Cutting force temperature decoupling method and system based on LSTM and PINN neural network

Abstract

The application relates to the field of machine tool cutter monitoring and intelligent machining, and provides a cutting force and temperature decoupling method and system based on LSTM and PINN neural networks. The method comprises the following steps: establishing a physical coupling model of resistance changes of a cutter sensitive area, cutting force and cutting temperature in a cutting process; constructing a hybrid neural network composed of a time sequence feature extraction branch and a physical constraint branch; the time sequence feature extraction branch adopts LSTM, is used for receiving a standardized original cutting voltage time sequence signal, extracting time sequence features and outputting preliminary decoupling; the physical constraint branch adopts PINN, is used for receiving the standardized original cutting voltage time sequence signal, the output of the LSTM, and taking the residual error of the physical coupling model as a physical loss to constrain and correct the output of the LSTM; and a decoupling layer of the hybrid neural network performs weighted fusion on the output of the LSTM and the output of the PINN to obtain a final cutting force signal and a final cutting temperature signal.

Description

Technical Field

[0001] This invention relates to the field of machine tool monitoring and intelligent machining, and in particular to a cutting force-temperature decoupling method and system based on LSTM and PINN neural networks. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] In precision and ultra-precision machining, cutting force and cutting temperature are key variables affecting tool wear, part surface finish, and machining process stability. For measuring cutting force and cutting temperature signals during the cutting process, external force sensors are used to measure cutting force, and thermocouples or infrared thermometers are used to measure cutting temperature. However, external sensors are bulky and have limited response bandwidth. Furthermore, the spatial distance between the external sensor and the cutting point region causes hysteresis in the sensor's mechanical and thermal responses, making it difficult to obtain accurate cutting force and cutting temperature signals under strong transient cutting conditions.

[0004] To address the limitations of external sensors in testing cutting force and temperature signals, the industry has increasingly adopted tool-integrated sensing technology, embedding piezoresistive, piezoelectric, or thermoelectric elements into the tool holder or insert to achieve near-field sensing of the cutting process. However, force and temperature signals are strongly coupled during cutting, and under extreme conditions such as high temperature, high pressure, and severe vibration, the physical properties of sensing materials are prone to drift. Traditional signal preprocessing methods such as filtering and wavelet transform, as well as empirical models, are insufficient for effectively decoupling and analyzing cutting force and temperature signals.

[0005] With the rapid development of deep learning technology, Long Short-Term Memory (LSTM) networks have been widely applied to the processing and analysis of real-time changing signals due to their technical advantages in processing time-series data. LSTM networks can capture the long-term dependence characteristics of signals, and have significant modeling effects on non-stationary and nonlinear force or temperature time-series signals, while also exhibiting good robustness in noisy environments. However, purely data-driven LSTM models lack an accurate reflection of physical laws, and their prediction results are prone to bias if the training data coverage is insufficient or sensor drift is severe.

[0006] On the other hand, Physical Information Neural Networks (PINNs) combine prior physical knowledge with the fitting ability of neural networks by introducing physical constraints such as thermal resistance and piezoresistive effects into the loss function. This demonstrates good generalization ability and physical consistency in fields such as engineering heat conduction and mechanics. However, the PINN method alone relies on accurate physical models and boundary conditions when dealing with highly nonlinear and strongly random transient time-series signals. When actual working conditions are complex and physical modeling is difficult, the convergence speed and accuracy of PINNs are easily limited. Summary of the Invention

[0007] To address the technical problems mentioned above, this invention provides a cutting force-temperature decoupling method and system based on LSTM and PINN neural networks, which can achieve synchronous and high-precision online monitoring of cutting force and cutting temperature under strong transient cutting conditions.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of the present invention provides a cutting force-temperature decoupling method based on LSTM and PINN neural networks.

[0009] A cutting force-temperature decoupling method based on LSTM and PINN neural networks includes: A physical coupling model is established between the change in resistance of the tool's sensitive area and the cutting force and cutting temperature during the cutting process, and physical constraint terms for neural network training are constructed based on this model. A hybrid neural network consisting of a temporal feature extraction branch and a physical constraint branch is constructed. The temporal feature extraction branch uses a Long Short-Term Memory (LSTM) network to receive the standardized original cutting voltage time-series signal, extract temporal features, and output a preliminarily decoupled first cutting force signal and first cutting temperature signal. The physical constraint branch uses a Physical Information Neural Network (PINN) to receive the standardized original cutting voltage time-series signal, the first cutting force signal, and the first cutting temperature signal, and uses the residual of the physical coupling model as a physical loss to constrain and correct the first cutting force signal and the first cutting temperature signal, obtaining a second cutting force signal and a second cutting temperature signal. The decoupling layer of the hybrid neural network performs weighted fusion of the first cutting force signal, the first cutting temperature signal, the second cutting force signal, and the second cutting temperature signal to obtain the final cutting force signal and the final cutting temperature signal.

[0010] Furthermore, a two-stage training strategy is adopted to train the hybrid neural network. The first stage is a pre-training stage that uses only the data loss term, and the second stage is a joint training stage that uses the total loss function of both the data loss term and the physical loss term.

