Grinding parameter optimization method for full-automatic rough and fine grinding integrated machine based on particle swarm optimization

Abstract

The present application relates to the technical field of control parameter optimization, and in particular to a full-automatic rough and fine grinding integrated machine grinding parameter optimization method based on particle swarm optimization. The method comprises the following steps: obtaining device operation multi-modal data and performing multi-thread aggregation to obtain structured grinding field perception data; performing grinding parameter semantic abstraction on the structured grinding field perception data to obtain a grinding parameter intention nested mapping table; performing rough solution space boundary evolution on the grinding parameter intention nested mapping table to obtain a coarse-grained grinding parameter trajectory population; performing soft clustering bias learning on the coarse-grained grinding parameter trajectory population to obtain a high-dimensional grinding parameter candidate atlas; and performing parameter tensor loop optimization based on the high-dimensional grinding parameter candidate atlas to obtain a stable grinding parameter set. The present application helps to improve the precision and production efficiency of the device, and also reduces energy consumption, equipment wear and maintenance costs, thereby achieving higher benefits and sustainable development in industrial production.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of control parameter optimization, and in particular to a full-automatic rough and fine grinding integrated machine grinding parameter optimization method based on particle swarm optimization. BACKGROUND

[0002] As an intelligent equipment integrating rough grinding and fine grinding processes, the full-automatic rough and fine grinding integrated machine has advantages of high efficiency, high precision and process integration. In actual application process, reasonable configuration of grinding process parameters (including grinding speed, feed speed, grinding depth, grinding wheel linear speed, etc.) has a direct impact on workpiece surface quality, processing efficiency, equipment stability and energy consumption control indicators. In the prior art, the selection of grinding parameters depends on manual experience or rule-based decision system, which is difficult to effectively cope with complex working condition regulation and control problems under multi-objective and multi-constraint conditions. Some technical solutions introduce traditional optimization methods such as genetic algorithm and simulated annealing for parameter adjustment. However, due to the highly nonlinear, dynamic coupling and uncertainty characteristics of the grinding process, the traditional optimization methods have low search efficiency, are prone to local optimization, and are sensitive to initial population, which cannot meet the global optimization requirements in high-dimensional parameter space. SUMMARY

[0003] Therefore, it is necessary to provide a full-automatic rough and fine grinding integrated machine grinding parameter optimization method based on particle swarm optimization to solve at least one of the above technical problems.

[0004] To achieve the above-mentioned purpose, a full-automatic rough and fine grinding integrated machine grinding parameter optimization method based on particle swarm optimization comprises the following steps:

[0005] Step S1: Obtain device running multi-modal data, and perform edge-level preliminary processing on the device running multi-modal data to extract implicit physical disturbance features; perform multi-thread aggregation on the implicit physical disturbance features to obtain structured grinding field perception data;

[0006] Step S2: Perform multi-modal grinding behavior feature analysis on the structured grinding field perception data to obtain grinding behavior mode data; perform grinding parameter semantic abstraction according to the grinding behavior mode data to obtain a grinding parameter intention nested mapping table;

[0007] Step S3: Perform initialization of the basic particle swarm algorithm on the grinding parameter intention nested mapping table to obtain an initialized basic particle swarm; perform rough solution space boundary evolution based on the initialized basic particle swarm to obtain a coarse-grained grinding parameter trajectory group;

[0008] Step S4: Perform subgraph nested reconstruction on the coarse-grained grinding parameter trajectory group, and construct a heterogeneous trajectory graph encoding vector set; perform soft clustering bias learning on the heterogeneous trajectory graph encoding vector set to obtain a high-dimensional grinding parameter candidate graph.

[0009] Step S5: According to the high-dimensional grinding parameter candidate atlas, multi-dimensional working condition simulation is carried out, and parameter tensor loop optimization is carried out according to the multi-dimensional working condition simulation result, to obtain a stable grinding parameter set;

[0010] Step S6: The stable grinding parameter set is transmitted to the full-automatic rough and fine grinding integrated machine management platform to execute the control parameters, and iterative precision compensation mapping is carried out according to the real-time acquisition of the actual grinding deviation feedback data, to obtain an optimized grinding parameter set.

[0011] The present application effectively solves the problems of instability and low efficiency in the traditional grinding process through accurate data acquisition, disturbance identification and optimization adjustment. Firstly, in the original data acquisition stage, through the cooperative work of multi-modal sensors, the relevant information of the equipment, workpiece and environment state is comprehensively captured, providing a solid data foundation for subsequent analysis. By reasonably setting the time window and standard deviation threshold, the cleanliness and reliability of the data are guaranteed, the noise interference is excluded, and the accuracy of subsequent processing is ensured. In the disturbance identification stage, the accurate disturbance detection capability enables the early identification of abnormal fluctuations that may occur in the grinding process, providing a basis for subsequent precise adjustment and avoiding errors caused by external factors or equipment instability. In the behavior modeling and atlas generation process, by using center transfer learning and fuzzy clustering technology, the key features in the grinding process can be effectively extracted, and high-quality behavior atlases are generated. These behavior atlases provide detailed references for each stage of the grinding process, effectively guiding the optimization and adjustment of parameters, and ensuring the stability and precision of the grinding process. The set clustering parameters and clustering threshold help the system accurately divide different working conditions, making the atlas more consistent with actual operation, enabling quick response to working condition changes and real-time adjustment, thereby ensuring the adaptability and precision of the process. In the grinding parameter optimization stage, the particle swarm search method is used to intelligently adjust the grinding parameters, which can find the optimal processing parameters under different working conditions, avoiding the subjectivity and uncertainty of manual adjustment in traditional methods. This optimization method not only improves the grinding efficiency and reduces energy consumption, but also effectively reduces the product rejection rate and improves production quality. In addition, through the multi-thread aggregation and feedback compensation mechanism, the optimization system can adapt to changes in different working conditions in real time, keeping the grinding process in the optimal state at all times. In summary, through the comprehensive application of these technical means, the optimization technology greatly improves the stability, efficiency and intelligence level of the full-automatic rough and fine grinding integrated machine. Reasonable parameter setting and precise working condition feedback mechanism not only improve the precision and production efficiency of the equipment, but also reduce energy consumption, equipment wear and maintenance cost, thereby achieving higher benefits and sustainable development in industrial production. BRIEF DESCRIPTION OF DRAWINGS

[0012] Other features, objects, and advantages of the application will become more apparent from the following detailed description when read in connection with the following drawings:

[0013] Fig. 1 A schematic diagram of the step flow of the grinding parameter optimization method of the full-automatic rough and fine grinding integrated machine based on particle swarm optimization of the application is shown in the figure.

[0014] Fig. 2 A schematic diagram of the detailed step flow of step S1 in the application is shown in the figure.

[0015] Fig. 3 A schematic diagram of the detailed step flow of step S3 in the application is shown in the figure.

[0016] The implementation of the object of the application, functional features and advantages will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0017] The technical method of the application will be described clearly and completely below in combination with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

[0018] In addition, the accompanying drawings are only schematic illustrations of the application, and are not necessarily drawn to scale. The same reference numerals in the drawings represent the same or similar parts, and thus repeated descriptions thereof will be omitted. Some block diagrams shown in the drawings are functional entities, which do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in the form of software, or in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0019] It should be understood that although the terms "first", "second" and the like can be used herein to describe various elements, these elements should not be limited by these terms. These terms are only used to distinguish one element from another. For example, without departing from the scope of the exemplary embodiments, a first element can be referred to as a second element, and similarly a second element can be referred to as a first element. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0020] To achieve the above-mentioned object, please refer to Figs. 1 to 3 The application provides a grinding parameter optimization method of a full-automatic rough and fine grinding integrated machine based on particle swarm optimization, which comprises the following steps:

[0021] Step S1: Obtain device operation multi-modal data, and perform edge-level preliminary processing on the device operation multi-modal data to extract implicit physical disturbance features; perform multi-thread aggregation on the implicit physical disturbance features to obtain structured grinding field perception data;

[0022] In this embodiment, for the real-time running state of the full-automatic rough and fine grinding integrated machine, the thermal flow, current, vibration and acoustic spectrum multi-source data generated in a continuous 20-minute grinding period are collected, the sampling frequencies are set to be 10 Hz for thermal flow, 100 Hz for current, 1 kHz for vibration and 8 kHz for acoustic spectrum, and the first-order noise weakening and time sequence normalization processing are performed on the data of each channel. After preliminary processing, through parallel multi-thread structure, the feature point sequence highly related to the contact state change of the grinding wheel surface is extracted, including the thermal flow sudden rise interval, the current fluctuation segment, the vibration peak value slope mutation point and the abnormal frequency band of the acoustic spectrum. Align and fuse the above feature point sequences according to the time clue to form structured grinding field perception data based on the grinding wheel linear speed, which is convenient for subsequent behavior dynamic feature analysis.

[0023] Step S2: Perform multi-modal grinding behavior feature analysis on the structured grinding field perception data to obtain grinding behavior mode data; perform semantic abstraction of grinding parameters according to the grinding behavior mode data to obtain a grinding parameter intention nested mapping table;

[0024] In this embodiment, the structured grinding field perception data is imported into the time sequence recognition process, each group of data is processed by grinding behavior slicing labeling, the continuous behavior frame segment is divided according to every 4 seconds as a time window, and the critical change threshold is set to identify the contact arc disturbance features in the grinding zone. Taking the grinding wheel linear speed as the main shaft reference, map each modal disturbance response to the unified behavior time axis, and identify the high-impact behavior segments including abrasive particle breakage, shedding and local thermal burn in the contact area. Based on this, the semantic response mapping relationship between behavior and parameter group is established, and the grinding parameter intention nested mapping table is output, wherein each behavior segment is associated with more than three adjustable parameter dimensions, and the response weight is generated according to the frequency and amplitude change strength.

[0025] Step S3: Perform initialization of the basic particle swarm algorithm on the grinding parameter intention nested mapping table to obtain an initialized basic particle swarm; perform rough solution space boundary evolution based on the initialized basic particle swarm to obtain a rough granularity grinding parameter trajectory group;

[0026] In this embodiment, according to the grinding parameter intention nested mapping table, the initial particle number of the particle swarm is set to 30, the particle dimension is one-to-one corresponding to the effective adjustable dimension in the grinding parameter group (such as linear velocity, feed rate, grinding depth, etc.), and the behavior response weight is used as the initial position bias for parameter group initialization. During operation, the boundary relaxation coefficient of each round of evolution is set to 0.1, and the parameter update offset of each time is limited to not more than 20% of the local optimum of the last round. In the 10 rounds of preliminary evolution, the densest trajectory community is extracted to form a coarse-grained grinding parameter trajectory family, wherein the trajectory has a dynamic characteristic with a behavior response significance threshold exceeding 0.8, which is used to guide the construction and optimization of the next step trajectory structure.

