Cascade correction method for weak abnormal sensing data of composite coordinate grinding machine

Abstract

This application discloses a cascaded correction method for weak anomaly sensing data of a composite coordinate grinding machine, relating to the field of equipment sensor data processing technology. The method includes: real-time acquisition of multi-source joint sensing time-series data of the composite coordinate grinding machine; obtaining the original sensing sequence to be corrected through a sliding window of preset length; inputting the original sensing sequence into a sensing prediction network to obtain a predicted sensing sequence of the same length as the original sensing sequence; splicing the two together, and outputting an anomaly confidence sequence of the same length as the spliced ​​sequence through an anomaly identification network; finally, using a consensus correction network, employing the anomaly confidence as a differentiable continuous weighting coefficient, performing point-by-point adaptive weighted fusion of the original sensing sequence and the predicted sensing sequence, and outputting the corrected data. This application resolves the inherent contradiction between missed detections of weak anomalies and false alarms of normal fluctuations, overcomes the technical bottleneck of prediction accuracy limiting correction effectiveness, and can effectively ensure accurate perception of the processing status of the composite coordinate grinding machine.

Description

Technical Field

[0001] This application relates to the field of equipment sensor data processing technology, and in particular to a cascaded correction method for weak abnormal sensor data of a composite coordinate grinding machine. Background Technology

[0002] Composite coordinate grinding machines are core equipment in the precision and ultra-precision machining fields, often referred to as "industrial mother machines." Their positioning accuracy directly determines the manufacturing ceiling for critical components in aerospace and optical devices. To ensure micron-level or even sub-micron-level machining accuracy, these machine tools are equipped with multi-source sensors for temperature, stress, and other parameters, achieving dynamic accuracy assurance through fully closed-loop feedback control. However, under long-term continuous machining conditions, sensors are inevitably affected by factors such as thermal radiation, cutting fluid erosion, and adhesive layer aging, leading to subtle anomalies such as temperature drift, time drift, and creep. These anomalies initially have amplitudes below the noise floor, evolve slowly, and are highly concealed, making them easily misinterpreted as normal fluctuations. If they are not identified and corrected in time, these subtle anomalies will propagate and accumulate within the closed-loop control system, resulting in an inability to accurately perceive the machining status of the composite coordinate grinding machine.

[0003] Regarding the problem of anomaly correction in sensor data, existing technologies suffer from three core defects: First, traditional threshold methods rely on manually defined empirical limits, leading to long-term underreporting of weak anomalies due to insufficient amplitude. Lowering the threshold, on the other hand, triggers numerous false alarms, failing to balance detection accuracy and anti-interference capability. Second, the prior correction function method assumes that sensor data follows a fixed evolutionary law; once actual operating conditions deviate from this assumption, correction fails, resulting in extremely poor robustness. Third, the existing intelligent prediction and substitution method's "prediction-threshold-substitution" paradigm has a fundamental flaw. Discrete threshold switching causes abrupt changes in the correction boundary, disrupting the temporal smoothness of sensor data. Simultaneously, the prediction network input is easily contaminated by anomalies, and the fragmented optimization of the prediction and correction networks makes it impossible to simultaneously achieve "accurate prediction" and "good correction." Summary of the Invention

[0004] The purpose of this application is to provide a cascaded correction method for weak anomaly sensing data of a composite coordinate grinding machine, which can solve the inherent contradiction between underreporting of weak anomalies and false alarms of normal fluctuations, and address the temporal misalignment problem between joint correction prediction and the original sequence.

[0005] To achieve the above objectives, this application provides the following solution: A cascaded correction method for weak anomaly sensing data of a composite coordinate grinding machine includes the following steps: The multi-source joint sensing time series data of the composite coordinate grinding machine is acquired in real time, and the original sensing sequence is extracted from the multi-source joint sensing time series data of historical time periods according to the sliding window of preset length; the multi-source joint sensing time series data includes temperature sensing time series data and stress sensing time series data.

[0006] The original sensing sequence is input into the trained sensing prediction network, which outputs a predicted sensing sequence of the same length as the original sensing sequence.