[0011] Furthermore, the total loss function is expressed by the following formula:

[0012]

[0013]

[0014] in, Represents the physical loss function; Represents the loss function for supervised data; , Indicates the weighting coefficient; Indicates the first t Predicted cutting force value at time 1. Indicates the first t Predicted cutting temperature at each moment.

[0015] Furthermore, the physical coupling model is expressed by the following formula:

[0016] in, R meas (t) represents the equivalent resistance of the tool-sensitive area measured at time t; F ( t () represents the equivalent cutting force. T ( t ) represents the equivalent cutting temperature, σ( t ) represents the equivalent conductivity; α is the equivalent piezoresistive coefficient, β is the thermistor coefficient, γ is the force-temperature coupling coefficient, and ε is the reference bias term.

[0017] Furthermore, the original cutting voltage timing signal is a time sequence of the voltage in the tool sensitive area acquired under constant current excitation conditions, which is then input into the hybrid neural network after normalization processing.

[0018] Furthermore, the temporal feature extraction branch employs a two-layer stacked LSTM network, with 64 hidden units in each layer; the physical constraint branch uses the output of the LSTM network as input and participates in network parameter updates through the physical loss function.

[0019] A second aspect of the present invention provides a cutting force-temperature decoupling system based on LSTM and PINN neural networks.

[0020] A cutting force-temperature decoupling system based on LSTM and PINN neural networks includes: The physical coupling model construction module is configured to: establish a physical coupling model between the change in resistance of the tool sensitive area and the cutting force and cutting temperature during the cutting process, and construct physical constraint terms for neural network training based on the model; The hybrid neural network prediction module is configured to: construct a hybrid neural network consisting of a temporal feature extraction branch and a physical constraint branch; the temporal feature extraction branch uses a Long Short-Term Memory (LSTM) network to receive the standardized original cutting voltage time-series signal, extract temporal features, and output a preliminarily decoupled first cutting force signal and first cutting temperature signal; the physical constraint branch uses a Physical Information Neural Network (PINN) to receive the standardized original cutting voltage time-series signal, the first cutting force signal, and the first cutting temperature signal, and uses the residual of the physical coupling model as a physical loss to constrain and correct the first cutting force signal and the first cutting temperature signal to obtain a second cutting force signal and a second cutting temperature signal; the decoupling layer of the hybrid neural network performs weighted fusion of the first cutting force signal, the first cutting temperature signal, the second cutting force signal, and the second cutting temperature signal to obtain the final cutting force signal and the final cutting temperature signal.

[0021] A third aspect of the present invention provides a computer device comprising: A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps in the cutting force-temperature decoupling method based on LSTM and PINN neural networks as described in the first aspect above.

[0022] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program adapted to be loaded by a processor and to execute steps in the cutting force-temperature decoupling method based on LSTM and PINN neural networks as described in the first aspect above.

[0023] The fifth aspect of the present invention provides a computer program product or computer program.

[0024] This invention provides a computer program product or computer program comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps in the cutting force-temperature decoupling method based on LSTM and PINN neural networks as described in the first aspect above.

[0025] Compared with the prior art, the beneficial effects of the present invention are: This invention provides a cutting force-temperature decoupling method and system based on LSTM and PINN neural networks. First, LSTM extracts and performs preliminary separation of the temporal features of the original cutting process signal. Then, PINN uses a physical model to physically constrain and perform secondary correction on the LSTM output. This hybrid neural network can fully utilize the data-driven model's ability to fit transient coupled signals, and also leverages the physical information network to ensure that the prediction results conform to constitutive relations and physical laws. This achieves high-precision, synchronous, online decoupling monitoring of cutting force and cutting temperature, overcoming the shortcomings of existing methods that rely on single applications. It has significant advantages such as high robustness, good real-time performance, and wide applicability to various working conditions. Attached Figure Description

[0026] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0027] Figure 1 This is a flowchart illustrating the cutting force-temperature decoupling method based on LSTM and PINN neural networks, as shown in an embodiment of the present invention. Figure 2 This is a diagram illustrating the model of valence band structure splitting and reconstruction under stress, as shown in an embodiment of the present invention. Figure 3 This is a diagram illustrating the LSTM-PINN hybrid model framework of an embodiment of the present invention; Figure 4 This is a schematic diagram of a long short-term memory neural network shown in an embodiment of the present invention; Figure 5 This is a schematic diagram of a physical information neural network shown in an embodiment of the present invention; Figure 6 This is a schematic diagram illustrating the relationship between resistance and cutting force at T=0, as shown in an embodiment of the present invention. Figure 7 This is a schematic diagram illustrating the relationship between resistance and cutting temperature when F=0, as shown in an embodiment of the present invention. Figure 8 This is a statistical diagram of the cutting force prediction decoupling error shown in an embodiment of the present invention; Figure 9 This is a statistical diagram of the cutting temperature prediction decoupling error shown in an embodiment of the present invention; Figure 10 This is a structural diagram of the cutting force-temperature decoupling system based on LSTM and PINN neural network shown in an embodiment of the present invention; Figure 11 This is a structural diagram of a computer device shown in an embodiment of the present invention. Detailed Implementation