[0027] Step S4: Subgraph nested reconstruction is performed on the coarse-grained grinding parameter trajectory family, and a heterogeneous trajectory graph coding vector set is constructed; soft clustering bias learning is performed on the heterogeneous trajectory graph coding vector set to obtain a high-dimensional grinding parameter candidate graph;

[0028] In this embodiment, the coarse-grained grinding parameter trajectory family is constructed into a subgraph nested structure according to the trajectory continuity and parameter change coupling characteristics, wherein each trajectory subgraph is constructed with a parameter change sequence as an edge set, and only the connection edge with an adjacent behavior response intensity greater than 0.05 is included in the graph. The node represents the parameter state at the discrete time point, and the maximum number of nodes in a single graph is set to not more than 300. The formed multi-subgraph set represents the behavior response network under different parameter evolution paths. On this basis, window sampling is used to perform structure coding on each subgraph, extract its joint statistical features in the time evolution, intensity response and frequency fluctuation three domains, and form a unified dimension original coding vector. In order to eliminate the imbalance caused by the different dimension of the feature, the zero mean unit variance standardization method is used for normalization processing: first, the sample mean and standard deviation of each feature dimension are calculated, and then the original value is mapped to the standardized value, so that all feature dimensions conform to the 0 mean value and 1 standard deviation distribution, thereby generating the coding vector set under the unified scale. The behavior bias vector constructed based on the historical working condition model is introduced as the initial membership degree guide, which is used to guide the fuzzy membership process to approach the true behavior distinguishing characteristics. In the clustering process, the target cluster number is set to 10, and the fuzzy weighted index is set to 2.1 to enhance the boundary fuzzy control ability in soft clustering. The iterative center adjustment mechanism is adopted, and the maximum offset amplitude of the class center is limited to not more than 0.05 in each round of update, so as to ensure the stability of the clustering convergence path. After clustering, the stability of the clustering result is analyzed, the center vector variance of each class is counted, and the unstable class with a variance greater than 0.08 is removed. Finally, the remaining clustering center is used as the representative of the grinding parameter semantic clustering, and the multi-membership structure graph is constructed by tracking the membership of each coding vector to multiple centers. Based on the multiple correlations between vectors, the grinding parameter semantic space with high confidence is presented.

[0029] Step S5: Multi-dimensional working condition simulation is performed according to the high-dimensional grinding parameter candidate atlas, and parameter tensor loop optimization is performed according to the multi-dimensional working condition simulation result, to obtain a stable grinding parameter set;

[0030] In this embodiment, the high-dimensional grinding parameter candidate atlas is imported into a multi-dimensional working condition response simulation framework, relying on a typical working condition library (such as silicon carbide workpieces, high-strength steel workpieces, ceramic materials, etc.) set in the full-automatic rough and fine grinding integrated machine, setting the material matching factor to 0.9, simulating and reconstructing the grinding response under different parameter combinations, and outputting four indexes of grinding efficiency, thermal load, energy consumption and surface roughness. The simulation output is constructed into a tensor structure, the response error is taken as the loop update condition, the error convergence threshold is set to 0.02, three rounds of feedback optimization are performed, and finally a stable grinding parameter set is generated, and its feasibility and consistency in the actual environment are verified.

[0031] Step S6: The stable grinding parameter set is transmitted to the full-automatic rough and fine grinding integrated machine management platform to execute the control parameters, and iterative precision compensation mapping is performed according to the actual grinding deviation feedback data obtained in real time, to obtain an optimized grinding parameter set.

[0032] In this embodiment, the stable grinding parameter set is sent to the device management controller of the full-automatic rough and fine grinding integrated machine through an industrial Ethernet interface, and the controller allocates devices to execute the grinding task. Real-time collection of device return execution current, voltage fluctuation, machining precision and thermal load data, and calculation of the deviation degree thereof from the preset working condition. The deviation feedback iteration period is set to 5 minutes, the offset type and amplitude in the current parameter execution are identified every period, and fine-grained adjustment is performed through a bidirectional parameter mapping mechanism, to output the optimized grinding parameter set after deviation compensation, to continuously improve the device running stability and machining quality consistency. The bidirectional parameter mapping mechanism is a parameter updating strategy combining forward and reverse adjustment paths, and its core purpose is to accurately compensate the grinding parameter deviation caused by working condition fluctuations on the basis of maintaining the stability of the current device running state. Specifically, it includes two directions of data loop: 1) Forward mapping refers to starting from the original stable parameter set, according to the preset working condition model and grinding task requirements, mapping these parameters into executable machine tool control instructions through the controller. This process corresponds to the path of "parameter issuing" or "driving" the device. 2) Reverse mapping is to judge the deviation between the feedback and the expected target after collecting actual feedback data (such as current, voltage, machining deviation, thermal load, etc.) during device operation, and then based on the deviation type and amplitude, the direction and amplitude of the current parameter adjustment are back calculated, which belongs to the adaptive compensation path based on feedback.

[0033] Optionally, step S1 is specifically:

[0034] Step S11: Obtain equipment running multi-modal data, wherein the equipment running multi-modal data comprises a grinding parameter group, a high-frequency vibration signal, a spindle current fluctuation record, a temperature rise thermal flow field image, and an accompanying acoustic spectrum signal;

[0035] In this embodiment, multi-modal data during the running of the full-automatic rough-fine grinding integrated machine is obtained through the multi-type sensor array deployed on the full-automatic rough-fine grinding integrated machine. The data comprises a grinding parameter group (covering a grinding depth, a linear speed, and a feed rate), a three-axis acceleration vibration signal with a high-speed sampling frequency of 20 kHz, a current real-time fluctuation curve of a spindle driving system, a temperature rise thermal flow field dynamic image sequence (a frame rate is not less than 30 frames per second) collected based on infrared thermal imaging, and an acoustic spectrum signal in a range of 0-10 kHz obtained by a coupled acoustic sensor. All acquisition channels are uniformly time-stamped and synchronized through an embedded data acquisition unit, ensuring the time consistency of the data, and are encoded and stored in a unified data packet format for subsequent decoupling and analysis operations.

[0036] Step S12: Perform channel-decoupling processing on the equipment running multi-modal data, and identify a physical field disturbance trend of the decoupling processing result to obtain physical field disturbance trend data;

[0037] In this embodiment, the obtained multi-modal data is decoupled in the physical domain respectively according to data channels. The vibration signal channel uses short-time Fourier transform to extract excitation energy in a frequency band; the current signal channel uses a sliding window to statistically calculate a root mean square of current fluctuation amplitude; the thermal flow image channel uses a local area temperature gradient change rate as a time domain feature; and the acoustic spectrum channel obtains sound pressure power values of each frequency band through 1 / 3 octave decomposition. Subsequently, on each channel feature sequence, a position point with a continuous change gradient greater than a threshold value (0.15 g / s in the vibration domain, 0.12 A / s in the current domain, and 1.2 °C / s in the thermal flow domain) in the time sequence is determined as a disturbance trend key point, and after integration, cross-channel physical field disturbance trend data is formed.

[0038] Step S13: Extract an implicit physical disturbance section according to the physical field disturbance trend data to obtain implicit physical disturbance primary annotation data;

[0039] In this embodiment, based on the constructed physical field disturbance trend data, disturbance threshold criteria are set according to different physical channel characteristics, and implicit physical disturbance section extraction is performed in sequence. Specifically, if the temperature rise change rate is greater than 1.5°C / s at three or more consecutive time points, and the vibration change amplitude reaches 0.2g, the interval is determined as a potential thermal-mechanical coupling disturbance; if the current peak change rate is more than 0.15A / s, and the sound pressure mean square increment difference between adjacent two windows is more than 10%, the interval is further marked as a composite disturbance section. All identified disturbance sections are recorded according to the start and end time, channel attribute and disturbance intensity, and are marked with disturbance level to form a primary disturbance annotation data set.

[0040] Step S14: The implicit physical disturbance primary annotation data is thread split, and time segment indexes are constructed for heat flow, vibration, current and sound spectrum signal segments respectively. The aggregation window step is set to 0.5s for thread merging operation to obtain a disturbance event structure vector set;

[0041] In this embodiment, the extracted primary disturbance annotation data is thread split according to the data channel type, and time segment indexes are established for heat flow image frame sequence, vibration acceleration record, current fluctuation record and sound spectrum density function curve. Each type of segment is intercepted as a segment window within 1.5 seconds before and after the disturbance occurrence time point, and the aggregation window step is set to 0.5 seconds for thread merging operation. During the thread merging process, if multiple channels have interference intervals in the same time period, the maximum disturbance level is taken as the main event, and the other channel segments are used to establish a composite structure. Finally, a disturbance event set in uniform structure vector format is output. Each disturbance event structure includes disturbance source, action channel, maximum change amplitude, duration and channel coordination marker.

[0042] Step S15: The disturbance event structure vector set is encoded for structure consistency, and the obtained device running timestamp, working condition identifier and part index information are fused to obtain structured grinding field perception data.

[0043] In this embodiment, the constructed disturbance event structure vector set is processed for structure consistency. First, the vectors are sorted and classified according to channel sequence and event intensity, and then the vectors are unified into an encoding form containing six fixed fields: channel type identifier, disturbance start time, disturbance duration, maximum disturbance intensity value, impact level score and interference frequency range. The machining task number, machine body working part number and unified timestamp information synchronized with the main control PLC system at the time of event occurrence are obtained from the device running control system, and are embedded into the vector structure. The data is reconstructed in time, space and semantic three levels, and finally the structured grinding field perception data is generated, which is used for subsequent behavior analysis and grinding behavior feature modeling.

[0044] Optionally, step S13 is specifically:

[0045] Step S131: Set the window length as 2s and the step length as 0.5s, segment the physical field disturbance trend data by the sliding window, and obtain the sliding segment disturbance trend data;

[0046] In this embodiment, the sliding window length is set as 2s and the step length is set as 0.5s, and the physical field disturbance trend data obtained during the operation of the device is segmented. Each sliding window rolls on the time axis with an interval of 0.5s, and data in the corresponding 2s window is extracted at each rolling. Specifically, for multi-channel data such as heat flow, vibration, current and acoustic spectrum, they are distinguished by synchronous time markers. The data of each sliding segment will be analyzed as independent disturbance trend data. These data contain the physical quantity changes and the corresponding disturbance performance of each channel in the time period, which ensures that further disturbance analysis and feature extraction can be performed for each segment in the subsequent steps.

[0047] Step S132: Calculate the disturbance intensity differential value of each sliding segment in the sliding segment disturbance trend data, and set the disturbance intensity mutation threshold as 1.75 to identify the initial segment of potential disturbance anomaly, and generate the disturbance segment preliminary screening data set;

[0048] In this embodiment, the disturbance intensity of each channel in each sliding segment in the obtained sliding segment disturbance trend data is calculated by differential calculation, and the change rate is particularly concerned. The disturbance intensity mutation threshold is set as 1.75, which is used as a standard to detect the disturbance intensity fluctuation of each segment. Once the disturbance intensity change mutation in the segment exceeds the threshold, it is considered that the segment has potential disturbance abnormal behavior, which is marked as a preliminary abnormal segment. In specific implementation, by calculating the change rate of the disturbance intensity in each time step (for example, the change rate of the current amplitude and the temperature gradient), once it exceeds the set threshold, it is considered that a significant disturbance change occurs in the region. All segments meeting this condition will be collected into a preliminary screening data set for subsequent disturbance behavior analysis and pattern recognition.

[0049] Step S133: Group the disturbance segment preliminary screening data set by modal channel type, respectively construct vibration frequency-energy graph, current fluctuation spectrum graph, heat flow intensity spectrum and acoustic spectrum density graph, and perform frequency domain entropy increment analysis on each group of graphs to screen the segments meeting the preset multi-channel synchronous entropy fluctuation condition, and generate multi-channel coupled disturbance candidate segment;

[0050] In this embodiment, the screened disturbance segments are further subdivided into different channel types for grouping. For vibration signals, a vibration frequency-energy graph is constructed to analyze the energy distribution of each frequency band using FFT; for current fluctuations, a current fluctuation spectrum graph is constructed to evaluate the frequency components of current fluctuations; for heat flow data, a heat flow intensity map is generated to reflect the intensity of local temperature changes; for acoustic spectrum data, an acoustic spectrum density graph is constructed to measure the distribution of sound wave signals in different frequency bands. Then, frequency domain entropy increment analysis is performed on each graph respectively to calculate the entropy value fluctuations in each segment. A preset multi-channel synchronous entropy fluctuation condition is set, when the entropy value changes of multiple channels exceed a certain fluctuation threshold at the same time, it is considered that the segment has significant disturbance behavior. All segments that meet this condition are collected as multi-channel coupled disturbance candidate segments for further analysis.

