A motor part manufacturing process optimization method based on digital twinning

Abstract

The application discloses a motor parts manufacturing process optimization method based on digital twinning, comprising the following steps: S1, collecting multi-source process parameters in the motor parts manufacturing process; S2, performing process data preprocessing on the multi-source process parameters to generate a standardized process data set; S3, performing manufacturing state evolution modeling and motor parts performance evaluation on the standardized process data set through an improved Mamba network, and outputting manufacturing process performance prediction results; S4, constructing a multi-objective process optimization function, and generating a process parameter optimization solution set by using a hybrid crossover NSGA-II algorithm; S5, selecting an optimal process parameter scheme and performing manufacturing application; S6, collecting feedback data of a target process execution stage in real time, and inputting the feedback data into the improved Mamba network for online optimization iteration; and S7, visually displaying in a graphical form. The application improves the modeling precision, parameter optimization efficiency and process control intelligent level of the motor parts manufacturing process.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent manufacturing and industrial process optimization, and particularly relates to a motor part manufacturing process optimization method based on digital twinning. BACKGROUND

[0002] With the continuous deepening of intelligent manufacturing and industrial digital transformation, the demand for high-precision modeling and process optimization of motor part manufacturing processes is increasing. The existing manufacturing process design and adjustment process mainly relies on operating experience or offline testing, and it is difficult to realize real-time correlation modeling and precise control between process parameters and part performance. In practical applications, the following problems generally exist:

[0003] The multi-source process data in the manufacturing process is asynchronous, the modal distribution is uneven, and there are abnormal fluctuations, which cannot be directly used as modeling basis; traditional process optimization methods mostly use linear regression, response surface method or heuristic search based on experience rules, and lack the ability to describe complex manufacturing stages and state transitions, making it difficult to adapt to typical motor part processing scenarios with frequent working condition disturbances and complex stage switching; although some models can realize physical modeling and performance prediction, they lack online feedback updating mechanism, and cannot perform model self-correction and optimization iteration according to performance deviations occurring in the manufacturing execution process, resulting in limited consistency and yield rate of the final product. In addition, the existing digital twinning modeling methods are mostly used for equipment-level or process-level simulation, and cannot deeply integrate multi-modal process data and microscopic manufacturing behavior characteristics, making it difficult to realize fine process optimization control driven by multi-objective performance indicators in complex processing stages.

[0004] Therefore, how to provide a motor part manufacturing process optimization method based on digital twinning is a problem that those skilled in the art need to solve. SUMMARY

[0005] An object of the present application is to provide a motor part manufacturing process optimization method based on digital twinning. The present application fully integrates an improved Mamba network, a multi-objective process optimization algorithm and an online feedback closed loop, and describes in detail the digital twinning modeling process from multi-source process parameter collection, performance prediction modeling, process optimization decision making to manufacturing feedback closed loop in the motor part manufacturing process. The present application has the advantages of high prediction accuracy, high optimization efficiency and high intelligentization degree of process control.

[0006] According to the motor part manufacturing process optimization method based on digital twinning of the present application, the following steps are included:

[0007] S1, collecting multi-source process parameters in the motor part manufacturing process;

[0008] S2, preprocessing the multi-source process parameters to generate a standardized process data set;

[0009] S3, performing manufacturing state evolution modeling and motor part performance evaluation by the improved Mamba network based on the standardized process data set, and outputting manufacturing process performance prediction results;

[0010] S4, constructing a multi-objective process optimization function based on the manufacturing process performance prediction results, and generating a process parameter optimization solution set by using a hybrid crossover NSGA-II algorithm;

[0011] S5, selecting an optimal process parameter scheme from the process parameter optimization solution set and performing manufacturing application;

[0012] S6, in the execution process, real-time acquisition of feedback data of the target process execution stage, deviation comparison of the feedback data and the manufacturing process performance prediction results, generation of a feedback error vector and input of the feedback error vector to the improved Mamba network for online optimization iteration;

[0013] S7, visualizing the manufacturing process performance prediction results, the optimal process parameter scheme, the feedback error vector and the online optimization iteration results in a graphical form, and constructing a manufacturing process visualization interface.

[0014] Optionally, the multi-source process parameters include temperature parameters, spindle speed parameters, cutting force parameters, feed speed parameters, machining pressure parameters, cooling liquid flow parameters, equipment vibration parameters, displacement deviation parameters and clamping state parameters.

[0015] Optionally, the process data preprocessing specifically includes:

[0016] S21, grouping the multi-source process parameters by modes, the modes including temperature mode, speed mode, mechanical mode, flow mode and displacement mode;

[0017] S22, respectively assigning channel numbers, task numbers and time indexes to each type of process parameters, and generating multi-source process parameters with structured index information;

[0018] S23, filling in missing multi-source process parameters by a weighted moving average method, eliminating invalid values of multi-source process parameters with redundancy or logical conflicts by a time stamp window comparison method, and correcting abnormal values of multi-source process parameters deviating from the statistical range by a modal threshold method;

[0019] S24, time aligning the multi-source process parameters according to a set sampling period;

[0020] S25, normalizing the time-aligned multi-source process parameters, and performing channel alignment, dimension pruning and feature padding operations on the data of each channel, and encapsulating to generate a standardized process data set.