[0028] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0029] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0030] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0031] Figure 1 This is a flowchart illustrating the cutting force-temperature decoupling method based on LSTM and PINN neural networks, as shown in an embodiment of the present invention; see reference. Figure 1 The method includes: A physical coupling model is established between the change in resistance of the tool's sensitive area and the cutting force and cutting temperature during the cutting process, and physical constraint terms for neural network training are constructed based on this model. A hybrid neural network consisting of a temporal feature extraction branch and a physical constraint branch is constructed. The temporal feature extraction branch uses a Long Short-Term Memory (LSTM) network to receive the standardized original cutting voltage time-series signal, extract temporal features, and output a preliminarily decoupled first cutting force signal and first cutting temperature signal. The physical constraint branch uses a Physical Information Neural Network (PINN) to receive the standardized original cutting voltage time-series signal, the first cutting force signal, and the first cutting temperature signal, and uses the residual of the physical coupling model as a physical loss to constrain and correct the first cutting force signal and the first cutting temperature signal, obtaining a second cutting force signal and a second cutting temperature signal. The decoupling layer of the hybrid neural network performs weighted fusion of the first cutting force signal, the first cutting temperature signal, the second cutting force signal, and the second cutting temperature signal to obtain the final cutting force signal and the final cutting temperature signal.

[0032] In this embodiment, LSTM is first used to extract and initially separate the temporal features of the original cutting process signal. Then, PINN is used to apply physical constraints and secondary corrections to the LSTM output using a physical model. This hybrid neural network can fully exploit the data-driven model's ability to fit transient coupled signals, and also ensure that the prediction results conform to constitutive relations and physical laws by leveraging the physical information network. It achieves end-to-end decoupling mapping from a single-channel voltage signal to a bivariate force-temperature signal, eliminating the need for complex multi-sensor configurations. The system structure is simple and suitable for online monitoring under strong transient cutting conditions.

[0033] The cutting force-temperature decoupling method based on LSTM and PINN neural network described in this embodiment is described in detail below, including the following steps: Step 1: Physical Modeling After achieving surface semiconductorization of single-crystal diamond tools through boron-sulfur co-doping, the tool tip sensitive region can output an electrical signal containing force and temperature information. During cutting, force and thermal loads typically act simultaneously on the doped modified layer. Therefore, the resistance signal simultaneously includes piezoresistive and thermal resistance effects, exhibiting coupling, nonlinearity, and time-varying characteristics. To ensure the physical consistency and stability of subsequent decoupling results under varying operating conditions, a physical expression for resistance change needs to be established and transformed into a "soft constraint" term in the neural network training process. Simultaneously, a time-series model capable of handling sequential memory and multi-timescale features is introduced to adapt to the dynamic characteristics of the force-temperature coupled signal.

[0034] According to stress crystal field theory, the introduction of stress causes the degenerate states near the valence band apex to split and reconstruct, especially in hole-dominated doped diamond. Small changes in the valence band structure alter the equivalent mass and density of states, thus amplifying the conductivity's sensitivity to stress. Figure 2 As shown. At the quantitative level, the electrical conductivity of a material can be characterized by the Boltzmann transport equation (1), which reflects the combined influence of factors such as band structure, Fermi distribution and mobility on electrical transport.

[0035] (1) in, q Represents the elementary charge. E c This represents the lower integral limit or transport energy level boundary of the transport energy band. D ( E ) represents energy E Density of states at that point μ ( E ) represents energy E carrier mobility at that location f ( E ) represents the Fermi-Dirac distribution function.

[0036] The above equation can be written as a weighted average to obtain equation (2).

[0037] (2) In equation (2), It represents the equivalent average mobility of transport-related energy. This represents the equivalent density of states contributing to transport. and N T The expressions are equations (3) and (4).

[0038] (3) (4) When only a force load is applied and the temperature is constant, the stress or strain introduced by the external load will change the band structure and scattering process, thereby... and As the electrical conductivity changes, it becomes a load variable. X The function, i.e. For small load disturbances in X Expanding the equation to the vicinity of 0, we get equation (5).

[0039] (5) Where σ0 represents the initial conductivity when the reference load state X=0.

[0040] The equivalent resistance of the tool-sensitive area satisfies equation (6).

[0041] (6) in, R ( X ) represents the equivalent resistance of the tool sensitive area under the action of load variable X.

[0042] In equation (6), L and A These represent the equivalent conductive path length and cross-sectional area, respectively. Under the small deformation approximation, geometric changes can be neglected. R ( X After performing first-order linearization, we obtain equation (7).

[0043] (7) Will Substituting into equation (7), we obtain equation (8).

[0044] (8) Where, Δ R F This represents the resistance increment caused solely by mechanical load.

[0045] In equation (8), The equivalent piezoresistive coefficient is used to characterize the sensitivity of the resistor to force, as given by equation (9).

[0046] (9) Equivalent cutting force F Represents force load, where X = kF Substituting this mapping relationship into equation (10) yields the result.

[0047] (10) The displacement mechanism of light vacancy bands under stress and its modulation of carrier mobility are key physical mechanisms of the piezoresistive response of diamond materials. By combining valence band splitting theory and Boltzmann transport equations, an interpretable piezoresistive model is constructed, which not only provides a solid physical foundation for the design of decoupling algorithms, but also lays a theoretical basis for improving prediction accuracy and model generalization ability.