[0051] Step S134: co-occurrence pattern recognition is performed on the multi-channel coupled disturbance candidate segments to screen out isolated segments, and a modality consistency evaluation index greater than 0.6 is set to select high synergy interference component segments as the primary labeling segments of the hidden physical disturbance;

[0052] In this embodiment, based on the obtained multi-channel coupled disturbance candidate segments, co-occurrence pattern recognition is performed. By analyzing the time synchronization of disturbance events in different channels, segments with high synergy interference are identified, and segments without sufficient synergy or isolated disturbance are removed. In order to select segments with high synergy interference components, a modality consistency evaluation index greater than 0.6 is set, that is, only those segments with high consistency of vibration, heat flow, current and acoustic spectrum signals are identified as hidden physical disturbances. Here, the consistency evaluation is completed by calculating the similarity between the signals of each channel, and all segments that meet the conditions are marked as high synergy disturbance segments and further used as the primary labeling data of hidden physical disturbances.

[0053] Step S135: The primary labeling segments of the hidden physical disturbance are collected to obtain the primary labeling data of the hidden physical disturbance.

[0054] In this embodiment, the hidden physical disturbance segments marked as high synergy disturbance are collected in chronological order. In specific implementation, these segments are collected into a disturbance event set, each disturbance event includes its start time, duration, interference intensity, channel synergy degree and other information. Then, according to the physical characteristics and synergy characteristics of each disturbance event, all segments are summarized as the primary labeling data of the hidden physical disturbance. This data set will be used as the basis for subsequent analysis and optimization for further disturbance pattern recognition, working condition prediction and process control.

[0055] Alternatively, step S14 is specifically:

[0056] Step S141: Grouping and identifying according to the modal channel type of the implicit physical disturbance primary annotation data, obtaining the heat flow field segment, vibration frequency band segment, current disturbance segment and acoustic spectrum fluctuation segment;

[0057] In this embodiment, the data is grouped according to the modal channel type in the implicit physical disturbance primary annotation data. Specifically, all disturbance segment data is classified according to the modal signal (heat flow, vibration, current, acoustic spectrum) from which it originates. The heat flow field segment includes temperature change data recorded during device operation, the vibration frequency band segment covers vibration signals generated by the device, the current disturbance segment relates to current fluctuations of the device, and the acoustic spectrum fluctuation segment refers to acoustic spectrum changes generated by the device during operation. In this way, the data of each modal signal is divided into independent groups, facilitating subsequent individual analysis and processing of each channel. In specific operation, for each type of data, a dedicated processing channel is set up for subsequent parallel processing and synchronization.

[0058] Step S142: Distribute the heat flow field segment, vibration frequency band segment, current disturbance segment and acoustic spectrum fluctuation segment to the preset corresponding thread pool respectively, and activate four types of parallel processing threads to generate a multi-thread data stream initial channel set;

[0059] In this embodiment, for the heat flow field segment, vibration frequency band segment, current disturbance segment and acoustic spectrum fluctuation segment obtained in step S141, each type of data is distributed to the preset corresponding thread pool. Each thread pool contains multiple worker threads, and the number of worker threads in each thread pool is set to 4 according to actual needs. For example, heat flow signals are processed in parallel by a thread pool containing 4 threads, vibration signals are processed by another thread pool containing 4 threads, and so on. Each thread pool is initialized with a maximum of 4 threads, and the task queue of the thread pool uses a blocking queue (such as LinkedBlockingQueue) to ensure the orderly execution of tasks and prevent resource overload due to excessive task accumulation. To avoid thread contention, each thread in the thread pool is responsible for processing a specific subset of data, and the data is distributed to different threads through a predefined strategy. Each type of data is processed by an independent thread pool, and these thread pools will be activated and run in parallel. Specifically, heat flow signals are processed by one thread pool, vibration signals are processed by another thread pool, current signals are processed by a third thread pool, and acoustic spectrum signals are processed by a fourth thread pool. The worker threads in each thread pool will process the data segments assigned to them in parallel to generate a multi-thread data stream initial channel set. The size of the thread pool is set to 4 threads for each signal type to ensure efficient concurrent processing and avoid resource contention.

[0060] Step S143: Construct a standard time segment index for the perturbation segment data in each thread of the multi-thread data stream initial channel set, set a single index time window of 0.5 s, and perform sliding alignment on the perturbation segment data in each thread, remove segments with cross-channel time drift greater than 200 ms, and generate a synchronized time index segment set;

[0061] In this embodiment, the perturbation segment data in each thread of the multi-thread data stream initial channel set is subjected to standard time segment index processing. Specifically, the data in each thread is divided within a time window of 0.5 seconds to generate a time segment index. The length of each sliding window is 0.5 seconds, and the sliding step is 0.5 seconds, ensuring that the data in each time period is fully analyzed and synchronized. When performing time segment alignment, if the cross-channel time drift exceeds 200 milliseconds, the data is considered inconsistent and needs to be removed. This is done to ensure that the signals from different channels can be accurately synchronized, thereby ensuring the accuracy and reliability of subsequent analysis. Finally, after this processing step, a synchronized time index segment set is obtained.

[0062] Step S144: Set the aggregation window step to 0.5 s, perform thread aggregation operation on the synchronized time index segment set in each aggregation window, extract the cross-channel cooperative perturbation structure, and generate a perturbation thread fusion node set;

[0063] In this embodiment, the aggregation window step is set to 0.5 seconds to ensure that the disturbance fragments within each time window have sufficient temporal overlap. Within each aggregation window, all synchronized time segment data from different channels are aggregated, and the extraction of cross-channel coordinated disturbance begins. When performing thread aggregation operations, cross-modal feature correlation analysis techniques are used. Within each aggregation window, for the disturbance data in the synchronized time segment, first, the time domain and frequency domain analysis of the data in each channel is performed to extract its key features in time, frequency, and amplitude, such as the frequency components of the vibration signal, the fluctuation amplitude of the current signal, etc. Subsequently, using collaborative analysis techniques, the cross-channel disturbance structure is identified by calculating the similarity and correlation between different modal signals. For example, methods such as Pearson correlation coefficient or mutual information can be used to measure the degree of association between different signal channels. If the correlation between the vibration signal and the current signal is high within a certain time window, and the disturbance intensity of both reaches the preset threshold, it is considered that there is cross-channel coordinated disturbance in this time period. After completing the coordinated disturbance analysis, the cross-channel coordinated disturbance structure is extracted based on the coordination relationship between different channels. Using multi-modal coupling analysis, the disturbance features of different signals are fused within the same time window to form a fusion node of the multi-modal disturbance structure. The fusion node is labeled according to the correlation, coordination strength, and other characteristics of the signal, and a set of disturbance thread fusion nodes is generated. In order to further ensure the reliability of the extracted coordinated disturbance structure, each structure node is filtered during the extraction process. Specifically, a filtering condition is set, such as only when the coordination degree of cross-channel disturbance is greater than 0.6, can the disturbance structure be considered valid. At the same time, it also needs to meet the threshold requirement of disturbance intensity to ensure that only disturbance nodes with high influence are retained.

[0064] Step S145: The set of disturbance thread fusion nodes is serialized and integrated, and each disturbance thread fusion node is assigned a unique structure label to obtain a set of disturbance event structure vectors.

[0065] In this embodiment, the set of disturbance thread fusion nodes obtained is serialized and integrated, and each disturbance thread fusion node is assigned a unique structure label. These fusion nodes represent the coordinated disturbance of each modal signal during device operation, reflecting the mutual relationship and influence between different signals. In specific operations, each node will generate a unique label according to its position in the time sequence, disturbance intensity, cross-channel coordination degree, etc. For example, by combining device state, working condition information, and disturbance intensity, different disturbance modes are labeled. Finally, all the disturbance thread fusion nodes are integrated into a set of disturbance event structure vectors for further analysis and decision support. Each disturbance event contains its coordinated characteristics in the multi-modal signal and provides structured data support for subsequent prediction and optimization.

[0066] Optionally, the multi-modal grinding behavior feature analysis in step S2 is specifically:

[0067] The structured grinding field perception data is expanded in the time domain, and time sequence frames are reconstructed synchronously according to the grinding wheel line speed in the grinding parameter group, to obtain a grinding field perception time sequence frame set;

[0068] In this embodiment, the collected structured grinding field perception data is processed in the time domain. This process linearly recalibrates the time axis of multi-modal perception data such as heat flow, current, vibration, and acoustic spectrum, with the grinding wheel line speed in the grinding parameter group as the synchronous reference benchmark. To ensure time sequence consistency, the sampling period is unified to 1 ms, and data points with a sampling error greater than 5 ms are time-aligned through interpolation correction. Subsequently, the effective perception sequence dominated by the grinding wheel line speed in each section is extracted according to the continuous processing section in the grinding process, and different modal data are arranged into a unified frame structure based on the synchronous frame reconstruction method, generating a grinding field perception time sequence frame set with continuous time sequence numbers and consistent data formats.

[0069] Obtain abrasive structure data, and analyze the grinding ratio change trend based on the abrasive structure data and the grinding parameter group to obtain cutting zone grinding ratio change trend data;

[0070] In this embodiment, the process of obtaining abrasive structure data includes reading the preset abrasive parameter file in the system and synchronously analyzing the current grinding wheel brand and abrasive grain size grade. The abrasive structure data must include the abrasive topography factor and the abrasive grain arrangement density (unit: g / cm³). This data is combined with the real-time grinding parameter group for analysis, and a timed rolling window method (window length 10 seconds, step 2 seconds) is used to calculate the grinding ratio (G ratio) change trend of the grinding wheel in the cutting section. In the trend analysis, the grinding depth threshold is set to 0.03 mm, and the cutting speed fluctuation rate is limited to ±5% to exclude error interference under abnormal working conditions. The final output result is the trend curve of the cutting zone grinding ratio evolution with time, which is used to indicate the abrasive consumption stage and abrasive state change characteristics.

[0071] Combine the grinding field perception time sequence frame set with the cutting zone grinding ratio change trend data to perform multi-modal time synchronization alignment of heat flow, current, vibration, and acoustic spectrum, to generate a modal reconstruction sequence group;

[0072] In this embodiment, the grinding field perception timing frame set is matched with the extracted cutting zone grinding ratio change trend data. The synchronous alignment operation is based on the grinding ratio mutation point as the anchor point, combined with the peak response in the same time period in the multi-modal perception signal, to complete the main time axis reconstruction of the data frame. The maximum alignment offset is set to ±20ms, and the data segments that do not meet the alignment conditions are smoothed and interpolated. The aligned modal data forms a unified modal reconstruction structure in each frame, and is organized into a modal reconstruction sequence group according to the time number. The sequence group has the fields of heat flux density, current instantaneous value, vibration frequency band power spectral density and acoustic spectrum energy distribution, providing a consistent timing basis for subsequent disturbance recognition.

[0073] The modal reconstruction sequence group is subjected to contact arc zone disturbance feature recognition, and the transient component and energy mutation point of the modal disturbance are extracted.

[0074] In this embodiment, on the basis of the modal reconstruction sequence group, the contact arc zone disturbance recognition process is performed. First, the heat flux density change rate is greater than 40W / cm² / s, the current surge rate is more than 1.5A / ms, and the instantaneous power of the 2~5kHz frequency band in the acoustic spectrum is increased by more than 8dB as the preliminary disturbance candidate condition. On this basis, the transient component that meets the disturbance feature condition in each modal signal is extracted, and the energy mutation point is marked by combining wavelet analysis.

[0075] The disturbance significance threshold is set to 0.85, the characteristics of abrasive particle breakage, shedding and local burn in the grinding process are recognized according to the transient component and energy mutation point, and the grinding modal response significance map is formed.

