Milling force prediction method based on transfer learning enhanced WM-GRU neural network

Abstract

The invention discloses a milling force prediction method based on a transfer learning enhanced WM-GRU neural network, and the method comprises the steps: considering a tool wear mechanism, building a WM-GRU prediction model through the influence relation of the tool wear to the milling force in a machining process, predicting the tool wear through GRUW, transmitting a prediction result to GRUM, and carrying out the milling force prediction through the GRUM according to the tool wear and cutting parameters; a migration enhancement training strategy is formulated, fusion with a WM-GRU prediction model is carried out through parameter migration, feature sharing and field adaptation, high-precision adaptation of a target domain is realized based on progressive migration fine tuning, construction of a milling force prediction model based on migration learning and WM-GRU neural network deep fusion is completed, and finally milling force prediction is realized. According to the method, the defects that a milling force prediction model established based on a mechanism is incomplete and a milling force prediction model established based on pure data driving is poor in interpretability are overcome, and the prediction performance of the model on a target domain is improved by introducing transfer learning.

Description

Technical Field

[0001] The present invention belongs to the technical field of milling force prediction in metal processing processes, and specifically relates to a milling force prediction method based on transfer learning enhanced WM-GRU neural network. Background Art

[0002] During the milling process, the magnitude and variation of milling force directly impacts quality indicators such as workpiece dimensional accuracy. Excessive and unstable milling force can lead to excessive deformation of the workpiece, causing the finished dimensions to deviate from design requirements and fail to meet precision standards. Furthermore, excessive milling force increases the load on the tool, exacerbating tool wear, such as accelerated edge chipping and blunting, thereby shortening the tool's lifespan.

[0003] Currently, most technologies focus on analyzing tool wear using milling force as an input parameter. For example, patent application CN115186406A discloses a tool wear prediction method based on transfer learning, patent application CN119622652A discloses a tool wear state prediction method based on multi-source feature fusion, and patent application CN114619292B discloses a tool wear prediction method, device, and computer application based on transfer learning. However, in actual machining processes, it is difficult to accurately predict milling force, and the practical guidance is limited.

[0004] The invention with publication number CN119036199B proposes an indirect prediction method for milling force based on power signals taking into account tool wear. By establishing a theoretical power consumption model and considering the influence of tool wear on cutting force, the influence relationship from tool wear to cutting force to machine tool power is established at the mechanism level, and a variable working condition milling force prediction model based on machine tool power consumption is established, thereby realizing accurate prediction of tool wear and cutting force under any working conditions.

[0005] Through the analysis of milling force prediction technology, it was found that: (1) The prediction model established based on the mechanism has approximate assumptions in the model, which makes the prediction model unable to adapt to the actual processing conditions, thereby affecting the prediction accuracy of the milling force. (2) The milling force prediction model established based on pure data drive needs to rely on a large amount of training data, and the interpretability and generalization ability of the model are poor. (3) In the process of using neural networks to predict milling force, under the same working conditions, after the prediction model is trained on the training set, the data feature distribution in the newly collected data set is different, resulting in the inability to accurately predict the milling force during the processing process. Summary of the Invention

[0006] In order to make up for the shortcomings of the mechanism model and the pure data-driven model in milling force prediction, and at the same time solve the problem of poor prediction performance of the prediction model in newly collected data, the purpose of the present invention is to provide a milling force prediction method based on transfer learning enhanced WM-GRU neural network, which combines the advantages of the mechanism model and the data-driven model, and can accurately predict the milling force in the metal processing process, so that the force changes can be obtained in advance, and the cutting parameters can be adjusted according to the prediction results to avoid the tool from being in a harsh working condition of excessive force, extend the service life of the tool, and effectively ensure the processing quality of the workpiece.