[0048] To describe the effect of temperature on resistance, according to the Fermi-Dirac distribution, under the non-degenerate approximation, the free hole concentration satisfies the Boltzmann-type thermal activation relation.

[0049] (11) In equation (11), Hole concentration, For the valence band effective states density, The activation energy for doping, k Boltzmann's constant, T The absolute temperature is then used. The conductivity is further calculated using equation (12).

[0050] (12) In equation (12), q is the elementary charge. This represents the migration rate.

[0051] Depend on The form of resistivity change with temperature can be obtained (13).

[0052] (13) resistance R ( T ) and resistivity If they are directly proportional, then we can obtain R ( T The expression for ) is (14).

[0053] (14) This expression reflects two types of temperature effects: increased temperature enhances thermal excitation and increases carrier concentration, but may slightly decrease mobility. For doped diamond, the enhanced conductivity due to thermal excitation dominates within a certain temperature range, therefore the material resistance decreases non-linearly with temperature.

[0054] The thermal resistance effect is linearized to facilitate parameter calibration and online calculation. Using the same method described above, in... X A first-order Taylor expansion is performed near 0, and the following approximate form is obtained to describe the temperature-induced resistance increment as Equation (15).

[0055] (15) The thermal resistance effect of diamond originates from the significant increase in carrier concentration caused by thermal excitation, which is an important response mechanism for the material's self-sensing temperature. Starting from Fermi-Dirac statistics, under the non-degenerate approximation, the Fermi distribution can be approximated as the Boltzmann distribution, thus obtaining the relationship between free carrier concentration and temperature, which clearly elucidates the intrinsic law of thermal response at the theoretical and modeling levels.

[0056] Under constant current excitation, the output voltage and resistance satisfy equation (16).

[0057] (16) In equation (16), I For a constant current, R ( t ) represents the equivalent resistance of the sensitive area at the knife tip.

[0058] because I Since the voltage change is constant, the voltage change is equivalent to the resistance change; this will be consistent in subsequent treatments. R ( t Modeling. The sensitive area of ​​the knife tip can be represented by an equivalent conductive channel, and its resistance can be written as .

[0059] (17) K The expression (18) is shown below.

[0060] (18) In equation (18), L This is the equivalent conductive path length. A This is the equivalent cross-sectional area.

[0061] Under the approximation of small deformation and low thermal expansion, the geometric term... L and A It can be considered a constant, and the load is mainly applied through σ( t The changes manifest themselves. Since stress modulates transport parameters such as mobility, and temperature modulates carrier activation and scattering processes, conductivity can be considered a function of force and temperature.

[0062] (19) In equation (19), F ( t () is the equivalent cutting force. T ( t ) is the equivalent cutting temperature, σ( t () represents the equivalent conductivity.

[0063] To obtain an analytical form that can be used to train constraints, at the reference operating point ( F 0,T 0) near σ( F,T Perform a binary Taylor expansion, retaining the first-order term and one cross term.

[0064] (20) In equation (20), σ0 = σ( F 0 ,T 0), defined as in equation (21).

[0065] (twenty one) make: (twenty two) Then we can obtain equation (23).

[0066] (twenty three) The aforementioned small perturbation in conductivity needs to be further propagated to the resistance expression. Because... R=K / σ , can be obtained in σ First-order linearization near 0.

[0067] (twenty four) make: R 0 =K / σ 0, Δ σ ( t )= σ ( t )- σ 0, we can derive equation (25).

[0068] (25) (26) Substituting equation (25) into equation (26), we can obtain the equivalent decomposition form of the resistance change.

[0069] (27) In equation (27), α, β, γ The specific expression is shown in equation (28).

[0070] (28) The results yield three types of contributions with clear physical meaning. α Δ F The first-order response characterizing the piezoresistive effect β Δ T The first-order response characterizing the thermal resistance effect γ Δ F Δ TThe force-temperature coupling term characterizes the coupling phenomena such as temperature-induced changes in piezoresistive sensitivity and force-induced changes in thermal response amplitude. Considering noise and model truncation errors, the resistance observation can be further written as equation (29).

[0071] (29) in, R meas (t) represents the equivalent resistance of the tool-sensitive area measured at time t; F ( t () represents the equivalent cutting force. T ( t ) represents the equivalent cutting temperature, σ( t ) represents the equivalent conductivity; α is the equivalent piezoresistive coefficient, β is the thermistor coefficient, γ is the force-temperature coupling coefficient, and ε is the reference bias term.

[0072] The role of physical modeling is not only to provide an interpretable response structure, but also to provide soft constraints for PINN training. Therefore, it is written as physical residuals and incorporated into the loss function, and the neural network output is guided to satisfy the physical structure by penalizing the residuals. Let the force and temperature estimates of the neural network output be... , Then, physical residuals can be constructed.

[0073] (30) The physical loss function is defined as equation (31).

[0074] (31) The training data comes from controlled loading and calibration and includes supervised labels. , Construct a supervised data loss function.

[0075] (32) The total loss is weighted as shown in equation (33).

[0076] (33) The total loss function combines the constraints of data training and physical consistency into a single optimization objective. Under this dual-constraint framework, data loss and physical loss play different roles in model convergence, and their weights... λ 1 and λ 2. It can be flexibly allocated according to changes in the dataset to play a dynamic constraint role.