[0076] For the recognition of mutation points in this embodiment, the local energy change rate threshold is set to 150%, and the effective disturbance duration is not less than 5 ms. The transient components of each modal data such as heat flow, current, vibration and acoustic spectrum in the modal reconstruction sequence group are extracted. With a sliding time window of 10 ms, the instantaneous energy change rate curve of each modal signal is calculated, and the average change rate is taken as the benchmark to mark the mutation section which exceeds the benchmark by 150% and lasts for more than 5 ms. Then, the disturbance type is identified in combination with the typical response characteristics of each modal. 1) For the abrasive particle breakage type disturbance, the following three conditions must be met simultaneously: first, the power spectral density in the 13 kHz frequency band in the vibration modal is detected to be increased by more than 15 dB and lasts for not less than 7 ms; second, there is a rapid current drop slope of more than -1.0 A / ms in the current modal; third, there is a high-frequency response of energy jump of more than 12 dB in the 57 kHz frequency band in the acoustic spectrum signal. If the above three conditions appear simultaneously, it is determined as the abrasive particle breakage type disturbance frame. 2) The identification of the abrasive particle shedding type disturbance includes the following: the instantaneous energy of the 12 kHz frequency band in the vibration modal is increased by between 810 dB and lasts for more than 10 ms; the current drop slope is between -0.3 and -0.6 A / ms; the heat flow density is decreased by more than 20% within 50 ms, and the average energy of the 13 kHz frequency band in the acoustic spectrum appears a bottom noise lift of 36 dB. The frames meeting the above conditions are classified as the abrasive particle shedding type disturbance. 3) For the local burn type disturbance, the heat flow data is taken as the leading indicator, and the judgment condition is that the heat flow density increases to more than 80 W / cm² within 30 ms, and the increase gradient reaches 2.5 W / cm² / ms, at the same time, there is a current peak of more than 1.8 A in the current modal, and the second-order derivative of the main shaft current is greater than 10 A / s². In terms of acoustic spectrum, there is a high-frequency wideband surge of more than 10 dB in the 68 kHz frequency band, accompanied by a transient decrease in energy in the low-frequency band (0.5-1.5 kHz) of vibration; the frames meeting the above comprehensive conditions are marked as the local burn type disturbance. Finally, all the identified disturbance frames are organized in time sequence, and their disturbance types and modal response intensities are marked to generate the grinding modal response significance map.

[0077] Based on the structured grinding field perception data, the cutting energy characteristics of each modal are extracted, and the cutting force fluctuation frequency, the acoustic pressure surge rate of the grinding area and the second-order derivative of the main shaft current are obtained.

[0078] In this embodiment, after extracting the grinding modal response significance map, further feature extraction of cutting energy dimension is carried out on the structured grinding field perception data. Based on the aligned modal data, the cutting force fluctuation frequency (based on the difference between the mean and standard deviation of the vibration frequency, unit: Hz), the grinding interval sound pressure surge rate (unit: dB / s) and the second derivative of the spindle current (unit: A / s2) in the window are calculated by using a 5-second sliding window (step 1 second). The time period when the cutting force fluctuation frequency is greater than 60 Hz, the sound pressure surge rate is more than 12 dB / s, and the absolute value of the second derivative of the current is more than 20 A / s2 is recorded as a high-energy disturbance segment. These features constitute the key indicators reflecting the process stability and abrasive stress response, providing high-precision energy description for further construction of energy synergy structure.

[0079] The cutting force fluctuation frequency, the grinding interval sound pressure surge rate and the second derivative of the spindle current are fused to construct a cross-modal energy synergy tensor;

[0080] In this embodiment, the energy feature data of each dimension is fused to construct a unified cross-modal energy synergy tensor. The construction of the tensor is based on a time window length of 10 seconds, and three modal features are fused: cutting force fluctuation frequency, grinding sound pressure surge rate and second derivative of spindle current. In order to ensure the numerical stability of the tensor, Z-score normalization processing is carried out first, and the dynamic feature channel mutual information screening method is used to exclude low contribution channels (channels with mutual information less than 0.1 will be discarded). After the tensor is constructed, it is divided into sub-tensor blocks with time continuity, which facilitates behavior clustering and change boundary identification analysis. The tensor structure is expressed in three-dimensional form: time frame x modal feature x state value, which completely covers the time sequence feature fluctuation and synergy response behavior of the entire grinding process.

[0081] Based on the cross-modal energy synergy tensor, behavior pattern aggregation and boundary identification are carried out to obtain a set of time sequence grinding behavior segments;

[0082] In this embodiment, behavior pattern aggregation and boundary identification tasks are carried out on the constructed cross-modal energy synergy tensor. The specific method includes dividing the tensor into 5-second sliding blocks, and identifying the behavior state boundary points by jointly modeling the relative amplitude and frequency of change of each modal feature value in each block. The behavior change threshold is set to be the feature value change amplitude exceeding 2 times the standard deviation of the mean value, and the number of mutation dimensions is not less than two modalities, that is, it is determined as a state transition boundary. The tensor segment with no significant boundary is defined as the same behavior segment, and the complete set of time sequence grinding behavior segments is finally output. Each behavior segment has a clear start and end frame, corresponding modal change trend and energy fluctuation structure, which facilitates subsequent behavior pattern recognition and tracing.

[0083] The set of time-sequential grinding behavior segments is mapped to the pre-connected grinding condition behavior library, and the behavior sequence pattern similarity is calculated to obtain the grinding behavior pattern data.

[0084] In this embodiment, the identified set of time-sequential grinding behavior segments is mapped to the pre-connected grinding condition behavior library. The behavior library predefines the structural feature templates of various standard condition behaviors, including grinding stable period, abrasive grain initial shedding, and severe burn precursor behavior units. During the mapping process, the shape similarity (calculated using DTW distance, and a similarity higher than 0.85 is determined as matching) of each behavior segment to the behavior templates in the library is calculated according to the energy fluctuation spectrum, modal synergy structure, and state duration of the behavior segment. Finally, the matched standard condition type of each behavior segment is output, along with the corresponding similarity score, to form the grinding behavior pattern data. This data can be used for further grinding process backtracking, quality prediction, and parameter optimization.

[0085] Especially important is that the cross-modal energy synergy tensor is constructed, specifically as follows:

[0086] Based on the cutting force fluctuation frequency, grinding interval sound pressure surge rate, and spindle current second derivative, a multi-modal energy feature matrix is constructed.

[0087] In this embodiment, after the cutting force fluctuation frequency, grinding interval sound pressure surge rate, and spindle current second derivative are time-axis synchronized and aligned, they are normalized to form an energy feature matrix with a matrix structure of N×3, where N is the time sampling frame number, the matrix row represents a time-sequential segment, and the columns are the cutting force fluctuation frequency, grinding interval sound pressure surge rate, and spindle current second derivative.

[0088] According to the grinding modal response significance spectrum, the multi-modal energy feature matrix is response-selected to eliminate energy features with a significance weight lower than the significance threshold of 0.85, to obtain a significant energy response set.

[0089] In this embodiment, the above energy feature matrix is one-to-one mapped to the time axis in the grinding modal response significance spectrum. In the specific operation, each row of feature vectors is mapped to the significance spectrum labeled frame according to its time stamp, and the maximum response value in the disturbance significance layer is extracted. The significance weight is normalized in the range of 0~1, and the response threshold is set to 0.85. Energy features with a value lower than this threshold are all removed and no longer participate in subsequent feature analysis. This processing method ensures that the selected feature matrix is composed of only energy segments that are representative in actual disturbance response, avoiding the fuzzy interference of low response segments in subsequent modeling. The retained matrix after screening is defined as the significant energy response set, and its size is determined according to the actual number of disturbances, usually accounting for 15%~25% of the total sampling frame number.

[0090] The modal coupling criterion is set to 0.7, the bilateral correlation analysis between modes is carried out in the significant energy response set, the synergistic weight matrix between different modal characteristic pairs is calculated, the modal pairs with coupling degrees greater than the modal coupling criterion are screened and reserved, and the high coupling degree modal pair is obtained;

[0091] In this embodiment, in order to identify the coupling behavior characteristics among multiple modes, bilateral correlation analysis is required between any two modal characteristics in the significant energy response set. Specifically, the modal characteristic pairs are composed of the cutting force fluctuation frequency and the sound pressure surge rate, the sound pressure surge rate and the spindle current second derivative, and the cutting force fluctuation frequency and the spindle current second derivative, respectively. The Pearson correlation coefficient is calculated in a sliding manner within a time window length of 200 ms, and the time window overlap rate is set to 50%. The calculation result forms a 3x1 coupling coefficient vector, which represents the average coupling strength between the three pairs of modes. The modal coupling criterion is set to 0.7. If the coupling strength is greater than the threshold, the modal combination is reserved as a high coupling degree modal pair. This analysis process can effectively identify the characteristic combination with synchronous and synergistic change trend among multiple modes during the grinding disturbance process, and provide a reliable modal structure basis for subsequent tensor construction.

[0092] Based on the high coupling degree modal pair, a third-order tensor structure is constructed, so as to obtain a cross-modal energy synergistic tensor.

[0093] In this embodiment, based on the high coupling modal pair screened out, a cross-modal energy synergistic tensor is constructed to express its synergistic characteristics. The construction method is to take the time step as the first dimension, the modal pair combination as the second dimension, and the coupling strength, synergistic energy jump amplitude and transient peak delay between each pair of modes as the third dimension, and the overall structure is like a third-order tensor structure of TxPxF, wherein T is the time series length, P is the number of high coupling modal pairs, and F is the number of synergistic characteristics of each pair of modes. In specific implementation, for each high coupling modal pair, three types of indexes are extracted: joint wave peak alignment error, energy same increase rate, and resonance frequency band synchronization ratio. After all the features are processed according to the standardized interval, they are written into the tensor. The finally constructed tensor structure can be used as an input carrier for subsequent behavior aggregation and grinding condition recognition, support the process of multi-dimensional pattern recognition, and provide a cross-modal coupling structure basis for grinding behavior mechanism analysis.

[0094] Alternatively, the grinding parameter semantic abstraction in step S2 is specifically:

[0095] The behavior-parameter mapping preliminary screening is performed on the grinding behavior mode data, and a potential association matrix between the behavior segments and the grinding parameters is constructed;

[0096] In this embodiment, in the process of grinding behavior-parameter mapping preliminary screening, the behavior pattern data obtained by disturbance analysis and energy coupling modeling needs to be first divided into a set of behavior segments according to time segments, and the processing parameter records in the corresponding time period are extracted. The processing parameters include but are not limited to feed speed, spindle speed, coolant flow, wheel linear speed and wheel dressing interval. For each behavior segment, the change interval and variation trend in the corresponding parameter dimension are counted by sliding window method, and the five-number summary method (minimum, lower quartile, median, upper quartile and maximum) is used to express the parameter distribution. When constructing the potential association matrix, the behavior segments are divided into three types: disturbance type, overload type and stable type. The co-occurrence frequency, interval overlap and trend alignment degree between each behavior class and parameter group are used as the three factors of the matrix elements. Finally, a three-dimensional potential association matrix with size MxNx3 is constructed, where M is the number of behavior segment classes, N is the number of processing parameters, and 3 is the dimension of the characteristic factor.

[0097] The local sensitivity threshold is set to 0.75, and the maximum mutual information criterion analysis is performed on the potential association matrix to identify significant behavior response zones, thereby generating a parameter candidate cross table;

[0098] In this embodiment, in the significant behavior response zone identification link, mutual information quantification processing needs to be performed on the above-mentioned potential association matrix to judge the association strength of each behavior type and different parameters in the local section. In the operation process, the local sensitivity threshold is set to 0.75, the mutual information value of each parameter and behavior segment combination is calculated in the time window of 0.5s, and the mutual information value is normalized. All combinations greater than the set threshold are marked as significant response zones to establish candidate association pairs. In order to improve the screening accuracy, the disturbance intensity index is combined to filter the candidate combinations, only the high-intensity parameter response pairs appearing in two or more behavior segments are reserved, and they are organized into a parameter candidate cross table. This cross table is used as the initial structure of subsequent parameter semantic modeling, where each row corresponds to a behavior pattern and each column corresponds to a potential response parameter.