[0021] Optionally, the step S3 specifically comprises:

[0022] S31, input the standardized process data set into the improved Mamba network, the improved Mamba network comprising a multi-modal embedding module, a physical bias injection module, a stage controller module, a state segmentation modeling module, a gating adjustment module and a performance prediction module;

[0023] S32, the multi-modal embedding module respectively maps the standardized process data set according to temperature modalities, speed modalities, mechanical modalities, flow modalities and displacement modalities to independent embedding channels, and performs channel encoding, modality identification encoding and time index encoding operations;

[0024] The channel encoding assigns a channel number to each modality, converts the channel number into a channel representation vector through one-hot encoding, the modality identification encoding presets an independent modality semantic embedding vector for each type of modality, and the time index encoding generates a position embedding vector using a fixed sine-cosine position encoding method;

[0025] The channel representation vector, the modality semantic embedding vector and the position embedding vector are spliced with the standardized process data of the corresponding time step in the feature dimension to generate a multi-modal embedding feature tensor;

[0026] S33, the physical bias injection module constructs a physical bias tensor according to process parameters and manufacturing equipment state parameters, the physical bias tensor comprising a thermal diffusion bias component, a mechanical coupling bias component, a rotational speed vibration coupling bias component, a fluid disturbance bias component and a displacement error bias component;

[0027] The physical bias tensor and the multi-modal embedding feature tensor are weighted and superimposed in the channel dimension to generate a bias-enhanced state tensor;

[0028] S34, the stage controller module divides the bias-enhanced state tensor in the time step according to the set stage recognition rule, and constructs a process stage guide vector and a process stage index, the process stage comprising a pre-processing stage, a stable processing stage and an end unloading stage, the stage guide vector using one-hot encoding form to encode the pre-processing stage, the stable processing stage and the end unloading stage;

[0029] The bias-enhanced state tensor is classified according to the process stage index to generate a pre-processing tensor, a stable processing tensor and an end unloading tensor;

[0030] S35, the state segmentation modeling module comprises a pre-processing modeling branch, a stable processing modeling branch and an end unloading modeling branch, respectively receiving the pre-processing tensor, the stable processing tensor and the end unloading tensor;

[0031] The pre-processing modeling branch performs a perturbation attention allocation operation on the pre-processing tensor through one-dimensional convolution and a Softmax function, obtains a perturbation attention score, and performs time step by time step feature weighting fusion on the perturbation attention score and the pre-processing tensor to generate a perturbation guide tensor.

[0032] The perturbation guide tensor extracts perturbation response features of the pre-processing stage through an SRU within a state recursion window of a set length, to generate a pre-processing feature tensor.

[0033] The stable processing modeling branch performs time series noise reduction and local feature extraction operations on the stable processing tensor through a multi-head separable convolution, which includes multiple parallel deep separable one-dimensional convolution branches. Each deep separable one-dimensional convolution branch performs intra-channel time series feature extraction through deep convolution and inter-channel modal interaction fusion operation through point-by-point convolution.

[0034] The output feature tensors of all branches are spliced in the channel dimension and mapped to the target embedding dimension through a linear mapping layer to generate a multi-scale local embedding tensor.

[0035] The multi-scale local embedding tensor is subjected to non-linear mapping and feature compression operations through a double-layer feedforward network structure to generate a stable processing feature tensor.

[0036] The end unloading modeling branch performs time series state modeling operations on the end unloading tensor through a standard state space convolution unit to generate an initial unloading feature tensor.

[0037] The initial unloading feature tensor and the end unloading tensor are connected in time steps, and are compressed to the target embedding dimension through linear mapping to generate an end unloading feature tensor.

[0038] S36, the gating adjustment module performs embedding dimension normalization operation on the pre-processing feature tensor, the stable processing feature tensor and the end unloading feature tensor, and the corresponding gating weight vector of the gating unit,

[0039] The pre-processing feature tensor, the stable processing feature tensor and the end unloading feature tensor are element-level weighted with the corresponding gating weight vector to generate a pre-processing adjustment tensor, a stable processing adjustment tensor and an end unloading adjustment tensor, respectively.

[0040] The pre-processing adjustment tensor, the stable processing adjustment tensor and the end unloading adjustment tensor are spliced in the channel dimension and input to a feedforward mapping network to generate a fusion state feature tensor.

[0041] S37, input the fusion state feature tensor to a performance prediction module, the performance prediction module generates a no-load back-EMF deviation through a double-layer feedforward network structure and a ReLU activation function, generates a main shaft imbalance vector amplitude through a GRU, and generates an assembly surface flatness error through one-dimensional convolution and maximum pooling operation;

[0042] The no-load back-EMF deviation, the main shaft imbalance vector amplitude and the assembly surface flatness error constitute a manufacturing process performance prediction result.

[0043] Optionally, the stage recognition rule in the step S35 is:

[0044] S351, according to the main shaft speed parameter, the cutting force parameter and the cooling liquid flow parameter, the main shaft speed change rate, the cutting force standard deviation change rate and the cooling liquid flow change rate are calculated respectively;

[0045] S352, the rising threshold, the stable interval threshold and the falling threshold of the main shaft speed change rate, the cutting force standard deviation change rate and the cooling liquid flow change rate are set:

[0046] When the main shaft speed change rate is greater than the rising threshold, it is determined that the current time step is a pre-machining stage, when the main shaft speed change rate, the cutting force standard deviation change rate and the cooling liquid flow change rate are all less than the stable interval threshold, and the continuous state exceeds the set time window, it is determined that the current time step is a stable machining stage, and when the main shaft speed change rate and the cutting force standard deviation change rate are less than the falling threshold, it is determined that the current time step is an end unloading stage;

[0047] S353, when there is a time step that cannot be determined, the stage controller module performs process stage interpolation processing according to the process stage index of the adjacent time step.