[0077] This invention provides a solid physical basis for decoupling cutting force and temperature by establishing a physical coupling model that includes piezoresistive effect terms, thermal resistance effect terms, and force-temperature coupling terms. This ensures that the decoupling results conform to constitutive relations and physical laws, and overcomes the shortcomings of insufficient physical consistency in purely data-driven models.

[0078] Step 2: Define the neural network framework This hybrid model constructs a two-branch fusion framework consisting of a temporal feature extraction branch and a physical constraint branch, used to achieve decoupled representation of mechanical and temperature information in the tool electrical signal. The framework is as follows: Figure 3 As shown, the LSTM neural network structure is as follows: Figure 4 As shown, the PINN neural network structure is as follows: Figure 5 As shown in the diagram, the raw voltage data is first standardized and then simultaneously fed into two sub-networks. The LSTM branch on the left focuses on mining temporal features, using a forget gate, input gate, and output gate gating mechanism to automatically identify two types of patterns from the non-steady-state sequence: rapid mutation and slow drift. The corresponding outputs are the first set of coarse decoupling results: force signal F1 and temperature signal T1. The PINN branch on the right synchronously receives the standardized voltage sequence at its input and uses the coarse decoupling results (force signal F1 and temperature signal T1) provided by the LSTM as a second data source. During the training phase, the neural network adds the residual of the resistance-force-temperature coupling equation as a physical loss in addition to the regular data error, thereby imposing soft physical constraints on the prediction results. After correction, it outputs high-precision force signal F2 and temperature signal T2. The results from the two branches are fused in the top decoupling layer, and the final cutting force and cutting temperature signals are obtained by weight allocation.

[0079] LSTM is responsible for capturing dynamic features across multiple time scales, while PINN ensures that the output satisfies the piezoresistive-thermal-coupling mechanism. This dual supervision not only improves the generalization stability in small sample and high-noise environments but also ensures the physical consistency of the decoupled results.

[0080] The model input is a time series of voltages in the tool's sensitive area acquired under constant current excitation. Since voltage and resistance satisfy a linear relationship under constant current conditions, the voltage signal directly reflects the dynamic changes in the tool's internal resistance, serving as the input to the force-temperature decoupling model. Let the sliding time window length be... L Then the input of a single sample window can be represented as .

[0081] (34) In equation (34), U t Indicates the first t Voltage value at each sampling time. L This indicates the number of sampling points contained within a single time window.

[0082] To reduce the impact of dimensional differences on network training, the input voltage sequence needs to be normalized before entering the model. The normalized input sequence is as follows.

[0083] (35) In equation (35), x t This is the normalized voltage value.

[0084] This sequence preserves complete dynamic information about the voltage signal changes over time during the cutting process, which forms the basis for subsequent temporal feature extraction and physical constraint modeling.

[0085] The model outputs the predicted cutting force and cutting temperature corresponding to the input time window. At each time step, a two-dimensional state variable is output: (36) In equation (36), For the first t Predicted cutting force value at time 1. For the first t The predicted cutting temperature at each time point. For the entire time window, the model output can be written as:

[0086] (37) The acquired real cutting force and temperature signals are used to supervise training, and their label form is represented as follows.

[0087] (38) (39) The model essentially establishes an end-to-end nonlinear mapping from a single-channel voltage sequence to a bivariate force-temperature sequence, as shown in equation (40).

[0088] (40) In equation (40), Indicated by network parameters θ The resulting hybrid decoupling model.

[0089] This model framework enables an end-to-end mapping from a single-channel voltage sequence to a bivariate force-temperature output.

[0090] This invention employs a dual-branch framework that integrates LSTM and PINN. The LSTM branch fully leverages the data-driven model's ability to fit transient coupled signals, while the PINN branch uses a physical model for constraint and correction. The two complement each other and work synergistically to improve generalization stability in small sample and high-noise environments, while also ensuring the physical consistency of the decoupling results.

[0091] Step 3: Construct the loss function To ensure that the LSTM-PINN hybrid model can accurately learn and train the mapping relationship between voltage sequences and force-temperature labels, this study introduces both data loss function and physical loss function during the training process, and adopts an optimization strategy with dual constraints of data and mechanism.

[0092] Based on the voltage response model established above, under constant current excitation conditions, with reference state ( F 0 ,T If 0) is the reference, then we have equation (41).

[0093] (41) Reconstructed from prediction results It should be consistent with the actual input voltage. Maintain consistency. The physical residual can be defined as equation (42).

[0094] (42) At the supervised training level, the network output consists of predicted cutting force and predicted cutting temperature values ​​corresponding to the input time window, denoted as follows: The actual values ​​of cutting force and cutting temperature collected are denoted as follows: F ( t )and T ( t The data loss function is given by equation (32).

[0095] (32) In equation (32), N This represents the number of valid samples in a batch.

[0096] This loss term directly constrains the network output to approximate the true force and temperature labels, and is the main driving force for the model to learn the mapping relationship between the input voltage sequence and the target variable.

[0097] The physical loss function is given by equation (31).

[0098] (31) This approach does not require the network output to strictly satisfy the physical equations. Instead, it minimizes the physical residuals to gradually converge the prediction results in a direction that satisfies the physical laws.