[0007] The original sensor sequence and the predicted sensor sequence are concatenated. The concatenated sequence is then input into the trained anomaly detection network, which outputs an anomaly confidence sequence of the same length as the concatenated sequence. The anomaly confidence sequence takes the value of a continuous floating-point number from 0 to 1 at each time step.

[0008] The original sensing sequence, the predicted sensing sequence, and the anomaly confidence sequence are input into the trained consensus correction network, which outputs the original sensing sequence after cascade correction. In the consensus correction network, the anomaly confidence sequence is used as a differentiable continuous weighting coefficient to perform point-by-point adaptive weighted fusion of the original sensing sequence and the predicted sensing sequence, thereby realizing the correction of the sensing data of the original sensing sequence.

[0009] Optionally, the cascaded correction method for weak anomaly sensing data of the composite coordinate grinding machine further includes the following steps: Construct a sensor prediction network, an anomaly identification network, and a consensus correction network.

[0010] The sensor prediction network, anomaly detection network, and consensus correction network are jointly trained end-to-end on a labeled dataset. The labeled dataset includes several samples, each of which includes a historical sensor sequence, a predicted sensor sequence corresponding to the historical sensor sequence, an anomaly confidence sequence, and a corrected sensor sequence. In the end-to-end joint training, the sensor prediction network simultaneously receives gradient backpropagation from both the prediction loss and the correction loss, and the weight of the correction loss is set higher than that of the prediction loss. The prediction loss is the loss value calculated based on the prediction output generated by the sensor prediction network with the historical sensor sequence as input and the predicted sensor sequence corresponding to the historical sensor sequence. The correction loss is the loss value calculated based on the correction output of the consensus correction network and the corrected sensor sequence corresponding to the historical sensor sequence.

[0011] After the end-to-end joint training is completed, the trained sensor prediction network, the trained anomaly identification network, and the trained consensus correction network are deployed on the composite coordinate grinding machine CNC system.

[0012] Optionally, after the end-to-end joint training is completed, the trained sensor prediction network, anomaly identification network, and consensus correction network are trimmed and lightweighted before being deployed on the composite coordinate grinding machine CNC system.

[0013] Optionally, temperature sensing time-series data and stress sensing time-series data of the composite coordinate grinding machine are acquired in real time to construct multi-channel, multi-physical quantity, multi-source joint sensing time-series data. Among them, temperature sensing time-series data are acquired by temperature sensors deployed in the spindle bearing, guide rail slider, hydraulic system and cutting zone of the composite coordinate grinding machine, and stress sensing time-series data are acquired by stress sensors deployed in the spindle tool holder interface, worktable support and grinding contact area of ​​the composite coordinate grinding machine.

[0014] Optionally, the lengths of both the original sensing sequence and the predicted sensing sequence are equal to the length of the sliding window, containing data from n sampling times; the original sensing sequence is multi-source joint sensing time-series data from n consecutive sampling times before the current time; the predicted sensing sequence is multi-source joint sensing time-series data from n consecutive sampling times after the current time generated by prediction.

[0015] Optionally, the original sensing sequence and the predicted sensing sequence are concatenated, and the concatenated sequence is input into the trained anomaly detection network to output an anomaly confidence sequence of the same length as the concatenated sequence. This process includes the following steps: The original sensing sequence and the predicted sensing sequence are concatenated to obtain the concatenated sequence.

[0016] The spliced ​​sequence is input into the trained anomaly detection network, which extracts the local deviation features and long-term evolution trend differences between the original sensing sequence and the predicted sensing sequence in the time domain.

[0017] The extracted features are mapped to the 0-1 interval using the Sigmoid activation function, and an anomaly confidence sequence of the same length as the concatenated sequence is output. The magnitude of the values ​​at each time step in the anomaly confidence sequence represents the confidence level of the sensing data that needs to be corrected at the corresponding time step.