[0048] Optionally, the step S4 specifically includes:

[0049] S41, the multi-objective process optimization function takes the no-load back-EMF deviation, the main shaft imbalance vector amplitude and the assembly surface flatness error as target items:

[0050] S42, a hybrid crossover NSGA-II algorithm is used to search and control the convergence of the multi-objective optimization function, and a process parameter optimization solution set is generated.

[0051] Optionally, the hybrid crossover NSGA-II algorithm in the step S42 specifically includes:

[0052] S421, initialize a population and set the maximum number of iteration rounds, the population includes N individuals, and each individual represents a set of process parameter vectors;

[0053] S422、In each iteration process, the multi-objective process optimization function value corresponding to each individual in the current population is calculated respectively, and the current population is sorted according to the order from low to high of the non-dominated ranking and from large to small of the crowding distance, and the Top-K individuals are selected to form the parent population;

[0054] S423、The individuals of the parent population are crossed and recombined using the simulated binary crossover operator, and the differential mutation operation is performed on the offspring individuals after crossing to generate a child population of the same size;

[0055] S424、The parent population and the child population are merged to form a joint population, and the Top-K individuals selected by the non-dominated sorting and the crowding distance are selected again to form a new round of population;

[0056] S425、If the current iteration round does not reach the maximum iteration number, the updated new round of population is taken as the parent population for continuous iteration, otherwise the solution set with the first non-dominated level is extracted from the final joint population, and is taken as the process parameter optimization solution set.

[0057] Optionally, the step S5 specifically comprises:

[0058] S51、The optimal process parameter scheme is a process parameter scheme with the minimum multi-objective performance deviation metric value;

[0059] S52、The multi-objective performance deviation metric value is obtained according to the weighted values of the no-load back electromotive force deviation, the main shaft imbalance vector amplitude and the assembly plane flatness error;

[0060] S53、The optimal process parameter scheme is issued to the motor parts manufacturing, and a manufacturing execution instruction is triggered, the manufacturing execution instruction includes generating a machining path, setting a machining speed, adjusting a cooling parameter and configuring an assembly tension.

[0061] Optionally, the step S6 specifically comprises:

[0062] S61、The feedback data and the manufacturing process performance prediction result are compared item by item to generate a feedback error vector, the feedback data includes the no-load back electromotive force value, the main shaft vibration vector and the assembly surface topographic parameter;

[0063] S62、The feedback error vector is input to the physical bias injection module of the improved Mamba network, is spliced with the original physical bias tensor to generate an extended bias tensor, and the bias enhancement state tensor is updated, and the modeling and optimization iteration process is continued.

[0064] The beneficial effects of the application are:

[0065] Firstly, the application collects multiple-source process parameters such as temperature parameters, main shaft speed parameters, cutting force parameters, feed speed parameters, machining pressure parameters, cooling liquid flow parameters, equipment vibration parameters, displacement deviation parameters and clamping state parameters by deploying multiple types of sensors at key positions of motor part manufacturing equipment, and constructs a standardized process data set by modal grouping, time alignment and Z-score standardization processing, thereby providing a data basis for high-quality input of the digital twin model.

[0066] Secondly, the application introduces an improved Mamba network as a core modeling component in the digital twin model of the motor part manufacturing process, which includes a multi-modal embedding module, a physical bias injection module, a stage controller module, a state segmented modeling module, a gating adjustment module and a performance prediction module in the improved Mamba network structure, and can realize segmented modeling and performance prediction of stages such as pre-processing, stable processing and end unloading, thereby effectively improving the refinement degree of manufacturing state modeling and the adaptability to complex process stages.

[0067] In addition, based on the manufacturing process performance prediction results, a multi-objective process optimization function is constructed with no-load back electromotive force deviation, main shaft imbalance vector amplitude and assembly surface flatness error as target items, and a hybrid crossover NSGA-II algorithm is introduced to search for a process parameter solution set, and after selecting the optimal process parameter scheme by combining multi-objective performance deviation measurement, manufacturing application is performed, and then feedback data is collected and a feedback error vector is generated in the manufacturing execution stage, and online optimization iteration is realized by injecting the physical bias injection module of the improved Mamba network, thereby constructing a complete process-performance-feedback closed loop mechanism.

[0068] In summary, the application realizes dynamic modeling and closed-loop optimization between process parameters and performance indicators in the motor part manufacturing process, improves the real-time performance and accuracy of process optimization, and significantly enhances the stability of the manufacturing process and the performance consistency of the parts. BRIEF DESCRIPTION OF DRAWINGS

[0069] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, together with the embodiments of the application, to explain the application, and do not constitute a limitation on the application. In the drawings:

[0070] Fig. 1 is a motor part manufacturing process optimization method based on digital twin proposed by the application;

[0071] Fig. 2 is an improved Mamba network modeling flowchart in a motor part manufacturing process optimization method based on digital twin proposed by the application;

[0072] Fig. 3Is a kind of mixed cross NSGA-II optimization flow chart in the motor parts manufacturing process optimization method based on digital twinning proposed in the application. DETAILED DESCRIPTION

[0073] The application will now be described in further detail with reference to the drawings. These drawings show only the basic structure of the application and therefore only show the components relevant to the application.