[0099] Combining data loss and physical loss, the total loss function is constructed as equation (33).

[0100] (33) In equation (33), λ 1 and λ2 represents the weighting coefficients for data loss and physical loss, respectively.

[0101] In this dual-constraint structure, It is responsible for driving the network training to approximate the experimental labels, enabling the model to have basic predictive capabilities, and It is responsible for ensuring that the prediction results have physical consistency.

[0102] Regarding weight allocation, λ 1 and λ 2. The loss directly impacts the model's convergence state. If the data term dominates, the network is prone to over-reliance on the sample distribution, neglecting physical constraints. If the physical term dominates, the network's early training will be subject to strong residual perturbations, affecting the stable learning of temporal features. To maintain coordination between the two types of losses during training, this study employs an adaptive weight adjustment strategy. By monitoring the gradient norm or validation set error corresponding to the two losses, their contributions are kept at the same order of magnitude, and the weights are dynamically updated accordingly. λ 1 and λ 2. This avoids a single loss term dominating parameter updates for a long time, allowing data fitting and prediction capabilities to improve in tandem with physical consistency during training.

[0103] This invention employs an optimization strategy with dual constraints of data loss and physical loss, and introduces an adaptive weight adjustment mechanism to simultaneously improve data fitting ability and physical consistency during training, thus avoiding model bias caused by a single loss term.

[0104] Step 4: Data Training Model training was completed using the PyTorch deep learning framework, with a runtime environment including an NVIDIA RTX 3060 GPU, an Intel i9-12400F processor, and Python 3.9. The main computing libraries called included torch, numpy, scipy, and matplotlib.

[0105] The main body of the model consists of two stacked LSTM layers and a physical constraint branch. The LSTM part is responsible for the main tasks of extracting the time-series features of the voltage sequence and predicting force and temperature, with 64 hidden units in each layer. The input is a single-channel normalized voltage sequence with an input dimension of 1, and the output is the predicted sequence of cutting force and cutting temperature corresponding to the time window with an output dimension of 2. The time window length is set to... L =50, corresponding to a continuous voltage evolution process of approximately 0.025s. The physical constraint branch does not set independent force and temperature output terminals. Instead, it substitutes the force and temperature prediction results given by LSTM into the voltage-force-temperature coupling expression established earlier to construct physical residuals, which participate in the network parameter update through the loss function.

[0106] During training, the total loss function is optimized by combining the data term and the physical term, as shown in equation (33).

[0107] (33) The training process is divided into two phases. The first phase is the pre-training phase, in which only the data loss term is enabled, i.e., let... λ When 2=0, the total loss function can be written as equation (43).

[0108] (43) In this stage, the Adam optimizer is used for training, with a learning rate set to 1x10. -3 The training run consists of 50 rounds. The main purpose of this stage is to enable the LSTM to learn the fundamental correspondence between voltage sequences and force / temperature outputs on labeled samples, forming a stable temporal feature representation. In this case, no physical residual constraints are introduced, the network parameter update direction is relatively concentrated, and initial convergence can be achieved relatively quickly.

[0109] The second stage is the joint training stage. Based on the pre-trained parameters, a physical loss term is introduced to recover the complete loss function form, which is Equation (33).

[0110] This stage employs the L-BFGS optimizer for fine-tuning, with the number of training epochs controlled within the range of 150–300. Compared to the pre-training stage, this stage focuses not only on label approximation but also on simultaneously compressing both data residuals and physical residuals, ensuring that the network output converges synchronously in terms of both numerical accuracy and mechanistic consistency. The initial loss weights are set to λ1:λ2=0.7:0.3, and are dynamically adjusted during training based on the gradient contributions of the two types of losses, ensuring that the data and physical terms maintain the same order of magnitude of optimization effect.

[0111] This invention employs a two-stage training strategy. First, the LSTM learns the basic mapping relationship through pre-training, and then physical constraints are introduced through joint training for fine-tuning, which effectively improves the convergence speed and final accuracy of the model.

[0112] Step 5: Training Results and Error After completing the two-stage training, the LSTM-PINN hybrid network achieved stable convergence under the joint constraints of data loss and physical loss. Consistent with the previous resistance physical model, the piezoresistive coefficient α, the thermal coefficient β, the coupling coefficient γ, and the baseline bias term ε were treated as learnable parameters during training and updated synchronously with the network weights. After Adam pre-training and L-BFGS joint fine-tuning, the model finally converged to the resistance-force-temperature coupling expression as shown in equation (44).

[0113] (44) In equation (44), α = 1.24, β = -3.85, γ = 0.21, ε = 1053.15; "-" indicates that within the temperature range involved in this embodiment, the conductivity enhancement effect caused by thermal excitation is dominant, causing the resistance to decrease as the temperature increases.

[0114] First, we analyze α and β under univariate conditions. When T=0 is chosen, the change in resistance is only related to the cutting force, and the data is trained to obtain equation (45).

[0115] (45) The specific training results regarding the mapping relationship between cutting force and resistance are as follows: Figure 6 As shown, under univariate force conditions, the model can extract the piezoresistive-dominated electrical response characteristics, and the order of α is consistent with the piezoresistive effect mechanism.