[0018] Optionally, the sensing prediction network, anomaly identification network, and consensus correction network all adopt a parallel neural network architecture with dual-path interaction. The parallel neural network architecture with dual-path interaction includes a temporal local feature extraction path, a global cross-modal dependency modeling path, and a cross-level feature interaction gating unit. The temporal local feature extraction path uses dilated causal convolution groups to construct hierarchical receptive fields with an exponentially increasing dilation rate to capture the long-range trend of thermal drift and transient features of stress impact during the machining process of the composite coordinate grinding machine. The global cross-modal dependency modeling path uses a sparse global self-attention module to model the long-range cross-modal dependency relationship of multi-source joint sensing temporal data. The cross-level feature interaction gating unit is set between the temporal local feature extraction path and the global cross-modal dependency modeling path to realize bidirectional feature injection and fusion between the two paths.

[0019] Optionally, the cross-level feature interaction gating unit includes a temporal feature modulation gate and an attention sparse gate; the temporal feature modulation gate receives the intermediate feature map output by the dilated causal convolution group, generates channel weights through a channel-wise sigmoid activation function, and performs dynamic recalibration of the channel dimension of the Value feature output by the sparse global self-attention module; the attention sparse gate receives the attention weight matrix output by the sparse global self-attention module, and performs element-wise weighted modulation of the attention weight matrix with the local response intensity output by the dilated causal convolution group to achieve sparsity of the attention distribution.

[0020] Optionally, after outputting the cascaded corrected sensor data, the cascaded corrected sensor data can be used to replace the original sensor sequence and participate in the thermal error compensation, load monitoring and position closed-loop control of the composite coordinate grinding machine.

[0021] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides a cascaded correction method for weak anomaly sensing data of a composite coordinate grinding machine. The method first acquires real-time multi-source joint sensing time-series data of machine tool temperature and stress, and extracts the original sensing sequence through a preset sliding window to ensure the temporal integrity and processing standardization of the sensing data. This original sensing sequence is then input into a trained sensing prediction network, which outputs a predicted sensing sequence of equal length, providing a stable benchmark for subsequent anomaly identification and correction. Next, the original and predicted sensing sequences are concatenated and input into an anomaly identification network, which outputs an anomaly confidence sequence of equal length with continuous floating-point values ​​of 0-1. Continuous confidence replaces discrete threshold judgment, completely solving the problems of false alarms and abrupt changes in correction boundaries in traditional threshold methods. Finally, the original sequence, predicted sequence, and confidence sequence are input into a consensus correction network, where point-by-point adaptive weighted fusion is performed using confidence as a differentiable continuous weighting coefficient, and a correction sequence is output. Relying on end-to-end joint optimization of the three cascaded networks, this method overcomes the fixed constraints of prior correction functions, solves the problem of fragmented optimization between prediction and correction modules, and comprehensively improves the robustness and accuracy of weak anomaly correction. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 The flowchart illustrates a cascaded correction method for weak abnormal sensor data of a composite coordinate grinding machine, provided as an embodiment of this application.

[0024] Figure 2This is a flowchart of step A3 in a cascaded correction method for weak abnormal sensing data of a composite coordinate grinding machine provided in an embodiment of this application. Detailed Implementation

[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0027] This application provides a cascaded correction method for weak anomaly sensing data in a composite coordinate grinding machine. Addressing the difficulty in identifying and correcting weak anomalies such as temperature drift, time drift, and creep generated by sensors during long-term continuous machining operations of composite coordinate grinding machines, this method constructs an end-to-end cascaded correction architecture with a three-level linkage of "prediction-identification-correction." The weak sensing anomalies described in this embodiment refer to data anomalies caused by sensor temperature drift, time drift, and creep during precision machining in a composite coordinate grinding machine. These anomalies initially have amplitudes below the noise floor, evolve slowly, and are highly concealed, easily misjudged as normal fluctuations and thus uncorrectable. Consequently, they are continuously inherited and amplified during subsequent data acquisition and closed-loop control, hindering the accurate perception of the machine tool's machining status.

[0028] In one exemplary embodiment, such as Figure 1 As shown, it includes the following steps: A1. Real-time acquisition of multi-source joint sensing time series data of composite coordinate grinding machine, and extraction of original sensing sequences from multi-source joint sensing time series data of historical time periods by sliding window of preset length; multi-source joint sensing time series data includes temperature sensing time series data and stress sensing time series data.