[0074] Reference Figs. 1-3 A motor parts manufacturing process optimization method based on digital twinning, comprising the following steps:

[0075] S1, collect multi-source process parameters in motor parts manufacturing process;

[0076] S2, process data preprocessing is carried out to multi-source process parameters, and standardization process data set is generated;

[0077] S3, the standardization process data set is executed manufacturing state evolution modeling and motor parts performance evaluation through improved Mamba network, and manufacturing process performance prediction result is output;

[0078] S4, based on manufacturing process performance prediction result, construct multi-objective process optimization function, and generate process parameter optimization solution set using mixed cross NSGA-II algorithm;

[0079] S5, select optimal process parameter scheme from process parameter optimization solution set and execute manufacturing application;

[0080] S6, in the execution process, feedback data of target process execution stage is collected in real time, and deviation comparison is carried out between feedback data and manufacturing process performance prediction result, feedback error vector is generated and input to improved Mamba network for online optimization iteration;

[0081] S7, manufacturing process performance prediction result, optimal process parameter scheme, feedback error vector and online optimization iteration result are visualized in graphical form, and manufacturing process visualization interface is constructed.

[0082] The application constructs a motor part manufacturing process optimization flow based on digital twinning. First, multi-source process parameters in the manufacturing process are collected, and a standardized data set is generated through preprocessing. Then, the manufacturing state is modeled based on an improved Mamba network to predict key performance indicators. On this basis, a multi-objective optimization function is constructed, and a process parameter optimization scheme is generated by combining a hybrid crossover NSGA-II algorithm. After selecting the optimal parameter scheme to drive the actual manufacturing process, feedback data are collected in real time and compared with the predicted results to generate an error vector for online model optimization. Finally, manufacturing performance prediction, parameter optimization results and feedback iteration information are dynamically displayed through a visual interface to form a manufacturing digital twinning closed-loop system covering the whole process of "perception - modeling - optimization - feedback - display".

[0083] In the embodiment, the step S1 specifically comprises:

[0084] S11, temperature sensors, speed sensors, force sensors, flow sensors, vibration sensors and displacement sensors are deployed in the spindle assembly, clamp system, tool path area, cooling system and workpiece machining area of the motor part manufacturing equipment, and a unified sampling period is set to collect multi-source process parameters in the motor part manufacturing process;

[0085] S12, the multi-source process parameters include temperature parameters, spindle speed parameters, cutting force parameters, feed speed parameters, machining pressure parameters, cooling liquid flow parameters, equipment vibration parameters, displacement deviation parameters and clamping state parameters.

[0086] In the embodiment, the process data preprocessing specifically comprises:

[0087] S21, the multi-source process parameters are grouped by modalities, including temperature modality, speed modality, mechanical modality, flow modality and displacement modality; wherein the temperature modality includes temperature parameters; the speed modality includes spindle speed parameters and feed speed parameters; the mechanical modality includes cutting force parameters, machining pressure parameters and clamping state parameters; the flow modality includes cooling liquid flow parameters; and the displacement modality includes equipment vibration parameters and displacement deviation parameters;

[0088] S22, each type of process parameter is respectively given a channel number, a task number and a time index to generate multi-source process parameters with structured index information;

[0089] S23, the multi-source process parameters with missing values are filled by weighted moving average method, the multi-source process parameters with redundancy or logical conflict are removed by time stamp window comparison method, and the multi-source process parameters deviating from the statistical range are modified by modal threshold method.

[0090] S24, time aligning the multi-source process parameters according to a set sampling period;

[0091] S25, normalizing the time-aligned multi-source process parameters by a Z-score standardization method, and performing channel alignment, dimension pruning and feature padding operations on data of each channel to encapsulate a standardized process dataset.

[0092] In the embodiment, the step S3 specifically comprises:

[0093] S31, inputting the standardized process dataset into an improved Mamba network, the improved Mamba network comprising a multi-modal embedding module, a physical bias injection module, a stage controller module, a state segmentation modeling module, a gating adjustment module and a performance prediction module;

[0094] S32, the multi-modal embedding module respectively maps the standardized process dataset to independent embedding channels according to temperature modalities, speed modalities, mechanical modalities, flow modalities and displacement modalities, and performs channel encoding, modality identification encoding and time index encoding operations;

[0095] The channel encoding assigns a channel number to each modality, converts the channel number into a channel representation vector by one-hot encoding, the modality identification encoding presets an independent modality semantic embedding vector for each type of modality, and the time index encoding generates a position embedding vector by using a fixed sine-cosine position encoding method;

[0096] The channel representation vector, the modality semantic embedding vector and the position embedding vector are spliced with the standardized process data of the corresponding time step in the feature dimension to generate a multi-modal embedding feature tensor;

[0097] For example, for the finishing process of a certain type of motor rotor parts, 10 batches of manufacturing samples are collected, each batch of samples contains 100 time step processing state data, and is divided into 5 modality channels of temperature modalities, speed modalities, mechanical modalities, flow modalities and displacement modalities. In each time step, a channel representation vector with a dimension of 5 is generated by one-hot encoding, a preset modality semantic embedding vector with a dimension of 8 is generated, a time index embedding vector with a dimension of 16 is generated by sine-cosine position encoding, and is spliced with a standardized process parameter vector with a dimension of 24 in the feature dimension to form a feature input vector under each modality at each time step. The total dimension after splicing is 53. The finally constructed multi-modal embedding feature tensor has a shape of 10x100x5x53;

[0098] S33, the physical bias injection module constructs a physical bias tensor according to process parameters and manufacturing equipment state parameters, the physical bias tensor includes a thermal diffusion bias component, a mechanical coupling bias component, a rotational speed vibration coupling bias component, a fluid disturbance bias component and a displacement error bias component;

[0099] The physical bias tensor and the multi-modal embedded feature tensor are weighted and superimposed in the channel dimension to generate a bias enhanced state tensor;

[0100] S34, the stage controller module divides the bias enhanced state tensor into process stages according to the set stage recognition rule, and constructs a process stage guide vector and a process stage index, the process stage includes a pre-processing stage, a stable processing stage and an end unloading stage, the stage guide vector encodes the pre-processing stage, the stable processing stage and the end unloading stage in the form of one-hot encoding;