[0007] In order to achieve the above technical objectives, the technical solution adopted by the present invention is:

[0008] A milling force prediction method based on transfer learning enhanced WM-GRU neural network, the method comprising the following steps:

[0009] S1, collects a multi-condition sample data set consisting of cutting parameters, machining signals, tool wear data and milling force signals in the source domain, and collects a small amount of target condition sample data in the target domain;

[0010] S2, performs joint preprocessing of source and target domain sample data through denoising and filtering, aligns multi-source data based on dynamic time warping, and finally unifies the data distribution through domain adaptive normalization method;

[0011] S3, based on GRU neural network, construct tool wear prediction model GRU W and milling force prediction model GRU M The tool wear prediction model and the milling force prediction model are connected in series to establish a WM-GRU prediction model. The tool wear prediction model receives the cutting parameters, machining signals, and historical wear data of the previous moment as input, and outputs the tool wear prediction value of the current moment through the time series modeling capability of the GRU network. The milling force prediction model takes the cutting parameters, machining signals, and the tool wear prediction value of the current moment as input, and outputs the predicted milling force of the current moment.

[0012] S4, formulate transfer enhancement training strategy, in the source domain training stage, use the adversarial domain discriminator to constrain the domain invariant feature learning of the feature extraction layer, and improve the tool wear prediction model GRU W The parameters of the milling force prediction model GRU are retained. M The parameters of the source domain and the target domain are migrated, and the loss function is introduced to keep the topological consistency of the wear-milling force mapping relationship between the source domain and the target domain.

[0013] S5, implements progressive transfer fine-tuning to achieve high-precision adaptation to the target domain by fine-tuning local model parameters;

[0014] S6 embeds the migrated and optimized WM-GRU prediction model into the CNC system, receives cutting parameters, machining signals, and milling force signals in real time, and outputs tool wear status and milling force prediction values, while continuously updating the target domain model parameters through online learning.

[0015] Step S1 further comprises:

[0016] Build a test platform, select cutting parameters, and design a three-factor three-level orthogonal test table;

[0017] Milling experiments were carried out according to the orthogonal test table to obtain a sample data set consisting of various cutting parameters, machining signals, tool wear data and milling force signals.

[0018] Furthermore, the cutting parameters include feed per tooth f z , spindle speed n and axial cutting depth a p .

[0019] Step S2 further comprises:

[0020] S21 uses the truncation method to intercept the machining signals when the tool cuts in and out. The machining signals, including power, current, vibration, and cutting force, are filtered using a wavelet packet-Butterworth hybrid filter to obtain discrete wear data. The number of wavelet packet decomposition layers is adaptively adjusted according to the signal's main frequency, and the Butterworth filter cutoff frequency is dynamically set according to the signal's characteristic frequency.

[0021] S23, implements physical constraint sliding average on discrete wear data to make the wear curve conform to the Taylor formula trend;

[0022] S24, using the current signal phase as a timing reference, performing multi-signal alignment through dynamic time warping, and adjusting the wear tag timing according to the alignment result;

[0023] S25, normalizing the sample data set after the preprocessing operation, and dividing it into a training set and a test set in proportion.

[0024] Furthermore, in step S3, the expression of the WM-GRU prediction model is:

[0025] F (t) =GRU WM (x (t) , GRU W (x (t) , h W(t-1) ), h M(t-1) );

[0026] Where h M(t-1) and h W(t-1) Represents the milling force prediction model GRU Mand tool wear prediction model GRU W In the hidden state of the previous moment, the initial hidden state is set to zero vector; X M(t) =[f z ,n,a p ,VB (t) ] T , x (t) =[f z ,n,a p ] T , f z Indicates the feed per tooth, n indicates the spindle speed, a p Indicates the axial cutting depth.

[0027] Step S4 further comprises:

[0028] S41, retain the source domain pre-trained tool wear prediction model GRU W The weights of the middle and bottom GRU units are unfrozen. Only the last two layers of GRU units and the fully connected layer are used for target domain fine-tuning. Milling force prediction model GRU M Regularized parameter transfer is used to impose Fisher information matrix constraints on the source domain pre-training parameters, limiting the update amplitude of important parameters to protect the physical laws learned in the source domain from being destroyed;

[0029] S42, in the tool wear prediction model GRU w and milling force prediction model GRU M Designing a shared-private feature space between two models and enforcing the tool wear prediction model GRU through orthogonal constraint loss w Hidden layer output and milling force prediction model GRU M The input features of the GRU are kept linearly correlated in the shared subspace, and the gated feature sharing unit is used to dynamically adjust the tool wear prediction model GRU according to the target domain data distribution. W and milling force prediction model GRU M The feature transfer weight between

[0030] S43, using the maximum mean difference to calculate the MMD distance between the source domain and target domain data feature distributions, and minimize the MMD distance through optimization. The calculation formula of the MMD distance is:

[0031]

[0032] Among them, X s is the source domain dataset; X y is the target domain dataset; n s is the number of samples in the source domain dataset; n y is the number of samples in the target domain dataset; is a nonlinear feature mapping from the input space to the Hilbert space.