[0116] When F=0 is selected, training the data yields equation (46).

[0117] (46) The specific training results regarding the mapping relationship between cutting temperature and resistance are as follows: Figure 7 As shown, the resistance decreases as the temperature increases, indicating that the conductivity enhancement effect caused by thermal excitation dominates within the temperature range covered in this embodiment.

[0118] To evaluate the decoupling accuracy, 30 sets of test samples were input into the constructed neural network model, and the actual values ​​and predicted values ​​were compared and analyzed. The difference between the network's predicted decoupling value and the corresponding experimentally measured actual value was defined as the prediction decoupling error. The results of the cutting force prediction decoupling error are shown in [the table below]. Figure 8 The results of the cutting temperature prediction decoupling error are shown below. Figure 9 Quantitative statistical analysis shows that, in all test samples, the decoupling error of cutting force prediction is less than 0.1N, and the decoupling error of cutting temperature prediction is less than 0.3℃. Statistical results indicate that the established neural network model has high decoupling accuracy and can effectively decouple the cutting force signal from the cutting temperature signal.

[0119] The above combination Figure 1 The cutting force-temperature decoupling method based on LSTM and PINN neural networks provided in the embodiments of the present invention has been described in detail. Next, the cutting force-temperature decoupling system based on LSTM and PINN neural networks provided in the embodiments of the present invention will be described in conjunction with the accompanying drawings.

[0120] Figure 10 This is a schematic diagram of the cutting force-temperature decoupling system based on LSTM and PINN neural networks, as shown in an embodiment of the present invention. (Refer to...) Figure 10 The system described in this invention includes: The physical coupling model construction module is configured to: establish a physical coupling model between the change in resistance of the tool sensitive area and the cutting force and cutting temperature during the cutting process, and construct physical constraint terms for neural network training based on the model; The hybrid neural network prediction module is configured to: construct a hybrid neural network consisting of a temporal feature extraction branch and a physical constraint branch; the temporal feature extraction branch uses a Long Short-Term Memory (LSTM) network to receive the standardized original cutting voltage time-series signal, extract temporal features, and output a preliminarily decoupled first cutting force signal and first cutting temperature signal; the physical constraint branch uses a Physical Information Neural Network (PINN) to receive the standardized original cutting voltage time-series signal, the first cutting force signal, and the first cutting temperature signal, and uses the residual of the physical coupling model as a physical loss to constrain and correct the first cutting force signal and the first cutting temperature signal to obtain a second cutting force signal and a second cutting temperature signal; the decoupling layer of the hybrid neural network performs weighted fusion of the first cutting force signal, the first cutting temperature signal, the second cutting force signal, and the second cutting temperature signal to obtain the final cutting force signal and the final cutting temperature signal.

[0121] In some embodiments, a two-stage training strategy is employed to train the hybrid neural network. The first stage is a pre-training stage that uses only the data loss term, and the second stage is a joint training stage that uses the total loss function of both the data loss term and the physical loss term.

[0122] In some embodiments, the total loss function is expressed by the following formula:

[0123]

[0124]

[0125] in, Represents the physical loss function; Represents the loss function for supervised data; , Indicates the weighting coefficient; Indicates the first t Predicted cutting force value at time 1. Indicates the first t Predicted cutting temperature at each moment.

[0126] In some embodiments, the physical coupling model is represented by the following formula:

[0127] in, R meas(t) represents the equivalent resistance of the tool-sensitive area measured at time t; F ( t () represents the equivalent cutting force. T ( t ) represents the equivalent cutting temperature, σ( t ) represents the equivalent conductivity; α is the equivalent piezoresistive coefficient, β is the thermistor coefficient, γ is the force-temperature coupling coefficient, and ε is the reference bias term.

[0128] In some embodiments, the original cutting voltage timing signal is a time sequence of the voltage in the tool sensitive area acquired under constant current excitation conditions, which is then input into the hybrid neural network after normalization processing.

[0129] In some embodiments, the temporal feature extraction branch employs a two-layer stacked LSTM network, with 64 hidden units in each layer; the physical constraint branch takes the output of the LSTM network as input and participates in network parameter updates through the physical loss function.

[0130] According to embodiments of the present invention, the cutting force-temperature decoupling system based on LSTM and PINN neural networks can correspond to the execution of the method described in the embodiments of the present invention, and the above and other operations and / or functions of each module of the cutting force-temperature decoupling system based on LSTM and PINN neural networks are respectively for implementing Figure 1 For the sake of brevity, the corresponding processes of each method in the code will not be elaborated here.

[0131] See Figure 11 The diagram shows the structure of a computer device, which includes a processor, a communication interface, and a computer-readable storage medium. The processor, communication interface, and computer-readable storage medium are connected via a bus or other means. The communication interface is used to receive and send data. The computer-readable storage medium can be stored in the computer device's memory. The computer-readable storage medium stores computer programs, including program instructions, and the processor executes the program instructions stored in the computer-readable storage medium. The processor (or CPU, Central Processing Unit) is the computing and control core of the computer device, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions to implement the corresponding steps in the embodiment of the cutting force-temperature decoupling method based on LSTM and PINN neural networks.

[0132] This embodiment provides a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the processing system of the computer device.