[0101] The bias enhanced state tensor is classified according to the process stage index to generate a pre-processing tensor, a stable processing tensor and an end unloading tensor;

[0102] S35, the state segmentation modeling module includes a pre-processing modeling branch, a stable processing modeling branch and an end unloading modeling branch, which respectively receive the pre-processing tensor, the stable processing tensor and the end unloading tensor;

[0103] The pre-processing modeling branch performs perturbation attention allocation operation on the pre-processing tensor through one-dimensional convolution and Softmax function, obtains perturbation attention score, and performs time step feature weighted fusion on the perturbation attention score and the pre-processing tensor to generate a perturbation guide tensor;

[0104] The perturbation guide tensor extracts the perturbation response feature of the pre-processing stage in the state recursive window of the set length through SRU to generate a pre-processing feature tensor;

[0105] The stable processing modeling branch performs time series noise reduction and local feature extraction operation on the stable processing tensor through multi-head separable convolution, the multi-head separable convolution includes a plurality of parallel depth separable one-dimensional convolution branches, each depth separable one-dimensional convolution branch performs intra-channel time series feature extraction through depth convolution and inter-channel modal interaction fusion operation through point-by-point convolution;

[0106] The output feature tensors of all branches are spliced in the channel dimension, and mapped to the target embedding dimension through the linear mapping layer to generate a multi-scale local embedding tensor;

[0107] The multi-scale local embedding tensor is subjected to nonlinear mapping and feature compression operation through a double-layer feedforward network structure to generate a stable processing feature tensor;

[0108] The end unloading modeling branch performs a time step state modeling operation on the end unloading tensor through a standard state space convolution unit to generate an initial unloading feature tensor;

[0109] The initial unloading feature tensor is connected with the end unloading tensor in a time step by residual connection, and is compressed to a target embedding dimension through a linear mapping to generate an end unloading feature tensor;

[0110] S36, the gating adjustment module performs embedding dimension normalization operation on the pre-processing feature tensor, the stable processing feature tensor and the end unloading feature tensor, and performs element-level weighting on the pre-processing feature tensor, the stable processing feature tensor and the end unloading feature tensor through the corresponding gating weight vector,

[0111] The pre-processing feature tensor, the stable processing feature tensor and the end unloading feature tensor are respectively generated by the pre-processing adjustment tensor, the stable processing adjustment tensor and the end unloading adjustment tensor;

[0112] The pre-processing adjustment tensor, the stable processing adjustment tensor and the end unloading adjustment tensor are spliced according to the channel dimension, and are input into the feedforward mapping network to generate a fusion state feature tensor;

[0113] S37, the fusion state feature tensor is input into the performance prediction module, and the performance prediction module generates the no-load back electromotive force deviation through a double-layer feedforward network structure and a ReLU activation function, generates the main shaft imbalance vector amplitude through a GRU, and generates the assembly plane flatness error through a one-dimensional convolution and a maximum pooling operation;

[0114] The no-load back electromotive force deviation, the main shaft imbalance vector amplitude and the assembly plane flatness error constitute the manufacturing process performance prediction result.

[0115] In the embodiment, the stage recognition rule in step S35 is:

[0116] S351, according to the main shaft speed parameter, the cutting force parameter and the cooling liquid flow parameter, the main shaft speed change rate, the cutting force standard deviation change rate and the cooling liquid flow change rate are calculated respectively;

[0117] S352, the rising threshold, the stable interval threshold and the falling threshold of the main shaft speed change rate, the cutting force standard deviation change rate and the cooling liquid flow change rate are set:

[0118] When the main shaft speed change rate is greater than the rising threshold, the current time step is determined to be the pre-processing stage; when the main shaft speed change rate, the cutting force standard deviation change rate and the cooling liquid flow change rate are all less than the stable interval threshold, and the continuous state exceeds the set time window, the current time step is determined to be the stable processing stage; when the main shaft speed change rate and the cutting force standard deviation change rate are less than the falling threshold, the current time step is determined to be the end unloading stage;

[0119] S353、When there is a time step that cannot be determined, the stage controller module performs process stage interpolation processing according to the process stage indexes of the adjacent time steps.

[0120] In this embodiment, the step S4 specifically includes:

[0121] S41、The multi-objective process optimization function F(x) takes the no-load back-EMF deviation, the spindle unbalance vector amplitude and the assembly plane flatness error as the objective terms, and specifically:

[0122] Set the process parameter vector , the no-load back-EMF deviation is denoted as , the spindle unbalance vector amplitude is denoted as , and the assembly plane flatness error is denoted as , and the multi-objective process optimization function is constructed as:

[0123] ;

[0124] S42, a hybrid crossover NSGA-II algorithm is used to search and control the convergence of the multi-objective optimization function F(x) to generate an optimized solution set of the process parameters.

[0125] In this embodiment, the hybrid crossover NSGA-II algorithm in the step S42 specifically includes:

[0126] S421, initialize a population and set the maximum number of iteration rounds, the population includes N individuals, and each individual represents a set of process parameter vectors;

[0127] S422, in each iteration round, the multi-objective process optimization function value corresponding to each individual in the current population is calculated, and the current population is sorted according to the order from low to high of the non-dominated sorting level and from large to small of the crowding distance, and the Top-K individuals are selected to form the parent population;

[0128] The non-dominated sorting is as follows: in the multi-objective process optimization function F(x), for individuals and , if there exists any objective term satisfying , and at least one objective term satisfies , then it is determined that the individual is non-dominated to the individual , and all individuals are divided into several non-dominated levels based on the non-dominated relationship, wherein the first level contains all the solutions that are not dominated by any individual, the second level contains the solutions that are only dominated by the first level, and the non-dominated solution set is formed in this way.