[0029] Specifically, real-time temperature and stress sensing time-series data from the composite coordinate grinding machine are acquired to construct a multi-channel, multi-physical-quantity, multi-source joint sensing time-series data system. The temperature sensing time-series data is acquired through temperature sensors deployed in the spindle bearings, guide rails, hydraulic system, and cutting zone of the composite coordinate grinding machine. The stress sensing time-series data is acquired through stress sensors deployed in the spindle tool holder interface, table support, and grinding contact area of ​​the composite coordinate grinding machine. The temperature sensors have a sampling frequency of 10Hz, and the stress sensors have a sampling frequency of 1000Hz. Both are different time slices from the same data source and using the same format specification. Their data structure, channel dimensions, physical quantity types, and sampling frequencies are completely identical, differing only in time span, ensuring matching input and output dimensions of the sensing prediction network.

[0030] As an optional implementation, the length of the original sensing sequence to be corrected is equal to the length of the sliding window, and it contains data from n sampling times; the original sensing sequence is multi-source joint sensing time series data from n consecutive sampling times prior to the current time period.

[0031] A2. The original sensing sequence is input into the trained sensing prediction network, which outputs a predicted sensing sequence of the same length as the original sensing sequence. The sensing prediction network predicts future sensing sequences based on the original sensing sequence, reflecting the expected evolution trend of each sensing channel of the composite coordinate grinding machine under normal operating conditions. In this embodiment, the input to the prediction network is multi-source joint sensing time-series data for n time periods prior to the current time, and the output is multi-source joint sensing time-series data for n time periods after the current time. By extrapolating the evolution trend of sensing data under normal operating conditions using historical time-series data, a benchmark reference is provided for subsequent anomaly identification and correction. In this embodiment, n can be 30, meaning the preset sliding window length is 30.

[0032] A3. The original sensor sequence and the predicted sensor sequence are concatenated. The concatenated sequence is then input into the trained anomaly detection network, which outputs an anomaly confidence sequence of the same length as the concatenated sequence. The anomaly confidence sequence takes the value of a continuous floating-point number from 0 to 1 at each time step. In this embodiment, as... Figure 2 As shown, step A3 specifically includes the following steps: A31. The original sensing sequence and the predicted sensing sequence are concatenated to obtain the concatenated sequence.

[0033] A32. Input the spliced ​​sequence into the trained anomaly detection network, and extract the local deviation features and long-term evolution trend differences between the original sensing sequence and the predicted sensing sequence in the time domain through the anomaly detection network.

[0034] A33. The extracted features are mapped to the 0 to 1 interval using the Sigmoid activation function, and an anomaly confidence sequence of the same length as the concatenated sequence is output. The magnitude of the values ​​at each time step in the anomaly confidence sequence represents the confidence level of the sensing data that needs to be corrected at the corresponding time step.

[0035] Each moment in this anomaly confidence sequence is a continuous floating-point value between 0 and 1, rather than a discrete 0 / 1 label that has not undergone threshold binarization. The magnitude of the value directly represents the confidence level of the sensor data at that moment that needs to be corrected: the closer the value is to 1, the more confident the network is that the original sensor data at that moment is abnormal and more future predictions should be adopted to correct it; the closer the value is to 0, the more confident the network is that the original sensor data at that moment is normal and the original sensor data should be preserved to the greatest extent possible.

[0036] A4. Input the original sensing sequence, the predicted sensing sequence, and the anomaly confidence sequence into the trained consensus correction network, and output the original sensing sequence after cascade correction. In the consensus correction network, the anomaly confidence sequence is used as a differentiable continuous weighting coefficient to perform point-by-point adaptive weighted fusion of the original sensing sequence and the predicted sensing sequence, thereby realizing the correction of the sensing data of the original sensing sequence.

[0037] Specifically, in the consensus correction network, the abnormal confidence sequence is used as a differentiable continuous weighting coefficient to perform point-by-point adaptive weighted fusion of the original and predicted sensor values ​​at the same time. The sensor data of the original sensor sequence at each time step is used to determine whether to correct it based on the subsequent predicted sensor data without abnormalities, according to the abnormal confidence at the corresponding time step. When the abnormal confidence is in the middle region between 0 and 1, the correction output is a continuous proportional mixture of the original value and the predicted value. There is no discrete threshold switching or step change throughout the process, which not only ensures the smooth correction of abnormal data, but also achieves lossless preservation of the original information of normal data.