[0133] Furthermore, this storage space also contains one or more instructions suitable for loading and execution by the processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM memory or non-volatile memory, such as at least one disk storage device; optionally, it can also be at least one computer-readable storage medium located remotely from the aforementioned processor.

[0134] In one embodiment, the computer-readable storage medium stores one or more instructions; the processor loads and executes one or more instructions stored in the computer-readable storage medium to implement the corresponding steps in the above embodiment of the cutting force-temperature decoupling method based on LSTM and PINN neural networks.

[0135] This embodiment provides a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the corresponding steps in the above embodiment of the cutting force-temperature decoupling method based on LSTM and PINN neural networks.

[0136] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0137] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0138] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0139] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0140] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0141] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A cutting force-temperature decoupling method based on LSTM and PINN neural networks, characterized in that, include: A physical coupling model is established between the change in resistance of the tool's sensitive area and the cutting force and cutting temperature during the cutting process, and physical constraint terms for neural network training are constructed based on this model. A hybrid neural network consisting of a temporal feature extraction branch and a physical constraint branch is constructed. The temporal feature extraction branch uses a Long Short-Term Memory (LSTM) network to receive the standardized original cutting voltage time-series signal, extract temporal features, and output a preliminarily decoupled first cutting force signal and first cutting temperature signal. The physical constraint branch uses a Physical Information Neural Network (PINN) to receive the standardized original cutting voltage time-series signal, the first cutting force signal, and the first cutting temperature signal, and uses the residual of the physical coupling model as a physical loss to constrain and correct the first cutting force signal and the first cutting temperature signal, obtaining a second cutting force signal and a second cutting temperature signal. The decoupling layer of the hybrid neural network performs weighted fusion of the first cutting force signal, the first cutting temperature signal, the second cutting force signal, and the second cutting temperature signal to obtain the final cutting force signal and the final cutting temperature signal.

2. The cutting force-temperature decoupling method based on LSTM and PINN neural network according to claim 1, characterized in that, The hybrid neural network is trained using a two-stage training strategy. The first stage is a pre-training stage that uses only the data loss term, and the second stage is a joint training stage that uses the total loss function of both the data loss term and the physical loss term.

3. The cutting force-temperature decoupling method based on LSTM and PINN neural network according to claim 2, characterized in that, The total loss function is expressed by the following formula: in, Represents the physical loss function; Represents the loss function for supervised data; , Indicates the weighting coefficient; Indicates the first t Predicted cutting force value at time 1. Indicates the first t Predicted cutting temperature at each moment.

4. The cutting force-temperature decoupling method based on LSTM and PINN neural network according to claim 1, characterized in that, The physical coupling model is expressed by the following formula: in, R meas (t) represents the equivalent resistance of the tool-sensitive area measured at time t; F ( t () represents the equivalent cutting force. T ( t ) represents the equivalent cutting temperature, σ( t ) represents the equivalent conductivity; α is the equivalent piezoresistive coefficient, β is the thermistor coefficient, γ is the force-temperature coupling coefficient, and ε is the reference bias term.

5. The cutting force-temperature decoupling method based on LSTM and PINN neural network according to claim 1, characterized in that, The original cutting voltage timing signal is a time sequence of the tool sensitive area voltage acquired under constant current excitation conditions, which is then input into the hybrid neural network after normalization processing.

6. The cutting force-temperature decoupling method based on LSTM and PINN neural network according to claim 1, characterized in that, The temporal feature extraction branch uses a two-layer stacked LSTM network with 64 hidden units in each layer; the physical constraint branch takes the output of the LSTM network as input and participates in the network parameter update through the physical loss function.

7. A cutting force-temperature decoupling system based on LSTM and PINN neural networks, characterized in that, include: The physical coupling model construction module is configured to: establish a physical coupling model between the change in resistance of the tool sensitive area and the cutting force and cutting temperature during the cutting process, and construct physical constraint terms for neural network training based on the model; The hybrid neural network prediction module is configured to: construct a hybrid neural network consisting of a temporal feature extraction branch and a physical constraint branch; the temporal feature extraction branch uses a Long Short-Term Memory (LSTM) network to receive the standardized original cutting voltage time-series signal, extract temporal features, and output a preliminarily decoupled first cutting force signal and first cutting temperature signal; the physical constraint branch uses a Physical Information Neural Network (PINN) to receive the standardized original cutting voltage time-series signal, the first cutting force signal, and the first cutting temperature signal, and uses the residual of the physical coupling model as a physical loss to constrain and correct the first cutting force signal and the first cutting temperature signal to obtain a second cutting force signal and a second cutting temperature signal; the decoupling layer of the hybrid neural network performs weighted fusion of the first cutting force signal, the first cutting temperature signal, the second cutting force signal, and the second cutting temperature signal to obtain the final cutting force signal and the final cutting temperature signal.

8. A computer device, characterized in that, A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the steps of the cutting force-temperature decoupling method based on LSTM and PINN neural networks as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted to be loaded by a processor and execute the steps of the cutting force-temperature decoupling method based on LSTM and PINN neural networks as described in any one of claims 1-6.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps in the cutting force-temperature decoupling method based on LSTM and PINN neural networks as described in any one of claims 1-6.

Images