[0099] According to the parameter candidate cross table, the grinding behavior related parameter dynamic characteristics are extracted, and the grinding behavior related parameter dynamic characteristics are evaluated for parameter information density distribution to obtain a grinding parameter semantic weight atlas;

[0100] In this embodiment, based on the parameter candidate cross table, the dynamic changes of the parameters contained therein in all significant behavior segments are classified into trends for further refining the behavior association features of the parameters. Taking the main shaft speed as an example, the amplitude changes thereof in the contact instability section are statistically analyzed by standard deviation, the change amplitude and average gradient thereof in each disturbance period are extracted, the average information density thereof is calculated, and the information density is defined as the number of effective value jumps per unit time. Repeat the process to complete the dynamic trend feature extraction of all parameters. Subsequently, a three-factor semantic scoring system is constructed using the frequency of the parameters appearing in different behavior modes, the peak value and distribution diffusion degree of the information density, and a grinding parameter semantic weight atlas is formed by aggregating the scoring results. The atlas is presented in the form of a heat map, with the X-axis representing the behavior type and the Y-axis representing the parameter type, and the heat value representing the representative strength of the parameters in the semantic layer.

[0101] Performing behavior semantic projection operation on the parameter abnormal mode with high frequency in the grinding parameter semantic weight atlas, and converting the parameter abnormal mode into semantic behavior labels including the contact instability section, the local overload section and the thermal drift risk section.

[0102] In this embodiment, in order to land the key behavior semantics to identifiable labels, it is necessary to perform behavior semantic projection on the parameter abnormal mode with high weight in the semantic weight atlas. In the specific operation, the parameter behavior combination with a semantic intensity value greater than twice the global average value of the atlas is selected, and the dynamic profile thereof under different behavior types is extracted. Taking the sudden drop of the grinding wheel linear speed as an example, if it appears synchronously with the dramatic increase of sound pressure and the fluctuation of main shaft current in multiple disturbance segments, and the co-occurrence rate is higher than 80%, it is marked as the semantic feature of the “contact instability section”. Similarly, if the main shaft current slope continuously rises and is superimposed with the mutation signal of heat flow, it is classified as the “local overload section”; the temperature curve appears drift after multiple period processing, and is superimposed with the situation of slow rise of other parameters, then it is marked as the “thermal drift risk section”. After the semantic naming of all behavior-parameter combinations is completed, a semantic behavior label set is formed, which is the core basis for the next step of parameter intent analysis.

[0103] A multi-nested index structure between the semantic behavior labels and the grinding parameter groups is constructed, thereby generating a grinding parameter behavior intent mapping cluster;

[0104] In this embodiment, when constructing the mapping structure between semantic behavior labels and parameter groups, a clear hierarchical relationship needs to be established based on the process meaning and risk level embodied by the semantic behavior label. First, according to the behavior type, a primary classification is performed, such as "contact instability", "local overload", "thermal drift", etc., and each category is further set with specific labels, such as "micro-impact instability", "short cycle load peak", "thermal accumulation drift", etc., as the index primary key. For each label, by tracing the grinding process parameter data associated with the behavior segment, the parameter combination with the highest frequency of occurrence of the label in multiple samples is extracted, and a preliminary parameter list is formed. To enhance the semantic expression ability of the list, the parameters need to be functionally divided. The following criteria are set: if the disturbance change of the parameter significantly precedes the start of the semantic behavior label within 5 seconds, it is defined as a core parameter; if it fluctuates immediately after the core parameter during the behavior occurrence period, it is defined as a secondary parameter; if the parameter changes relatively stably but its setting affects the fluctuation amplitude of the core or secondary parameter, it is defined as a background parameter. For example, in the "local overload segment" label, the spindle current peak usually appears 1-2 seconds before the behavior segment starts, which is a core parameter; the dressing interval of the grinding wheel has a significant impact on the current fluctuation rhythm but changes slightly later, which is a secondary parameter; the coolant flow is stable but the adjustment level affects the current thermal effect, which is a background parameter. After the construction is completed, a three-level nested structure of behavior-parameters is formed, each layer having a resolvable process meaning. Taking the behavior label as the index primary key, the corresponding core, secondary, and background parameters are hierarchically aggregated to construct a parameter behavior intention mapping cluster. In actual process flow, this mapping cluster can be directly connected to the grinding process monitoring system to realize behavior recognition driven parameter linkage adjustment, such as when the "contact instability" label is triggered, the feed rate and grinding wheel linear speed can be adjusted according to the mapping cluster as part of the response strategy.

[0105] The grinding parameter behavior intention mapping cluster is converted into a multi-dimensional nested structure for parameter dimension semantic compression and up-down coding to obtain a grinding parameter intention nested mapping table.

[0106] In this embodiment, a plurality of mapping clusters are merged and encoded according to the semantic hierarchy and the superior-inferior relationship. In this process, the parameter set involved in all mapping clusters needs to be vectorized and normalized first, and the linear normalization of each parameter value is performed according to the standard interval (such as [0, 1]) to avoid the influence of parameter dimension difference on the accuracy of structure construction. The mapping clusters are divided into levels according to the risk level of the behavior embodied by the semantic label. The risk level evaluation standard refers to the parameter response strength, behavior duration and the influence of behavior on grinding stability. For example, the "thermal drift" label is classified as low risk, "contact instability" is set as medium risk, and "local overload" is classified as medium-high risk. On the basis of this structure, semantic compression is performed on the mapping clusters of the same behavior attribution. The specific way is to merge the multiple behavior labels corresponding to the mapping clusters into a unified semantic node, and extract the intersection of the core parameter set to form a "high semantic density" compressed node, avoiding semantic repetition. After completing the semantic merging, a multi-dimensional nested mapping table is constructed. The mapping table adopts a tree-like structure, with the risk level as the first layer of index node, and the semantic label layer, parameter type layer (core / secondary / background) and specific parameter item layer are set below. Each path represents a mapping chain from high-level behavior semantics to specific parameters. By introducing the nested structure, parameter commonality mining across semantic labels can be realized, such as under the "medium-risk behavior" node, both "contact instability segment" and "short cycle instability segment" can call two parameters of grinding wheel linear speed and spindle current, which facilitates strategy merging and process unification. The final grinding parameter intention nested mapping table not only supports upward derivation (from parameter fluctuation to identify risk behavior), but also supports downward analysis (from risk warning to derive parameters that need to be monitored), greatly improving the engineering practicability and response efficiency of the semantic parameter system.

[0107] Optionally, step S3 is specifically:

[0108] Step S31: tensorizing and expanding the grinding parameter intention nested mapping table to construct an initial parameter tensor field;

[0109] In this embodiment, in order to realize the high-dimensional data processing requirement of the intention nested mapping table, a tensorization expansion operation needs to be performed on the grinding parameter intention nested mapping table. Specifically, each "semantic behavior label-parameter combination path" in the mapping table is converted into a tensor element, wherein the semantic label code is taken as the first dimension, the parameter category (core, secondary, background) is taken as the second dimension, and the parameter value interval is divided into the third dimension. In order to improve the structural expression, the parameter value interval is divided into five level sections according to the quintile method, and the upper and lower bounds of the interval are marked. The tensor shape is designed as a three-dimensional tensor, such as L, C, and V, wherein L is the number of semantic labels, C is the number of parameter categories, and V is the interval division number of each category of parameters. Taking the "local overload section" label as an example, in the corresponding tensor unit, the main shaft current is taken as 0.9-1.2 A as the high load section, and is marked as the fifth level interval. The finally constructed initial parameter tensor field has a unified data format, which is convenient for subsequent structural disturbance and path encoding processing.

[0110] Step S32: According to the upper and lower boundary combination of each parameter dimension in the initial parameter tensor field, a constraint space boundary set is constructed, and the constraint space boundary set is subjected to non-uniform disturbance to obtain a coarse solution space structure grid;

[0111] In this embodiment, in order to model the parameter optimization path, a constraint space boundary set needs to be constructed based on the upper and lower boundary combination of each parameter dimension in the initial parameter tensor field. The specific operation is to extract the upper and lower boundaries of each parameter under each semantic label in the tensor field, combine to form a constraint boundary element group, and uniformly classify to form a multi-parameter boundary set. In order to improve the adaptability of the search space, a non-uniform disturbance strategy is further introduced, that is, unequal interval disturbance points are applied in each boundary interval, and the interval distance of the disturbance points is set to be 3%-10% of the original interval, and the density of the disturbance points is controlled by using a Gaussian offset (the mean value is set to be the interval center, and the standard deviation is 0.2 times the interval width). In the "thermal drift risk section", if the main shaft temperature rise is set to be 50℃±5℃, the generated disturbance points will cover the unequal density distribution of 46.5℃ to 53.5℃. Through the above method, a coarse solution space structure grid is obtained, which lays a preliminary space structure for subsequent parameter search.

[0112] Step S33: The coarse solution space structure grid is subjected to particle swarm space initialization, and each particle of the particle swarm is subjected to initial parameter path encoding to obtain an initialized basic particle swarm;

[0113] In this embodiment, after obtaining the coarse solution structure grid, the spatial group structure needs to be initialized. The particle swarm initialization method can be used to construct the group starting point structure of the parameter behavior path. First, set the number of group particles P=80, and each particle represents a group of parameter path combinations to be optimized. During initialization, each particle randomly selects a group of parameter upper and lower limit combinations from the coarse solution grid as the initial position, and encodes the three-dimensional path based on the tensor path structure, wherein the semantic label index, parameter category index and interval coding are used as the encoding vector input. To enhance the coverage of the search space, the initial distribution between particles in each dimension is as uniform as possible. The parameter initialization path length is fixed at 6 steps, and each step of the encoding combination contains 1 core parameter, 1 secondary parameter and 1 background parameter. The initial distribution state of this particle structure in space is the basic particle group, which has strong diversity and structure analysis capability.

[0114] Step S34: Perform coarse-grained search evolution on the initialized basic particle group to generate a coarse-grained grinding parameter trajectory group.

[0115] In this embodiment, after the basic particle group is constructed, preliminary path evolution exploration is needed to obtain the coarse-grained trajectory group of the grinding parameters. The initialized group is subjected to coarse-grained search evolution operation. The specific method is that in each iteration, each particle adjusts the forward direction according to the fitness of its current parameter combination under the semantic behavior target. The adjustment rule consists of two parts: one is to evaluate the correlation strength of the current parameter combination based on the historical mapping weight of the semantic label, and the other is to adjust the parameter jump amplitude according to the path smoothness in the tensor structure. In actual operation, the maximum parameter adjustment step in each round is set to 15% of the initial interval width, the search round number is 50 rounds, and the top 20% of the particles in each round are reserved as trajectory reservation groups, and the rest are randomly disturbed and regenerated. When the "contact unstable section" is used as the target semantic label, most of the reserved particles show a clear convergence trend in the feed speed and normal force direction, forming a stable coarse-grained parameter trajectory group, which provides a sample reference for the subsequent fine tuning stage.