[0038] Specifically, in this embodiment, after outputting the cascaded corrected sensor data, the cascaded corrected sensor data is used to replace the original sensor sequence and participate in the thermal error compensation, load monitoring and position closed-loop control of the composite coordinate grinding machine.

[0039] In an exemplary embodiment, before performing steps A1 to A4 above, the method further includes the following steps: B1. Construct a sensor prediction network, anomaly detection network, and consensus correction network. Specifically, in this embodiment, the sensor prediction network, anomaly detection network, and consensus correction network all adopt a parallel neural network architecture with dual-path interaction. The main architecture of the three networks is consistent, with only the output head and training optimization target differing, ensuring that gradients can be smoothly propagated back during end-to-end joint training.

[0040] The parallel neural network architecture with dual-path interaction consists of a temporal local feature extraction path, a global cross-modal dependency modeling path, and a cross-level feature interaction gating unit. The temporal local feature extraction path uses dilated causal convolutional groups to construct hierarchical receptive fields with an exponentially increasing dilation rate, capturing the long-range trend of thermal drift and transient features of stress impact during the machining process of a composite coordinate grinding machine. The causal convolution strictly follows the causal constraints of the temporal task to prevent future information leakage. The global cross-modal dependency modeling path uses sparse global self-attention modules to model the long-range cross-modal dependencies of multi-source joint sensing temporal data, capturing the coupling correlation characteristics between sensing data of different physical quantities, and making up for the deficiency of the long-range dependency modeling capability of the dilated causal convolutional group.

[0041] The cross-level feature interaction gating unit connects the temporal local feature extraction path and the global cross-modal dependency modeling path, enabling bidirectional feature injection and fusion between the two paths. This allows the local modal features extracted by the dilated causal convolution group to guide the sparsification of the attention matrix of the sparse global self-attention module, while the attention weights of the sparse global self-attention module dynamically adjust the convolution kernel response field of the dilated causal convolution group. The cross-level feature interaction gating unit comprises two sub-modules: a temporal feature modulation gate and an attention sparse gate. The temporal feature modulation gate receives the intermediate feature map output by the dilated causal convolution group, generates channel weights through a channel-wise sigmoid activation function, and dynamically calibrates the channel dimension of the Value feature output by the sparse global self-attention module, strengthening the contribution of effective feature channels and suppressing interference from redundant feature channels. The attention sparse gate receives the attention weight matrix output by the sparse global self-attention module, and modulates the attention weight matrix element-wise based on the local response intensity output by the dilated causal convolution group, achieving sparsification of the attention distribution, reducing the computational complexity of the sparse global self-attention module, and improving the model's ability to focus on local anomalous features.

[0042] As an exemplary embodiment, the input of the sensing prediction network is the original sensing sequence, which is a multi-source joint sensing time series data without abnormal contamination captured by a preset sliding window, containing temperature and stress dual physical quantity channels. The network first performs time alignment and channel normalization preprocessing on the input original sensing sequence, and then constructs a hierarchical receptive field with an exponentially increasing expansion rate through dilated causal convolution groups of the time series local feature extraction path. Simultaneously, it extracts long-range trend features of thermal drift and local transient features of stress impact. At the same time, it prevents future information leakage through causal convolution constraints. Through a sparse global self-attention module for global cross-modal dependency modeling path, it models the long-range cross-modal coupling correlation features of multi-source sensing data. The dual-path features are bidirectionally fused through a cross-level feature interaction gating unit. The time series feature modulation gate performs channel dimension dynamic calibration on the Value feature of the sparse global self-attention module. The attention sparse gate weights the attention weight matrix to achieve attention sparsity. Finally, the fused features are output by the time series prediction decoding layer as a predicted sensing sequence of the same length as the original sensing sequence.