[0033] Step S5 further comprises:

[0034] S51, fixed tool wear prediction model GRU w All parameters of the milling force prediction model GRU are trained using only target domain data M The last layer keeps the source domain knowledge intact;

[0035] S52, unfreeze the parameters of the shared feature layer, adopt a curriculum learning strategy to gradually increase the complexity of the target domain samples, and optimize the MMD loss;

[0036] S53, based on the dynamic weight averaging algorithm, adaptively adjusts the loss function weights of the source domain and the target domain.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] First, the milling force prediction method based on transfer learning enhanced WM-GRU neural network of the present invention establishes a neural network prediction model considering the mechanism, and designs two sub-GRUs, one for predicting tool wear and the other for predicting milling force. This prediction model not only overcomes the defects of poor interpretability and weak generalization ability of the prediction model, but also inherits the advantage of neural network with high prediction accuracy, and realizes high-precision prediction of milling force in the metal processing process, providing data support for extending tool life and improving processing quality.

[0039] Second, the present invention's method for predicting milling forces using a transfer learning-enhanced WM-GRU neural network incorporates model-based transfer learning into the prediction model. This method uses MMD to calculate the distance between the feature distributions of source and target domain data, minimizing this distance to optimize the model. The model is trained on a source domain dataset and then transferred to the target domain for prediction. The introduction of the transfer model addresses the problem of poor prediction performance on small and unlabeled datasets. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a flow chart of the milling force prediction method based on transfer learning enhanced WM-GRU neural network of the present invention;

[0041] Figure 2 Schematic diagram of a data acquisition test platform for the milling force prediction method based on transfer learning enhanced WM-GRU neural network of the present invention;

[0042] Figure 3 This is the basic structure diagram of the GRU neural network based on the transfer learning enhanced WM-GRU neural network of the present invention;

[0043] Figure 4 It is a basic structure diagram of the WM-GRU neural network of the milling force prediction method based on transfer learning enhanced WM-GRU neural network of the present invention;

[0044] Figure 5 It is a WM-GRU neural network calculation flow chart of the milling force prediction method based on transfer learning enhanced WM-GRU neural network of the present invention;

[0045] Figure 6 It is a transfer learning principle diagram of the milling force prediction method based on transfer learning enhanced WM-GRU neural network of the present invention. DETAILED DESCRIPTION

[0046] The embodiments of the present invention are described in further detail below with reference to the accompanying drawings.

[0047] like Figure 1 As shown, the present invention is based on the milling force prediction method of the WM-GRU neural network enhanced by transfer learning, and the specific steps are as follows:

[0048] S1: Build a test platform such as Figure 2 As shown in the figure, a data acquisition experiment was carried out, appropriate cutting parameters were selected, and a three-factor three-level orthogonal experimental table was designed. Then, a milling experiment was carried out according to the orthogonal experimental table, and processing signals, tool wear data, and milling force signals were collected. The sampling frequency of each signal source was 300Hz, and a sample data set consisting of various cutting parameters and milling forces was obtained. Among them, the tool material was cemented carbide with a specification of 12*75*R1, and the workpiece material was Ti6Al4V.