[0116] Alternatively, step S34 is specifically:

[0117] Step S341: Map the initial parameter path of the initialized basic particle group to the corresponding constraint space, and calculate the position similarity of the particles in the constraint space to assign the initial velocity and inertia weight to the initialized basic particle group to obtain an initial velocity particle group;

[0118] In this embodiment, the initial parameter path of the initialized basic particle swarm needs to be structured and mapped. Specifically, assuming that the total number of particles is N = 100, each particle is composed of a combination of three dimensions of grinding parameters: spindle speed, feed speed and grinding depth, and the initial values are randomly and uniformly generated in the intervals [5000, 8000] rpm, [100, 300] mm / min, [0.01, 0.05] mm, respectively. To achieve structural constraint mapping, these three combinations need to be embedded into a pre-defined non-uniform perturbation constraint space (coarse solution space structure grid). The tensor structure is divided by non-linear scale in each dimension according to the perturbation density, for example, the spindle speed dimension is divided by logarithmic section interval to improve the perturbation sensitivity in the middle and high interval. After mapping, each particle obtains its position vector in the tensor space, represented in three-dimensional coordinate form. Taking a particle parameter (6800 rpm, 200 mm / min, 0.03 mm) as an example, its mapping coordinates in the tensor space may be (0.62, 0.47, 0.55). Subsequently, based on the Euclidean distance formula, a 100x100-dimensional position similarity matrix is constructed, and a distance threshold is set. For any two particles, if the distance is less than 0.35 (0.65 similarity equivalent threshold after standardization), they are classified into the same neighborhood group, otherwise they are marked as dispersed. Further, for each particle, the initial velocity vector is calculated according to the average distance between it and its neighboring particles, and the vector size is limited in the interval [0.1, 0.9]. The velocity direction is determined according to the distance gradient direction, and a higher initial velocity is given to those far away from similar particles to encourage structural divergence behavior. The inertia weight is initially set to 0.7 as the velocity continuation factor in subsequent jump updates. Finally, the initial velocity particle swarm with structural embedding information, velocity vector and inertia weight is obtained.

[0119] Step S342: randomly jump updating the initial velocity particle swarm using the constraint space to obtain a perturbation response particle swarm;

[0120] In this embodiment, on the basis of the constraint space structure grid, the initial velocity particle swarm is subjected to a jump disturbance update operation. The specific implementation is as follows: according to the current velocity vector of the particle, a non-continuous jump disturbance is applied to the current position parameter combination of the particle, the disturbance scale is generated by random offset based on 25% or less of the current dimension interval width, and the disturbance direction is based on the superposition of a Gaussian offset factor (standard deviation is 0.2 times the disturbance scale) on the velocity vector direction. In order to control the rationality of the structure, the position of the particle after jumping must be kept within the range of the constraint boundary, and if it is out of range, a mirror reflection compensation process is performed. Taking the main shaft speed interval of 5000-8000 rpm as an example, if a particle is updated to 8100 rpm after jumping, it is re-adjusted to 7900 rpm according to the boundary symmetry point. After the disturbance process, the particle swarm position completes the first jump scattering in the search space, forming a disturbance response particle swarm, which provides diversified samples for subsequent fitness evaluation.

[0121] Step S343: Calculate the fitness value of the current solution of each particle in the disturbance response particle swarm;

[0122] In this embodiment, the fitness of each particle in the disturbance response particle swarm is evaluated to obtain its response quality under the current semantic behavior target. Taking “contact stability improvement” as the semantic behavior target, the feature tensors related to it are extracted from the structured perception data of the fully automatic coarse and fine grinding integrated machine, such as normal force fluctuation, main shaft current change rate, contact vibration frequency, etc. Each particle's current parameter combination is mapped to these feature tensor dimensions, and similarity matching is performed with historical behavior label data through sliding window comparison. The matching degree is calculated using dynamic time warping (DTW) distance weighted superposition evaluation, and the behavior matching score P_i is output, where P_i ∈ [0, 1]. Secondly, for the local feature sequence of the matching position, the standard deviation and mean rate of change in the fluctuation interval are evaluated to measure the stability of the behavior response, and the stability score S_i is also normalized to [0, 1]. The final fitness value F_i is composed of the weighted sum of the two: F_i = 0.6 × P_i + 0.4 × S_i; for example, if the parameter combination of a particle has a matching degree of 0.88 and a stability score of 0.75, then its fitness is F = 0.6 × 0.88 + 0.4 × 0.75 = 0.832. In order to facilitate classification processing, the fitness level division standard is set as follows: F ≥ 0.85 is a high matching particle (significant response); 0.6 ≤ F < 0.85 is a medium response particle; F < 0.6 is a low fitness particle (not recommended to be retained). All particle fitness values are finally stored in the structured fitness matrix, which is an important basis for subsequent optimal solution selection and behavior trajectory clustering, and can also be used for process monitoring and performance feedback of parameter trajectory learning.

[0123] Step S344: According to the fitness value of the current solution, the local optimal solution and the global optimal solution are selected for each particle, and the particle parameters are updated according to the local optimal solution and the global optimal solution, to obtain an updated particle swarm;

[0124] In this embodiment, after the fitness evaluation is completed, the particle swarm needs to be updated based on the target guidance structure. The optimal parameter combination in the history iteration of each particle can be found as the local optimal solution, and the one with the highest fitness in the current particle swarm is selected as the global optimal solution. The update operation is combined and corrected according to the local and global guidance directions. The new position vector of each particle is composed of the current position and a set of weighted difference vectors. The local guidance item weight is set to 0.4, the global guidance item weight is set to 0.6, and the step factor range is [0.1, 0.3] and is randomly generated. Taking the main shaft current as an example, if the current value is 1.0 A, the local optimal value is 1.2 A, and the global optimal value is 1.4 A, then the new value is updated to: 1.0 + 0.4 * (1.2 - 1.0) + 0.6 * (1.4 - 1.0) = 1.32 A. After the above process, a new round of particle swarm structure is obtained, and preparation is made for subsequent iterations.

[0125] Step S345: The updated particle swarm is returned to step S342, and steps S342 to S345 are one iteration. Ten iterations are performed. According to the solution and fitness evaluation result of each particle in the particle swarm, group collaborative clustering is performed, the particles are grouped, and a coarse-grained grinding parameter trajectory group is output.

[0126] In this embodiment, after the particle swarm completes one iteration update, it is subjected to a backflow operation and a multi-round behavior collaborative evolution. The updated particle swarm is returned to step S342, and random jumping, fitness evaluation and solution updating are performed for 10 rounds. The history optimal path and fitness value trend of each particle are recorded after each round. After 10 rounds, group collaborative clustering operation is performed on all particles based on their fitness performance and parameter path similarity. The clustering is based on the joint distance measurement of the encoding vector of the particle on the tensor path structure and the final fitness. The number of clusters is set to 5, and the joint criterion of density distribution and semantic label similarity is used. Finally, multiple coarse-grained parameter trajectory groups are formed, each group has consistent behavior response characteristics and high fitness characteristics, and is suitable for subsequent working condition matching and control strategy generation tasks.

[0127] Optionally, the soft clustering bias learning in step S4 is specifically:

[0128] The feature domain normalization mapping processing is performed on the heterogeneous trajectory graph encoding vector set to generate uniform scale parameter space data;

[0129] In this embodiment, in order to unify the numerical scale of different trajectory graph encoding vectors in each feature dimension, normalization processing is required for all encoding vectors. First, the maximum and minimum values of each parameter dimension in the encoding vector set are extracted, and the Min-Max standard scaling formula is used for mapping for all dimensions: normalized value = (original value - minimum value) / (maximum value - minimum value). In order to ensure the numerical stability of the normalization process, the minimum scaling interval width is set to 0.01. If the actual variation range of a certain parameter dimension is less than this threshold, the dimension is uniformly assigned a value of 0.5 after normalization to avoid weight imbalance caused by flat dimensions. In the processed uniform scale parameter space, each dimension value of all encoding vectors is limited to the interval [0, 1], thereby providing a consistent measurement basis for subsequent fuzzy clustering operations.

[0130] Based on the uniform scale parameter space data, perform biased fuzzy clustering to obtain a biased fuzzy initial membership set;

[0131] In this embodiment, in the uniform scale parameter space after normalization, biased fuzzy membership modeling processing is performed based on the historical working condition model to construct the initial membership degree structure. In the specific implementation process, first, the heterogeneous trajectory graph encoding vector set is standardized and transformed to ensure that all feature dimensions have zero mean and unit standard deviation in the standard scale space, excluding numerical shifts caused by dimensional differences. Then, a historical working condition model is constructed based on the grinding working condition behavior library, and at least 500 typical working condition sample sequences with time labels are extracted. The sliding window behavior segmentation and trend fitting method is used to construct the behavior bias vector set to depict the distribution tendency of the historical working condition evolution trend in the feature space. Finally, based on the uniform scale parameter space data, the number of fuzzy aggregation clusters is set to 10, the fuzzy weighted index is set to 2.1, the historical bias vector is introduced as the fuzzy membership initialization guide, the Gaussian decay weight function (bandwidth parameter σ = 0.25) is used to enhance the directional membership weight, and the fuzzy membership degree distribution interval is set to [0.3, 0.95]. Each trajectory vector can be attributed to 3 candidate classes, and the membership difference between classes is not less than 0.2. A label preference factor is introduced, with a value of 0.15, to improve the attractiveness of behavior-related class centers. The processing result forms a biased fuzzy initial membership set, each vector has membership probability weight in multiple fuzzy classes, meets the fuzzy recognition requirements of grinding behavior multi-semantic features, and lays the membership foundation for subsequent center transfer modeling and graph fusion.

[0132] According to the biased fuzzy initial membership set, perform iterative center transfer learning, constrain the center drift rate of each iteration to be not more than 0.05, and obtain a class center transfer trajectory set;

[0133] In this embodiment, the class center convergence process of the initial membership set is modeled. Each iteration is updated by weighted average of class center coordinates, and the weight coefficient is provided by the membership matrix. To ensure the stability and convergence of the clustering process, the drift rate of each class center update is not more than 0.05, and the drift rate is defined as the Euclidean distance between the centers of the previous two rounds divided by the maximum distance between the initial classes. If the drift rate of a class center in a certain iteration exceeds the limit, the buffer adjustment is performed by smoothing interpolation, and the buffer factor is set to 0.5. The iteration process is executed for a maximum of 30 rounds, or terminated in advance when the drift rate of all class centers is less than 0.01, and the dynamic movement path of each class center is finally recorded to construct the class center transfer trajectory set, providing the basis for historical evolution information for subsequent stability judgment.

[0134] The dynamic stability of the class center transfer trajectory set is evaluated, and the classes whose center trajectory variance exceeds the stability threshold of 0.08 are removed, and the remaining class centers are fitted and modeled to generate a convergent class center space.

[0135] In this embodiment, the reliability of the class center in the entire transfer trajectory needs to be determined according to the drift stability. Based on the time series of each class center in each dimension, the variance of the change is calculated, and the stability threshold is set to 0.08. The classes exceeding the threshold are considered unstable. For the class center trajectories that are not removed, further trajectory curve fitting processing is performed, and a quintic polynomial fitting method is used to model the center change path in each dimension. The fitting accuracy requires that the coefficient R² be higher than 0.93, and if it does not meet the requirement, local fitting is performed instead. Finally, a structure-stable and traceable convergent class center space is formed, providing a stable anchor point for subsequent membership backtracking and weight assignment.

[0136] According to the convergent class center space, the multi-class membership distribution of each trajectory vector is calculated in reverse, a multi-membership structure map is constructed, and each coding vector in the heterogeneous trajectory graph coding vector set is given a confidence scalar weight according to the multi-membership structure map, to obtain a high-confidence grinding parameter soft clustering structure.

[0137] In this embodiment, based on the fitted convergent class center structure, the reverse membership calculation of the original coding vector set is performed. Each coding vector calculates the initial membership according to the reciprocal of the distance from each class center, and uses normalization processing to obtain the complete multi-class membership probability distribution. The confidence weight allocation function is set as: confidence = max(membership distribution) x local density adjustment coefficient, and the local density adjustment coefficient is set according to the number of neighbors within a radius of 0.05 around the vector in the class, ranging from 0.6 to 1.0. Finally, each coding vector is marked with a confidence weight, and a soft clustering structure cluster with multi-class label support is obtained, preparing for the atlas merging and semantic compression.