[0043] The anomaly detection network takes as input the concatenated sequence of the original sensing sequence and the predicted sensing sequence. The network first performs temporal synchronization and feature normalization preprocessing on the concatenated sequence. Then, it extracts the local deviation and temporal gradient features of the original sensing sequence and the predicted sensing sequence through the dilated causal convolution group of the temporal local feature extraction path. It extracts the long-range trend difference between the two sequences and the cross-physical quantity anomaly correlation features through the sparse global self-attention module of the global cross-modal dependency modeling path. The dual-path features are deeply fused through the cross-level feature interaction gating unit to enhance the anomaly sensitive features and suppress the normal fluctuation redundant features. Finally, the fused features are mapped to the continuous interval of 0~1 through the anomaly confidence decoding layer. The output is an anomaly confidence sequence of the same length as the input concatenated sequence. This sequence uses continuous floating-point values ​​to represent the degree of anomaly correction of the data at the corresponding time. The subsequent use mainly uses a portion of the anomaly confidence that overlaps with the time of the original sensing sequence.

[0044] The consensus correction network takes as input a multi-source fusion input sequence, which is a concatenation of the original sensor sequence, the predicted sensor sequence, and the anomaly confidence sequence. The network first performs temporal alignment and weight normalization preprocessing on the multi-source fusion input sequence. Then, it extracts local temporal distortions and temporal smoothing features of the confidence weights from the sensor data using dilated causal convolutional groups in the temporal local feature extraction path. It models the global adaptation and multi-physical quantity correction coupling features between the confidence weights and the sensor data using a sparse global self-attention module in the global cross-modal dependency modeling path. The dual-path features are fused through a cross-level feature interaction gating unit to generate standardized correction features. Finally, the fused features are passed through an adaptive weighted fusion decoding layer, which uses the anomaly confidence as a differentiable continuous weighting coefficient to adaptively weight and fuse the original sensor sequence and the predicted sensor sequence at the same time point-by-point, achieving sensor data correction for the original sensor sequence. The output is the cascaded corrected original sensor sequence, which can be directly applied to thermal error compensation, load monitoring, and position closed-loop control of a composite coordinate grinding machine.

[0045] B2. Perform end-to-end joint training on a labeled dataset for the sensor prediction network, anomaly detection network, and consensus correction network. The labeled dataset includes several samples, each containing a historical sensor sequence, a corresponding predicted sensor sequence, an anomaly confidence sequence, and a corrected sensor sequence. In the end-to-end joint training, the sensor prediction network simultaneously receives gradient backpropagation from both the prediction loss and the correction loss, with the correction loss having a higher weight than the prediction loss. The prediction loss is calculated by combining the prediction output of the sensor prediction network (using the historical sensor sequence as input) with the corresponding predicted sensor sequence. The correction loss is calculated by combining the correction output of the consensus correction network with the corrected sensor sequence corresponding to the historical sensor sequence. The anomaly confidence sequence in the samples is used to calculate the detection loss based on the output of the anomaly detection network.

[0046] In end-to-end joint training, the sensor prediction network simultaneously receives gradient backpropagation from both the prediction loss and the correction loss, and sets the weight of the correction loss to be higher than that of the prediction loss. This cross-network gradient propagation mechanism forces the sensor prediction network to abandon the traditional goal of simply fitting the original sensor sequence, and instead learn to generate feature representations that are more conducive to the anomaly detection network in distinguishing between normal and abnormal, and more conducive to the consensus correction network in achieving smooth linear weighted fusion of the original and predicted values ​​with the anomaly confidence as a continuous weighting coefficient. This fundamentally solves the problem of accuracy reduction caused by anomaly contamination of the prediction network input, and breaks through the technical bottleneck of prediction accuracy restricting the correction effect.

[0047] The training process employs a sliding window mechanism identical to that used in online inference: for a given sample, the original sensor data with anomalies from time t-30 to t-1 in the dataset is used as the input to the sensor prediction network; the manually corrected sensor data without anomalies from time t to t+29 is used as the predicted output label of the sensor prediction network; and the manually corrected sensor data without anomalies from time t-30 to t-1 is used as the corrected sensor sequence. This ensures that the training objective is fully matched with the actual application objective and avoids distributional shifts between training and inference.

[0048] B3. After the end-to-end joint training is completed, the trained sensor prediction network, the trained anomaly identification network, and the trained consensus correction network are deployed on the composite coordinate grinding machine CNC system.