[0049] S2, performs joint preprocessing of source and target domain sample data through denoising and filtering, aligns multi-source data based on dynamic time warping, and finally unifies the data distribution through domain adaptive normalization method;

[0050] S3, based on the GRU neural network, the tool wear prediction model and the milling force prediction model are respectively constructed; considering the tool wear mechanism, the tool wear prediction model and the milling force prediction model are connected in series to establish the WM-GRU prediction model; wherein, the tool wear prediction model is represented by GRU W , receiving the cutting parameters, processing signals, and historical wear data of the previous moment as input, and outputting the tool wear prediction value of the current moment through the time series modeling capability of the GRU network; the milling force prediction model is represented by GRU M , taking the cutting parameters, machining signals at the previous moment and the tool wear prediction value at the current moment as input, outputs the predicted milling force at the current moment;

[0051] S4, formulate a transfer enhancement training strategy. In the source domain training phase, the adversarial domain discriminator is used to constrain the domain invariant feature learning of the feature extraction layer, and the GRUW The parameters of GRU are retained. M The parameters of the source domain and the target domain are migrated, and the loss function is introduced to ensure that the wear-milling force mapping relationship between the source domain and the target domain maintains topological consistency;

[0052] S5, implements progressive transfer fine-tuning to achieve high-precision adaptation to the target domain by fine-tuning local model parameters;

[0053] S6, embeds the migrated and optimized WM-GRU model into the CNC system, receives sensor data in real time and outputs tool wear status and milling force prediction values, while continuously updating the target domain model parameters through online learning.

[0054] In step S1, the test platform is as follows Figure 2 As shown, the cutting parameters include feed per tooth f Z , spindle speed n and axial cutting depth a P .

[0055] In step S2, the steps of preprocessing the sample data set are:

[0056] S21, using the truncation method to intercept the processing signals when the tool cuts in and out;

[0057] S22, wavelet packet-Butterworth hybrid filtering is used for processing signals (power, current, vibration, force). The number of wavelet packet decomposition layers is adaptively adjusted according to the main frequency of the signal, and the cutoff frequency of the Butterworth filter is dynamically set according to the characteristic frequency of the signal;

[0058] S23, implements physical constraint sliding average on discrete wear data to ensure that the wear curve conforms to the Taylor formula trend;

[0059] S24, using the current signal phase as the timing reference, multiple signals are aligned through dynamic time warping (DTW), and the wear tag timing is adjusted according to the alignment result;

[0060] S25, normalizing the sample data set after the preprocessing operation, and dividing the training data set and the test data set into a ratio of 8:2, that is, the training data set accounts for 80% of the total data set, and the test data set accounts for 20% of the total data set.

[0061] In step S3, the basic GRU neural network milling force prediction model mainly includes a standard GRU and a fully connected layer, such as Figure 4 As shown. At the current moment t, the GRU neural network receives the hidden state h of the previous moment (t-1) and the current collected signal x (t) , and then predict the milling force F at the current moment (t) And output the updated hidden state h at time t(t) , the specific calculation steps are as follows:

[0062] First, through h (t-1) and x (t) Calculate the reset gate output R at the current time t (t) :

[0063] R (t) =σ(W rh h (t-1) +W rx x (t) +b r )

[0064] Among them, W rh and W rx Represent the corresponding weight matrices respectively; b r represents the corresponding bias; σ is the sigmoid function.

[0065] Then, through R (t) Calculate the candidate hidden state at time t And through h (t-1) and x (t) Calculate the output Z of the update gate at the current time t (t) :

[0066]

[0067] Z (t) =σ(W zh h (t-1) +W iz X (t) +b z )

[0068] Where, ⊙ is the Hadamard product; Z (t) The meaning is the selection of forgetting the original hidden state, Z (t) The closer the elements in are to 0, the easier it is for the corresponding elements in the hidden state to be forgotten, so 1-Z (t) Delegates to adding storage to a candidate hidden state.

[0069] In getting and Z (t) After that, the hidden state h at the current moment can be calculated (t) :

[0070]

[0071] Finally, the milling force F at the current time t is (t) The fully connected layer can be used to (t) Transformation yields:

[0072] F(t) =W l h (t) +b l )

[0073] That is, the milling force predicted based on the GRU neural network can be expressed as:

[0074] F (t) =GRU(x (t) , h (t-1) )

[0075] Among them, x (t) =[f z ,n,a p ] T ; Initial hidden state h (0) Set to the zero vector.