[0138] The high-confidence grinding parameter soft clustering structure is subjected to similarity atlas merging and nested graph compression, and a high-dimensional grinding parameter candidate atlas is output.

[0139] In this embodiment, in order to aggregate the attributed code vectors into a more compact atlas expression form, atlas fusion processing needs to be performed based on the similarity between vectors. Taking cosine similarity as the core index, the atlas merging threshold is set to 0.85, and the vectors exceeding the threshold are aggregated into subgraph structures. Further, the subgraphs with overlapping boundaries are subjected to nested graph compression, and the compression strategy adopts two-stage operations of node merging and edge weight fusion. The maximum confidence selection principle is adopted for edge weight fusion. Finally, a high-dimensional grinding parameter candidate atlas with compact structure and clear semantics is output, which has the ability to support work condition reasoning, strategy formulation and other multi-scene applications.

[0140] Especially important is that the bias fuzzy clustering is specifically:

[0141] The feature domain normalization processing is performed on the heterogeneous trajectory graph coding vector set, and each dimension feature is mapped to a standard scale space with zero mean and unit variance, generating a unified scale parameter space data;

[0142] In this embodiment, in order to ensure the comparability of different feature dimensions in the metric space in the heterogeneous trajectory graph coding vector set, each feature needs to be standardized and mapped. In the specific implementation process, the mean and standard deviation of each dimension in the vector set are calculated, and then the original value is subtracted from the mean of the dimension and divided by the standard deviation of the dimension, so that each feature dimension presents a distribution state of zero mean and unit standard deviation in the standard scale space. To avoid numerical abnormalities caused by the denominator approaching zero, a minimum standard deviation threshold of 0.02 is set. If the actual standard deviation of a certain dimension is lower than the threshold, the standardization is performed after the minimum standard deviation threshold is replaced. After the above transformation, all vectors are mapped to a unified scale parameter space, providing a basic scale support for subsequent fuzzy processing and behavior modeling process, and excluding the numerical deviation risk caused by the inconsistency of different source feature dimensions. The so-called standard scale space, in this embodiment, refers to a multi-dimensional vector space in which all dimension features satisfy the distribution of zero mean and unit standard deviation. The main technical goal is to eliminate the clustering deviation or weight imbalance problem caused by the inconsistency of different feature values, and to enhance the stability and adaptability of the subsequent processing stage.

[0143] According to the grinding condition behavior library, a historical condition time sequence model is established, and a historical condition model is obtained;

[0144] In this embodiment, in order to effectively fuse the historical grinding behavior characteristics, it is necessary to extract typical time evolution samples based on the grinding condition behavior library constructed, and to perform behavior pattern modeling. In the specific implementation process, first, not less than 500 complete condition sequence samples are selected from the behavior library, and the samples need to include three key fields of time stamp, corresponding vector feature and condition label. A fixed time window sliding method is used to divide each 10 time points in the sequence into behavior segments, forming an equal-length time sequence fragment set. Then, the evolution trend of the key behavior indicators in each time sequence is extracted by feature statistics, and these evolution trends are fitted into a behavior bias vector model. The bias vector length needs to be consistent with the trajectory encoding vector dimension, which is used to express the historical distribution tendency of the condition behavior in the feature space. Finally, this model serves as an important reference for the initialization of subsequent fuzzy attribution, making the fuzzy aggregation process have a historical behavior orientation, and effectively improving the recognition rate of weak sample behavior in the grinding scene.

[0145] Based on the unified scale parameter space data, the number of clusters of the biased fuzzy clustering is set to 10, the fuzzy weighting index is set to 2.1, and the bias vector of the historical condition model is introduced as the membership initialization guide to perform membership initialization, thereby generating a biased fuzzy initial membership set.

[0146] In this embodiment, in the unified scale parameter space, in order to realize vector aggregation processing with behavior orientation, the biased fuzzy membership initialization needs to be performed in combination with the historical condition model. In the specific implementation, first, the number of clusters for fuzzy processing is set to 10, and the fuzzy weighting index is selected as 2.1 according to experience to control the fuzziness of the membership distribution. At the same time, the bias vector set of the constructed historical condition model is introduced as the initialization behavior guide weight, and each bias vector is taken as the initial center direction to enhance the attraction effect of the corresponding cluster at the initial stage of membership allocation. The membership initialization matrix is constructed by weighting the distance between the sample point and the bias vector, and the distance weight function is set as a Gaussian decay function with a bandwidth parameter σ value of 0.25 to improve the center attraction ability of the bias class. After the above initialization is completed, a biased fuzzy initial membership set covering the entire sample set is formed, so that each trajectory encoding vector obtains a set of fuzzy membership probabilities associated with the historical behavior model, laying a fuzzy structure foundation for subsequent convergence path learning and structure compression.

[0147] Optionally, step S5 specifically comprises:

[0148] Step S51: according to the granularity parameter, cutting depth, linear velocity, contact arc length and grinding ratio index corresponding to each candidate parameter structure in the high-dimensional grinding parameter candidate atlas, an initial condition state tensor is constructed;

[0149] In this embodiment, the basic working condition elements associated with each parameter structure need to be extracted from the constructed high-dimensional grinding parameter candidate atlas one by one, specifically including granularity parameter (unit: μm), cutting depth (unit: mm), linear velocity (unit: m / s), contact arc length (unit: mm) and grinding ratio (dimensionless). During the extraction process, the extraction batch unit is set to 20 groups of candidate parameter structures each time to ensure the parallel construction efficiency. In each group of candidate structures, the five indicators are integrated into a five-dimensional state vector in a fixed order, and the three-dimensional tensor structure is sequentially stacked vector by vector, with the dimension defined as [N×5×1], where N is the number of candidate parameters.

[0150] Step S52: Perform multi-physical field simulation based on the initial working condition state tensor, continuously simulate each group of parameters for 12S, and respectively evaluate the physical indicators of grinding ratio, surface roughness, heat affected zone width and grinding energy consumption to obtain working condition simulation response data;

[0151] In this embodiment, based on the constructed initial working condition state tensor, virtual simulation evaluation operation of candidate parameters is performed relying on a simulation environment with thermal-mechanical-grinding multi-field coupling capability. The simulation time of each group of parameters is set to 12 seconds, and the sampling frequency is 1000Hz during the simulation process to ensure that the time evolution curve of each physical indicator has sufficient resolution. The physical evaluation indicators include grinding ratio (calculated by the ratio of material removal amount to grinding wheel wear), surface roughness Ra (unit: μm), heat affected zone width (unit: μm) and energy consumption per unit time (unit: J / s). The average value of each physical quantity in the stable stage is recorded by the sampling end as the simulation response output of this group of parameters. To improve the simulation efficiency, a multi-thread parallel simulation framework is adopted, and each simulation batch is run concurrently once every 8 groups of parameters, with the average total simulation time controlled within 30 minutes. The final output is a structured response data table containing the corresponding parameter index and four physical indicator values, which serves as the working condition simulation response data set.

[0152] Step S53: Normalize the working condition simulation response data, set the index weight group [0.35, 0.25, 0.2, 0.2], calculate the performance score of each group of candidate parameters, and generate a candidate parameter performance evaluation table;

[0153] In this embodiment, the response data of all working conditions is normalized by linear interval standardization method, and all physical indicators are mapped to the interval [0, 1]. After normalization, the index weight set [0.35, 0.25, 0.2, 0.2] is introduced, which corresponds to the grinding ratio, surface roughness, heat affected zone width and energy consumption respectively, and the surface roughness and heat affected zone are the secondary indicators. The performance score is obtained by weighted summation, and each group of parameters corresponds to a score result. To ensure the physical interpretation of the scoring process, set the low value as the optimal indicator (such as Ra and energy consumption) to take the reciprocal processing before normalization. The final result takes the parameter index as the key and the performance score as the value to form the candidate parameter performance evaluation table. This embodiment processes 100 groups of parameters and outputs 100 score data to support subsequent parameter screening and judgment.

[0154] Step S54: According to the candidate parameter performance evaluation table, set the score threshold to 0.82 to screen the stable parameter subset;

[0155] In this embodiment, in order to realize the performance stability, the score in the candidate parameter performance evaluation table needs to be threshold screened. Combined with historical simulation experiments and experience data, the evaluation score screening threshold is set to 0.82, which is intended to eliminate parameter combinations with obvious performance disadvantages. In the screening process, the score of each group of parameters is determined in one direction, and any score below 0.82 is marked as a performance substandard group and is removed from the parameter candidate set. This batch processes 100 groups of parameters, and after screening, 37 groups of stable subsets with scores higher than the threshold are retained to form a stable parameter index set. This subset will be used for dimension loop evaluation processing in the subsequent steps. The threshold in this step can be flexibly adjusted according to specific application requirements, and the recommended value range is usually between 0.8 and 0.85. The specific value should be determined according to the robustness requirements of the target working condition.

[0156] Step S55: Perform dimension loop mapping and redundancy removal processing on the stable parameter subset to construct a stable grinding parameter loop tensor set, and perform dimension-by-dimension convergence evaluation on the stable grinding parameter loop tensor set, set the stability fluctuation threshold to ±3%, to eliminate periodic loop drift parameter combinations, and generate a stable grinding parameter set.

[0157] In this embodiment, further structured compression processing is performed. First, a stable grinding parameter loop tensor set is constructed, and each group of parameter vectors is remapped to a loop structure in the original five-dimensional index order, and the loop period is set to 10 groups, forming a tensor set structure with dimensions of [37x5x10]. Subsequently, the mean value and maximum fluctuation amplitude of each dimension in the loop period are calculated. If the relative fluctuation amplitude of any parameter dimension in the period exceeds the set threshold ± 3%, it is determined that there is a periodic drift, and this group of parameters will be marked as unstable structure and excluded. Finally, the remaining is the parameter structure that is stable in all dimensions in the loop period, a total of 24 groups. The stable grinding parameter set will be input to the actual process verification process or further working condition compression stage as a high reliability configuration scheme, ensuring that the subsequent process optimization process has high reliability and process consistency.

[0158] Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting, the scope of the application being defined by the appended claims rather than the above description, and it is intended to encompass all variations falling within the meaning and scope of the equivalent elements of the application file.

[0159] The above description is merely one specific implementation of the application, which enables those skilled in the art to understand or implement the application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the application. Therefore, the application will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for optimizing grinding parameters of a fully automated coarse and fine grinding integrated machine based on particle swarm optimization, characterized in that, Includes the following steps: Step S1: Acquire multimodal data of equipment operation and perform edge-level preliminary processing on the multimodal data of equipment operation to extract latent physical disturbance features; The implicit physical disturbance features are aggregated in multiple threads to obtain structured grinding field sensing data; Step S2: Perform multimodal grinding behavior feature analysis on the structured grinding field perception data to obtain grinding behavior pattern data; perform semantic abstraction of grinding parameters based on the grinding behavior pattern data to obtain a nested mapping table of grinding parameter intents; Step S3: Perform initialization basic particle swarm optimization on the nested mapping table of grinding parameters to obtain the initialization basic particle swarm; perform coarse solution space boundary evolution based on the initialization basic particle swarm to obtain the coarse-grained grinding parameter trajectory family. Step S4: Perform subgraph nesting reconstruction on the coarse-grained grinding parameter trajectory family and construct a heterogeneous trajectory map encoding vector set; perform soft clustering bias learning on the heterogeneous trajectory map encoding vector set to obtain a high-dimensional grinding parameter candidate map. Step S5: Perform multi-dimensional working condition simulation based on the candidate spectrum of high-dimensional grinding parameters, and perform parameter tensor loop closure optimization based on the multi-dimensional working condition simulation results to obtain a stable grinding parameter set; Step S6: Transmit the stable grinding parameter set to the management platform of the fully automatic rough and fine grinding machine to execute the control parameters, and perform iterative accuracy compensation mapping based on the real-time acquired actual grinding deviation feedback data to obtain the optimized grinding parameter set.