[0129] The crowding distance is calculated by calculating the normalized distance between individuals for each objective item in the multi-objective process optimization function and summing up the normalized distances to obtain the crowding distance of each individual in each non-dominated level;

[0130] S423, cross-recombine the individuals of the parent population using the simulated binary crossover operator, and perform differential mutation operation on the cross-recombined offspring individuals to generate an offspring population of the same size;

[0131] S424, merge the parent population and the offspring population to form a joint population, and select the Top-K individuals from the joint population by non-dominated sorting and crowding distance to form a new population for the next round of iteration;

[0132] S425, if the current iteration round does not reach the maximum number of iterations, the updated new population is used as the parent population for further iteration, otherwise, the solution set with the first non-dominated level is extracted from the final joint population, and is used as the process parameter optimization solution set.

[0133] In the embodiment, the step S5 specifically includes:

[0134] S51, the optimal process parameter scheme is a process parameter scheme with the minimum multi-objective performance deviation metric value;

[0135] S52, the multi-objective performance deviation metric value is obtained according to the weighted values of the no-load back electromotive force deviation, the main shaft imbalance vector amplitude and the assembly plane flatness error;

[0136] S53, the optimal process parameter scheme is sent to the motor part manufacturing, and a manufacturing execution instruction is triggered, the manufacturing execution instruction includes generating a machining path, setting a machining speed, adjusting a cooling parameter and configuring an assembly tension.

[0137] In the embodiment, the step S6 specifically includes:

[0138] S61, the feedback data and the manufacturing process performance prediction result are compared item by item to generate a feedback error vector, the feedback data including the no-load back electromotive force value, the main shaft vibration vector and the assembly surface topographic parameter;

[0139] S62, the feedback error vector is input into the physical bias injection module of the improved Mamba network, is spliced with the original physical bias tensor to generate an extended bias tensor, and the bias enhancement state tensor is updated, and the modeling and optimization iteration process is continued.

[0140] Example 1

[0141] To verify the feasibility of the application in implementation, the application is applied to the rotor machining process optimization task of a certain high-performance motor manufacturing enterprise. The enterprise mainly produces permanent magnet synchronous motors for new energy vehicle drive systems, and the rotor structure is complex and the manufacturing precision is high, especially in dynamic balance, electromagnetic performance and assembly precision, etc. In the existing manufacturing process, process parameter adjustment mainly depends on manual experience, and the manufacturing process lacks real-time feedback mechanism, resulting in performance fluctuation problems in some products, which restricts product consistency and yield.

[0142] In the implementation process, first, deploy multiple types of industrial sensors in the enterprise's numerical control machining production line, including temperature sensors, force sensors, vibration sensors, flow meters and displacement sensors, etc. Cover key nodes such as spindle assembly, cooling system, tool path and fixture area, etc. Collect real-time process parameters at a frequency of 10Hz, and upload them to the edge gateway for data aggregation. Through the process data preprocessing module, the collected temperature, speed, cutting force, cooling flow, vibration and displacement, etc. Raw process parameters are grouped, time-aligned, outlier repaired and standardized, forming a standardized process data set. Then, input the standardized process data set into the digital twin model constructed based on the improved Mamba network.

[0143] In practical application, the improved Mamba network models the state evolution and performance prediction of the motor rotor machining process, and outputs the no-load back-EMF deviation, spindle imbalance vector amplitude and assembly plane flatness error. Based on the prediction results, a multi-objective process optimization function is constructed, and a process optimization solution set containing 200 parameter solutions is generated using a hybrid crossover NSGA-II algorithm. After process deviation measurement and scoring, the optimal solution is selected and executed. During execution, real-time device response logs and performance feedback data are collected and compared with the prediction results for error comparison, and the improved Mamba network is re-injected to complete model adaptive optimization iteration.

[0144] To evaluate the actual effect of the application, the application method is compared with the traditional experience parameter method and the standard LSTM optimization method, and evaluated from six dimensions of prediction accuracy, optimization convergence speed, performance consistency, yield, parameter stability and feedback response time. The experimental results are shown in Table 1.

[0145] Table 1 Performance comparison table of the application and the comparison method

[0146]

[0147] From the data in Table 1, it can be seen that the method of the application is significantly better than the traditional empirical parameter method and the standard LSTM optimization method in many key performance indicators. In terms of prediction accuracy, the average absolute percentage error of the method of the application is only 3.85%, which is decreased by 7.87 percentage points compared with the traditional empirical method and decreased by 3.51 percentage points compared with the standard LSTM method, significantly improving the predictability of manufacturing performance. In terms of optimization convergence efficiency, the application only needs 51 iterations to converge to the optimal solution, which is much lower than the 143 iterations of the traditional method and the 89 iterations of the standard LSTM method, indicating that the optimization strategy has faster convergence speed and stronger search efficiency. In terms of dynamic balance performance consistency, the application controls the variance of the main shaft imbalance vector amplitude to 0.29g·mm, effectively reducing the vibration risk and product deviation in the manufacturing process, and improving the processing stability.