[0049] In another exemplary embodiment, after end-to-end joint training is completed, the trained sensor prediction network, anomaly detection network, and consensus correction network are pruned and lightweighted before being deployed on the composite coordinate grinding machine CNC system. The lightweighting process includes removing redundant parameters and computational branches, merging operators to optimize inference speed, and meeting the real-time computing power requirements of the CNC system.

[0050] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application constructs a three-level linked end-to-end correction architecture of "prediction-identification-correction," expanding the processing of abnormal sensor data from the traditional "threshold determination + prediction substitution" to a new paradigm of "continuous confidence weighting + adaptive fusion." This fundamentally eliminates the step abrupt change problem at the correction boundary, ensuring the temporal smoothness of the corrected sensor data. Through a consensus correction mechanism using the anomaly confidence as a differentiable weighting coefficient, it achieves point-by-point adaptive soft fusion of the original sensor value and the predicted sensor value, realizing the correction of the original sensor sequence and smoothly transitioning to the predicted value in the anomaly segment. The normal segment retains the original information without loss, which solves the inherent contradiction between the "missed detection" in the early stage of weak anomalies and the "false detection" of normal fluctuations in existing methods. The differentiated joint training strategy of correction accuracy as the main factor and prediction accuracy as the auxiliary factor is adopted, so that the prediction network receives the gradient backpropagation of prediction loss and correction loss at the same time. This forces the prediction network to abandon the traditional goal of simply fitting the original sequence and instead generate feature representations that are more conducive to anomaly identification and smooth fusion. This fundamentally breaks through the technical problem of "correct correction collapse due to prediction inaccuracy" and ensures the accurate perception of the processing status of the composite coordinate grinding machine.

[0051] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0052] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0053] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0054] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0055] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A cascaded correction method for weak anomaly sensing data of a composite coordinate grinding machine, characterized in that, include: The multi-source joint sensing time series data of the composite coordinate grinding machine is acquired in real time, and the original sensing sequence is extracted from the multi-source joint sensing time series data of historical time periods according to a sliding window of preset length; the multi-source joint sensing time series data includes temperature sensing time series data and stress sensing time series data. The original sensing sequence is input into the trained sensing prediction network, and a predicted sensing sequence of the same length as the original sensing sequence is output. The original sensing sequence and the predicted sensing sequence are concatenated, and the concatenated sequence is input into the trained anomaly detection network. The network outputs an anomaly confidence sequence of the same length as the concatenated sequence. The anomaly confidence sequence takes the value of a continuous floating-point number from 0 to 1 at each time step. The original sensing sequence, the predicted sensing sequence, and the anomaly confidence sequence are input into the trained consensus correction network, which outputs the cascaded corrected original sensing sequence. In the consensus correction network, the anomaly confidence sequence is used as a differentiable continuous weighting coefficient to perform point-by-point adaptive weighted fusion of the original sensing sequence and the predicted sensing sequence, thereby realizing the correction of the sensing data of the original sensing sequence.

2. The cascaded correction method for weak abnormal sensor data of a composite coordinate grinding machine according to claim 1, characterized in that, Also includes: Construct a sensor prediction network, an anomaly detection network, and a consensus correction network; The sensor prediction network, anomaly detection network, and consensus correction network are jointly trained end-to-end on a labeled dataset. The labeled dataset includes several samples, each of which includes a historical sensor sequence, a predicted sensor sequence corresponding to the historical sensor sequence, an anomaly confidence sequence, and a corrected sensor sequence. In the end-to-end joint training, the sensor prediction network simultaneously receives gradient backpropagations from both the prediction loss and the correction loss, and the weight of the correction loss is set higher than that of the prediction loss. The prediction loss is the loss value calculated based on the prediction output generated by the sensor prediction network with the historical sensor sequence as input and the predicted sensor sequence corresponding to the historical sensor sequence. The correction loss is the loss value calculated based on the correction output of the consensus correction network and the corrected sensor sequence corresponding to the historical sensor sequence. After the end-to-end joint training is completed, the trained sensor prediction network, the trained anomaly identification network, and the trained consensus correction network are deployed on the composite coordinate grinding machine CNC system.

