-layer structure, where the first hidden layer is used to extract features related to high-weight parameters, and subsequent hidden layers are used to fit nonlinear mapping relationships; The output layer uses a probabilistic output function, and the output result is the probability value of the current variable transmission ratio rack-sector pair belonging to various error types.

6. The method for detecting machining errors of a variable transmission ratio rack-sector gear pair as described in claim 1, characterized in that, The sensor-sensed operating condition data and response data are acquired in the following ways: The system acquires operating condition data, including steering speed and load torque, through torque sensors, angle sensors, and displacement sensors. Response data, including transmission ratio fluctuations, peak meshing force, and meshing points, are obtained through angle sensors, displacement sensors, and strain gauge sensors.

7. A variable transmission ratio rack-gear sector pair error detection system, employing the variable transmission ratio rack-gear sector pair machining error detection method as described in any one of claims 1-6, characterized in that, include: The data acquisition module is configured to acquire dynamic characteristic data of a known variable transmission ratio rack-sector pair under multiple errors and multiple working conditions. The dynamic characteristic data includes error categories, working condition data, and corresponding response data. The weighting module is configured to: use a two-dimensional comprehensive weighting index to weight the error type parameters and working condition parameters of the variable transmission ratio rack-sector pair, and build a dynamic characteristic database; The modeling module is configured to: construct a neural network model of the error-operating condition-response relationship based on the dynamic characteristic database; The detection module is configured to use an error-operating condition-response relationship neural network model to reverse-calculate the operating condition data and response data sensed by the sensor, and obtain the error types of the variable transmission ratio rack-sector pair.

Images