[0148] In addition, in terms of yield, the method of the application reaches 96.7%, which is increased by 9.4 percentage points compared with the traditional empirical parameter method and increased by 4.9 percentage points compared with the standard LSTM optimization method, reflecting the effective integration of manufacturing consistency and process intelligent control. In terms of process parameter fluctuation control, the application controls the process parameter fluctuation amplitude to ±3.1%, which is much lower than ±12.5% of the traditional method and ±8.3% of the standard LSTM method, significantly enhancing the stability and controllability of the process parameters. In terms of performance feedback response capability, the application compresses the response delay to 3.9 seconds through the construction of feedback closed loop and model online iteration mechanism, realizes the rapid response to abnormal working conditions and performance drift, and meets the high requirements of real-time in intelligent manufacturing scenarios.

[0149] The application shows significant advantages in manufacturing precision, optimization efficiency, consistency control and feedback response capability, fully indicating the wide applicability and engineering promotion value of the application in the scene of high-precision and high-stability part manufacturing.

[0150] The above is only the preferred specific embodiment of the application, but the protection scope of the application is not limited thereto, any person skilled in the art can make equivalent replacement or change according to the technical solution and inventive concept of the application within the technical range disclosed by the application, which should be covered within the protection scope of the application.

Claims

1. A method for optimizing motor component manufacturing process based on digital twin, characterized in that, The method comprises the following steps: S1, collecting multi-source process parameters in the manufacturing process of motor parts; S2, preprocessing the multi-source process parameters to generate a standardized process data set; S3, performing manufacturing state evolution modeling and motor part performance evaluation on the standardized process data set through an improved Mamba network, and outputting manufacturing process performance prediction results; Wherein, the step S3 specifically comprises: S31, inputting the standardized process data set into the improved Mamba network, the improved Mamba network comprising a multi-modal embedding module, a physical bias injection module, a stage controller module, a state segmentation modeling module, a gating adjustment module and a performance prediction module; S32, the multi-modal embedding module maps the standardized process data set to independent embedding channels according to temperature mode, speed mode, mechanical mode, flow mode and displacement mode respectively, and performs channel coding, mode identification coding and time index coding operations; The channel coding assigns a channel number to each mode, converts the channel number to a channel representation vector through one-hot coding, the mode identification coding presets an independent mode semantic embedding vector for each type of mode, and the time index coding generates a position embedding vector using fixed sine-cosine position encoding; The channel representation vector, the mode semantic embedding vector and the position embedding vector are spliced with the standardized process data of the corresponding time step in the feature dimension to generate a multi-modal embedding feature tensor; S33, the physical bias injection module constructs a physical bias tensor according to the process parameters and the manufacturing equipment state parameters, the physical bias tensor comprising a thermal diffusion bias component, a mechanical coupling bias component, a speed vibration coupling bias component, a fluid disturbance bias component and a displacement error bias component; The physical bias tensor and the multi-modal embedding feature tensor are weighted and superimposed according to the channel dimension to generate a bias-enhanced state tensor; S34, the stage controller module divides the bias-enhanced state tensor into process stages according to the set stage recognition rules, and constructs a process stage guide vector and a process stage index, the process stages including a pre-processing stage, a stable processing stage and an end unloading stage, and the stage guide vector encodes the pre-processing stage, the stable processing stage and the end unloading stage in the form of one-hot encoding; The bias-enhanced state tensor is classified according to the process stage index to generate a pre-processing tensor, a stable processing tensor and an end unloading tensor; S35, the state segmentation modeling module comprises a pre-processing modeling branch, a stable processing modeling branch and an end unloading modeling branch, which respectively receive the pre-processing tensor, the stable processing tensor and the end unloading tensor; The pre-processing modeling branch performs perturbation attention allocation operation on the pre-processing tensor through one-dimensional convolution and Softmax function to obtain perturbation attention score, and performs feature weighted fusion between the perturbation attention score and the pre-processing tensor step by step to generate a perturbation guide tensor; The perturbation guide tensor extracts the perturbation response feature of the pre-processing stage in the state recursive window of the set length through SRU to generate a pre-processing feature tensor; The stable machining modeling branch performs time series denoising and local feature extraction operations on the stable machining tensor through multi-head separable convolution, the multi-head separable convolution includes multiple parallel deep separable one-dimensional convolution branches, each deep separable one-dimensional convolution branch performs intra-channel time series feature extraction through deep convolution and inter-channel modal interaction fusion operation through point-by-point convolution; The output feature tensors of all branches are spliced in the channel dimension and mapped to the target embedding dimension through a linear mapping layer to generate a multi-scale local embedding tensor; The multi-scale local embedding tensor is subjected to nonlinear mapping and feature compression operations through a double-layer feedforward network structure to generate a stable machining feature tensor; The end unloading modeling branch performs time series state modeling operations on the end unloading tensor through a standard state space convolution unit to generate an initial unloading feature tensor; The initial unloading feature tensor is connected with the end unloading tensor step by step, and is compressed to the target embedding dimension through linear mapping to generate the end unloading feature tensor; S36, the gating adjustment module performs embedding dimension normalization operation on the pre-machining feature tensor, the stable machining feature tensor and the end unloading feature tensor, and through the gating weight vector of the corresponding gating unit, The pre-machining feature tensor, the stable machining feature tensor and the end unloading feature tensor are subjected to element-level weighting with the corresponding gating weight vector to generate a pre-machining adjustment tensor, a stable machining adjustment tensor and an end unloading adjustment tensor, respectively; The pre-machining adjustment tensor, the stable machining adjustment tensor and the end unloading adjustment tensor are spliced in the channel dimension and input to the feedforward mapping network to generate a fusion state feature tensor; S37, the fusion state feature tensor is input to the performance prediction module, the performance prediction module generates the no-load back electromotive force deviation through a double-layer feedforward network structure and a ReLU activation function, generates the main shaft imbalance vector amplitude through a GRU, and generates the assembly plane flatness error through one-dimensional convolution and maximum pooling operation; The no-load back electromotive force deviation, the main shaft imbalance vector amplitude and the assembly plane flatness error constitute the manufacturing process performance prediction result; S4, a multi-objective process optimization function is constructed based on the manufacturing process performance prediction result, and a hybrid crossover NSGA-II algorithm is used to generate a process parameter optimization solution set; S5, the optimal process parameter scheme is selected from the process parameter optimization solution set and manufacturing application is performed; S6, during the execution process, feedback data of the target process execution stage is collected in real time, and the feedback data is compared with the manufacturing process performance prediction result to generate a feedback error vector which is input to the improved Mamba network for online optimization iteration; S7, the manufacturing process performance prediction result, the optimal process parameter scheme, the feedback error vector and the online optimization iteration result are visualized in a graphical form, and a manufacturing process visualization interface is constructed.