3. The cascaded correction method for weak abnormal sensor data of a composite coordinate grinding machine according to claim 2, characterized in that, After the end-to-end joint training is completed, the trained sensor prediction network, anomaly identification network, and consensus correction network are trimmed and lightweighted before being deployed on the composite coordinate grinding machine CNC system.

4. The cascaded correction method for weak abnormal sensor data of a composite coordinate grinding machine according to claim 1, characterized in that, Real-time temperature and stress sensing time-series data of the composite coordinate grinding machine are acquired to construct multi-channel, multi-physical quantity, multi-source joint sensing time-series data. The temperature sensing time-series data is acquired by temperature sensors deployed in the spindle bearing, guide slide, hydraulic system, and cutting zone of the composite coordinate grinding machine, while the stress sensing time-series data is acquired by stress sensors deployed in the spindle tool holder interface, worktable support, and grinding contact area of ​​the composite coordinate grinding machine.

5. The cascaded correction method for weak abnormal sensing data of a composite coordinate grinding machine according to claim 1, characterized in that, The lengths of both the original sensing sequence and the predicted sensing sequence are equal to the length of the sliding window, and contain data from n sampling times. The original sensing sequence is multi-source joint sensing time-series data from n consecutive sampling times before the current time. The predicted sensing sequence is multi-source joint sensing time-series data from n consecutive sampling times after the current time, generated by prediction.

6. The cascaded correction method for weak abnormal sensor data of a composite coordinate grinding machine according to claim 1, characterized in that, The original sensing sequence and the predicted sensing sequence are concatenated, and the concatenated sequence is input into the trained anomaly detection network. The output is an anomaly confidence sequence of the same length as the concatenated sequence, specifically including: The original sensing sequence and the predicted sensing sequence are concatenated to obtain the concatenated sequence. The spliced ​​sequence is input into the trained anomaly detection network, which extracts the local deviation features and long-term evolution trend differences between the original sensing sequence and the predicted sensing sequence in the time domain. The extracted features are mapped to the 0-1 interval using the Sigmoid activation function, and an anomaly confidence sequence of the same length as the concatenated sequence is output. The magnitude of the values ​​at each time step in the anomaly confidence sequence represents the confidence level of the sensing data that needs to be corrected at the corresponding time step.

7. The cascaded correction method for weak abnormal sensing data of a composite coordinate grinding machine according to claim 1, characterized in that, The sensor prediction network, anomaly identification network, and consensus correction network all adopt a parallel neural network architecture with dual-path interaction. The parallel neural network architecture with dual-path interaction includes a temporal local feature extraction path, a global cross-modal dependency modeling path, and a cross-level feature interaction gating unit. The temporal local feature extraction path adopts dilated causal convolution groups to construct hierarchical receptive fields with an exponentially increasing dilation rate to capture the long-range trend of thermal drift and transient features of stress impact during the composite coordinate grinding process. The global cross-modal dependency modeling path employs a sparse global self-attention module to model the long-range cross-modal dependencies of multi-source joint sensing time-series data; the cross-level feature interaction gating unit is set between the time-series local feature extraction path and the global cross-modal dependency modeling path to realize bidirectional feature injection and fusion between the two paths.

8. The cascaded correction method for weak abnormal sensing data of a composite coordinate grinding machine according to claim 7, characterized in that, The cross-level feature interaction gating unit includes a temporal feature modulation gate and an attention sparse gate. The temporal feature modulation gate receives the intermediate feature map output by the dilated causal convolution group, generates channel weights through a channel-wise sigmoid activation function, and performs dynamic recalibration of the channel dimension of the Value feature output by the sparse global self-attention module. The attention sparse gate receives the attention weight matrix output by the sparse global self-attention module, and performs element-wise weighted modulation of the attention weight matrix with the local response intensity output by the dilated causal convolution group to achieve sparsity of the attention distribution.

9. The cascaded correction method for weak abnormal sensor data of a composite coordinate grinding machine according to claim 1, characterized in that, After outputting the cascaded corrected sensor data, the cascaded corrected sensor data is used to replace the original sensor sequence and participate in the thermal error compensation, load monitoring and position closed-loop control of the composite coordinate grinding machine.

Images