2. The grinding parameter optimization method for a fully automatic integrated roughing and finishing grinding machine based on particle swarm optimization as described in claim 1, characterized in that, Step S1 is as follows: Step S11: Acquire multimodal data of equipment operation, which includes grinding parameter set, high frequency vibration signal, spindle current fluctuation record, temperature rise heat flow field image and accompanying acoustic spectrum signal; Step S12: Perform channel-by-channel decoupling processing on the multi-modal data of equipment operation, and identify the physical field disturbance trend of the decoupling processing results to obtain physical field disturbance trend data; Step S13: Extract latent physical disturbance segments based on the physical field disturbance trend data to obtain the initial annotation data of latent physical disturbances; Step S14: The implicit physical disturbance primary annotation data is split into threads, and time segment indexes are constructed for heat flow, vibration, current and acoustic spectrum signal segments respectively. The aggregation window step size is set to 0.5s to perform thread merging operation, and the disturbance event structure vector set is obtained. Step S15: Perform structural consistency encoding on the set of disturbance event structure vectors, and fuse the acquired equipment operating timestamps, operating condition identifiers and part index information to obtain structured grinding field sensing data.

3. The grinding parameter optimization method for a fully automated coarse and fine grinding integrated machine based on particle swarm optimization according to claim 2, characterized in that, Step S13 is as follows: Step S131: Set the window length to 2s and the step size to 0.5s, and perform segmented sliding window division on the physical field disturbance trend data to obtain the sliding segment disturbance trend data; Step S132: Calculate the differential value of the disturbance intensity for each sliding segment in the disturbance trend data of the sliding segment, and set the disturbance intensity mutation threshold to 1.75 to identify the initial segment of potential disturbance anomaly and generate the initial screening dataset of disturbance segments. Step S133: Group the initial screening dataset of the disturbance section by modal channel type, construct vibration frequency-energy map, current fluctuation spectrum map, heat flow intensity map and acoustic spectral density map respectively, and perform frequency domain entropy increment analysis on each group of maps to screen out sections that meet the preset multi-channel synchronous entropy fluctuation conditions and generate multi-channel coupled disturbance candidate segments; Step S134: Perform co-occurrence pattern identification on candidate segments of multi-channel coupled perturbation, screen out isolated segments, and set the modal consistency evaluation index to be greater than 0.6 to select segments with high cooperability of interference components as primary labeled segments of latent physical perturbation; Step S135: Assemble the implicit physical perturbation primary annotation fragments to obtain implicit physical perturbation primary annotation data.

4. The grinding parameter optimization method for a fully automated coarse and fine grinding integrated machine based on particle swarm optimization according to claim 2, characterized in that, Step S14 is as follows: Step S141: Group and label the data according to the modal channel type of the primary annotation data of the implicit physical disturbance to obtain thermal flow field segment, vibration frequency band segment, current disturbance segment and acoustic spectrum fluctuation segment; Step S142: Assign the thermal flow field segment, vibration frequency band segment, current disturbance segment, and acoustic spectrum fluctuation segment to the corresponding preset thread pools, and activate the four types of parallel processing threads to generate the initial channel set of the multi-threaded data stream; Step S143: Construct a standard time segment index for the perturbation segment data in each thread of the initial channel set of the multi-threaded data stream, set the single index time window to 0.5s, and perform sliding alignment on the perturbation segment data in each thread to remove segments with cross-channel time drift greater than 200ms, and generate a synchronized time index segment set. Step S144: Set the aggregation window step size to 0.5s, perform thread aggregation operation on the synchronized time index fragment set within each aggregation window, extract the cross-channel cooperative perturbation structure, and generate a set of perturbation thread fusion nodes; Step S145: Serialize and integrate the set of perturbation thread fusion nodes, and assign a unique structural label to each perturbation thread fusion node to obtain a set of perturbation event structure vectors.

5. The grinding parameter optimization method for a fully automated coarse and fine grinding integrated machine based on particle swarm optimization according to claim 1, characterized in that, The multimodal grinding behavior feature analysis in step S2 is specifically as follows: The structured grinding field sensing data is expanded in the time domain, and the time frame is reconstructed synchronously according to the grinding wheel linear velocity in the grinding parameter group to obtain the grinding field sensing time frame set. Acquire abrasive structure data, and analyze the grinding ratio variation trend based on abrasive structure data and grinding parameter set to obtain grinding ratio variation trend data in the cutting zone; The grinding field sensing time frame set is combined with the grinding ratio change trend data in the cutting zone to perform multi-modal time synchronization alignment of heat flow, current, vibration and acoustic spectrum, and generate a modal reconstruction sequence set; The contact arc region perturbation features of the modal reconstruction sequence group are identified, and the transient components and energy mutation points of the modal perturbation are extracted. The perturbation significance threshold was set to 0.

85. Based on the transient components and energy mutation points, the characteristics of abrasive grain breakage, shedding and local burns during the grinding process were identified, thereby forming a grinding modal response significance map. Based on the structured grinding field sensing data, the cutting energy characteristics of each mode are extracted to obtain the cutting force fluctuation frequency, the sound pressure surge rate in the grinding interval, and the second derivative of the spindle current. A cross-modal energy co-existing tensor is constructed by integrating the cutting force fluctuation frequency, the sound pressure surge rate in the grinding interval, and the second derivative of the spindle current, and combining them with the grinding modal response significance spectrum. Behavioral pattern aggregation and boundary identification are performed based on cross-modal energy cooperative tensor to obtain a set of time-series grinding behavior fragments; The time-series grinding behavior fragment set is mapped to a pre-connected grinding condition behavior library, and the behavior sequence morphological similarity is calculated to obtain grinding behavior pattern data.

6. The grinding parameter optimization method for a fully automated coarse and fine grinding integrated machine based on particle swarm optimization according to claim 1, characterized in that, The semantic abstraction of grinding parameters in step S2 is specifically as follows: Initial screening of grinding behavior pattern data by behavior-parameter mapping was performed to construct a potential correlation matrix between behavior segments and grinding parameters; A local sensitivity threshold of 0.75 was set, and the maximum mutual information criterion analysis was performed on the potential association matrix to identify significant behavioral response regions, thereby generating candidate cross-tabulations for parameters. Dynamic features of grinding behavior-related parameters are extracted from the parameter candidate cross-tabulation, and the parameter information density distribution of the dynamic features of grinding behavior-related parameters is evaluated to obtain a semantic weight map of grinding parameters. Perform behavioral semantic projection operation on the high-frequency parameter anomaly patterns in the grinding parameter semantic weight map, and transform the parameter anomaly patterns into semantic behavioral labels including contact instability segment, local overload segment and thermal drift risk segment. Construct a multi-nested index structure between semantic behavior labels and grinding parameter groups to generate a grinding parameter behavior intent mapping cluster; The grinding parameter behavior intent mapping cluster is converted into a multi-dimensional nested structure for parameter dimension semantic compression and upper and lower bit encoding, resulting in a grinding parameter intent nested mapping table.

7. The grinding parameter optimization method for a fully automated coarse and fine grinding integrated machine based on particle swarm optimization according to claim 1, characterized in that, Step S3 is as follows: Step S31: Tensor expansion of the grinding parameter intention nested mapping table to construct the initial parameter tensor field; Step S32: Based on the upper and lower boundary combinations of each parameter dimension in the initial parameter tensor field, construct the constraint space boundary set, and perform non-uniform perturbation on the constraint space boundary set to obtain the coarse solution space structure mesh. Step S33: Initialize the particle swarm space of the coarse solution spatial structure mesh, and encode the initial parameter path of each particle in the particle swarm to obtain the initialized basic particle swarm; Step S34: Perform coarse-grained search evolution on the initial basic particle swarm to generate a family of coarse-grained grinding parameter trajectories.

8. The grinding parameter optimization method for a fully automated coarse and fine grinding integrated machine based on particle swarm optimization according to claim 7, characterized in that, Step S34 is as follows: Step S341: Map the initial parameter path of the initialization basic particle swarm to the corresponding constraint space, and calculate the position similarity of the particles in the constraint space to assign initial velocity and inertia weights to the initialization basic particle swarm, thereby obtaining the initial velocity particle swarm. Step S342: Use the constraint space to perform random jump updates on the initial velocity particle swarm to obtain the perturbation response particle swarm; Step S343: Calculate the fitness value of the current solution for each particle in the perturbation response particle swarm; Step S344: Based on the fitness value of the current solution, select the local optimum and the global optimum for each particle, and update the particle parameters collaboratively based on the local optimum and the global optimum to obtain the updated particle swarm; Step S345: Return the updated particle swarm to step S342. Perform 10 iterations, with each iteration consisting of steps S342 to S345. Based on the solution and fitness evaluation results of each particle in the particle swarm, perform group collaborative clustering to group the particles and output the coarse-grained grinding parameter trajectory family.

9. The grinding parameter optimization method for a fully automated coarse and fine grinding integrated machine based on particle swarm optimization according to claim 1, characterized in that, The soft clustering bias learning in step S4 specifically involves: The heterogeneous trajectory map encoding vector set is subjected to feature domain normalization mapping to generate uniform scale parameter space data; Biased fuzzy clustering is performed based on spatial data with uniform scale parameters to obtain biased fuzzy initial membership sets; Iterative center transfer learning is performed based on the biased fuzzy initial membership set, with the center drift rate constrained to not exceed 0.05 in each iteration, to obtain the class center transfer trajectory set; A dynamic stability assessment is performed on the class center transition trajectory set, and classes whose center trajectory variance exceeds the stability threshold of 0.08 are removed. The remaining class centers are then fitted and modeled to generate a convergent class center space. Based on the convergence class center space, the multi-class membership distribution of each trajectory vector is calculated in reverse to construct a multi-membership structure graph. Then, based on the multi-membership structure graph, a confidence scalar weight is assigned to each encoded vector in the heterogeneous trajectory graph encoding vector set to obtain a high-confidence grinding parameter soft clustering structure. The high-confidence grinding parameter soft clustering structure is subjected to similarity map merging and nested graph compression to output a high-dimensional grinding parameter candidate map.

10. The grinding parameter optimization method for a fully automated coarse and fine grinding integrated machine based on particle swarm optimization according to claim 1, characterized in that, Step S5 is as follows: Step S51: Extract the grain size parameter, cutting depth, linear velocity, contact arc length and grinding ratio corresponding to each candidate parameter structure from the high-dimensional grinding parameter candidate map, and construct the initial working condition state tensor; Step S52: Perform multiphysics simulation based on the initial working condition state tensor, perform continuous simulation for 12 seconds for each set of parameters, and evaluate the physical indicators of grinding ratio, surface roughness, heat-affected zone width and grinding energy consumption respectively to obtain working condition simulation response data. Step S53: Normalize the operating condition simulation response data, set the index weighting group [0.35, 0.25, 0.2, 0.2], calculate the performance score of each group of candidate parameters, and generate a candidate parameter performance evaluation table; Step S54: Based on the candidate parameter performance evaluation table, set the score threshold to 0.82 to filter a subset of stable parameters; Step S55: Perform dimensional loop mapping and redundancy removal on the stable parameter subset to construct a stable grinding parameter loop tensor set, and perform dimension-wise convergence evaluation on the stable grinding parameter loop tensor set. Set the stability fluctuation threshold to ±3% to eliminate periodic loop drift parameter combinations and generate a stable grinding parameter set.

Images