2. The method of claim 1, wherein the method is based on digital twin of motor parts manufacturing process optimization. The multi-source process parameters include temperature parameters, main shaft speed parameters, cutting force parameters, feed speed parameters, machining pressure parameters, cooling liquid flow parameters, equipment vibration parameters, displacement deviation parameters and clamping state parameters.

3. The method of claim 1, wherein the method is based on digital twin of motor parts manufacturing process optimization. The process data preprocessing specifically includes: S21, group the multi-source process parameters by modal, the modal includes temperature modal, speed modal, mechanical modal, flow modal and displacement modal; S22, assign a channel number, a task number and a time index to each type of process parameter to generate multi-source process parameters with structured index information; S23, interpolate and fill in the missing multi-source process parameters by weighted moving average method, remove invalid values of the multi-source process parameters with redundancy or logical conflict by time stamp window comparison method, and correct abnormal values of the multi-source process parameters deviating from the statistical range by modal threshold method; S24, time align the multi-source process parameters according to the set sampling period; S25, normalize the time-aligned multi-source process parameters, and perform channel alignment, dimension pruning and feature padding operations on the data of each channel to generate a standardized process data set.

4. The method of claim 1, wherein the method is based on digital twinning of motor parts manufacturing process optimization. The stage recognition rule in step S35 is: S351, calculate the spindle speed change rate, cutting force standard deviation change rate and cooling liquid flow change rate according to the spindle speed parameter, cutting force parameter and cooling liquid flow parameter respectively; S352, set the rising threshold, stable interval threshold and falling threshold of the spindle speed change rate, cutting force standard deviation change rate and cooling liquid flow change rate: When the spindle speed change rate is greater than the rising threshold, it is determined that the current time step is the pre-processing stage; when the spindle speed change rate, cutting force standard deviation change rate and cooling liquid flow change rate are all less than the stable interval threshold, and the continuous state exceeds the set time window, it is determined that the current time step is the stable processing stage; when the spindle speed change rate and cutting force standard deviation change rate are less than the falling threshold, it is determined that the current time step is the end unloading stage; S353, when there is a time step that cannot be determined, the stage controller module performs process stage interpolation processing according to the process stage index of the adjacent time step.

5. The method of claim 1, wherein the method is based on digital twin of motor parts manufacturing process optimization. The step S4 specifically includes: S41, the multi-objective process optimization function takes the no-load back electromotive force deviation, spindle imbalance vector amplitude and assembly surface flatness error as target items: S42, use hybrid crossover NSGA-II algorithm to search and control the convergence of the multi-objective optimization function to generate a process parameter optimization solution set.

6. The method of claim 5, wherein the method is based on digital twinning of the motor component manufacturing process. The hybrid crossover NSGA-II algorithm in step S42 specifically includes: S421, initialize the population and set the maximum number of iterations, the population includes N individuals, each individual represents a set of process parameter vectors; S422, in each iteration process, calculate the multi-objective process optimization function value corresponding to each individual in the current population, and sort the current population according to the non-dominated sorting level from low to high and the crowding distance from large to small, select the Top-K individuals to form the parent population; S423, use the simulated binary crossover operator to cross and recombine the individuals in the parent population, and perform difference mutation operation on the offspring individuals after crossing to generate an offspring population with the same size. S424, merge the parent population and the offspring population to form a joint population, and select the Top-K individuals with the best non-dominated ranking and crowding distance from the joint population to form a new population for the next iteration; S425, if the current iteration does not reach the maximum number of iterations, continue iteration with the updated new population as the parent population, otherwise extract the solution set with the first non-dominated level from the final joint population as the process parameter optimization solution set.

7. The method of claim 1, wherein the method is based on digital twinning of motor parts manufacturing process optimization. The step S5 specifically comprises: S51, the optimal process parameter scheme is the process parameter scheme with the minimum multi-objective performance deviation metric value; S52, the multi-objective performance deviation metric value is obtained according to the weighted values of the no-load back electromotive force deviation, the main shaft imbalance vector amplitude and the assembly plane flatness error; S53, the optimal process parameter scheme is sent to the motor part manufacturing, and a manufacturing execution instruction is triggered, which includes generating a machining path, setting a machining speed, adjusting a cooling parameter and configuring an assembly tension.

8. The method of claim 1, wherein the method is based on digital twinning of motor parts manufacturing process optimization. The step S6 specifically comprises: S61, compare the feedback data and the manufacturing process performance prediction result item by item to generate a feedback error vector, the feedback data including the no-load back electromotive force value, the main shaft vibration vector and the assembly plane geometric appearance parameter; S62, input the feedback error vector into the physical bias injection module of the improved Mamba network, splice the original physical bias tensor to generate an extended bias tensor, update the bias enhancement state tensor, and continue the modeling and optimization iteration process.

Images