[0076] In step S4, following the idea of ​​hybrid drive modeling, a WM-GRU neural network considering the mechanism is established. The sub-GRU neural network used to predict tool wear is called the wear neural network, and the sub-GRU neural network used to predict milling force is called the milling neural network. W Receive cutting parameters, processing signals and historical wear data as input, through the temporal modeling capability of GRU network, GRU W Output the current tool wear prediction value VB (t) :

[0077] VB (t) =GRU W (x (t) , h W(t-1) )

[0078] Among them, x (t) =[f z ,n,a p ] T .

[0079] The obtained VB (t) Add to x W(t) are used together as input to the milling GRU neural network to predict the milling force F (t) :

[0080] F (t) =GRU M (x M(t) , h M(t-1) )

[0081] Among them, x M(t) =[f z ,n,a p ,VB (t) ] T ;h M(t-1)Represents GRU M The hidden state at the previous moment.

[0082] Based on the above, the proposed WM-GRU neural network can be expressed as:

[0083] F (t) =GRU WM (x (t) , GRU W (x (t) , h W(t-1) ), h M(t-1) )

[0084] In step S4, the migration enhancement training strategy steps are:

[0085] S41, retain the source domain pre-trained GRU w The weights of the bottom GRU units are unfrozen, and only the last two layers of GRU units and the fully connected layer are used for target domain fine-tuning. M Regularized parameter migration is used to impose Fisher information matrix constraints on the source domain pre-training parameters, limiting the update amplitude of important parameters and protecting the physical laws learned in the source domain from being destroyed.

[0086] S42, in GRU W and GRU M Designing a shared-private feature space between two nodes and enforcing GRU through orthogonal constraint loss W The hidden layer output and GRU M The input features of the GRU are kept linearly correlated in the shared subspace, and the gated feature sharing unit is used to dynamically adjust the GRU according to the target domain data distribution. W and GRU M The feature transfer weight between

[0087] S43, using the maximum mean difference to calculate the MMD distance between the source domain and target domain data feature distributions, and minimize the MMD distance through optimization. The calculation formula of the MMD distance is:

[0088]

[0089] Among them, X s is the source domain dataset; X y is the target domain dataset; n s is the number of samples in the source domain dataset; n y is the number of samples in the target domain dataset; is a nonlinear feature mapping from the input space to the Hilbert space.

[0090] In step S5, the steps of progressive migration fine-tuning are:

[0091] S51, fixed GRU w All parameters of GRU are trained using only target domain data. M The last layer keeps the source domain knowledge intact;

[0092] S52, unfreeze the parameters of the shared feature layer, adopt a curriculum learning strategy to gradually increase the complexity of the target domain samples, and optimize the MMD loss;

[0093] S53, based on the dynamic weight averaging algorithm, adaptively adjusts the loss function weights of the source domain and the target domain.

[0094] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present application.

[0095] Obviously, those skilled in the art may make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is intended to include these modifications and variations.

Claims

1. A milling force prediction method based on transfer learning enhanced WM-GRU neural network, characterized in that: The method comprises the following steps: S1, collects a multi-condition sample data set consisting of cutting parameters, machining signals, tool wear data and milling force signals in the source domain, and collects a small amount of target condition sample data in the target domain; S2, performs joint preprocessing of source and target domain sample data through denoising and filtering, aligns multi-source data based on dynamic time warping, and finally unifies the data distribution through domain adaptive normalization method; S3, based on GRU neural network, construct tool wear prediction model GRU W and milling force prediction model GRU M The tool wear prediction model and the milling force prediction model are connected in series to establish a WM-GRU prediction model. The tool wear prediction model receives the cutting parameters, machining signals, and historical wear data of the previous moment as input, and outputs the tool wear prediction value of the current moment through the time series modeling capability of the GRU network. The milling force prediction model takes the cutting parameters, machining signals, and the tool wear prediction value of the current moment as input, and outputs the predicted milling force of the current moment. S4, formulate transfer enhancement training strategy, in the source domain training stage, use the adversarial domain discriminator to constrain the domain invariant feature learning of the feature extraction layer, and improve the tool wear prediction model GRU W The parameters of the milling force prediction model GRU are retained. M The parameters of the source domain and the target domain are migrated, and the loss function is introduced to keep the topological consistency of the wear-milling force mapping relationship between the source domain and the target domain. S5, implements progressive transfer fine-tuning to achieve high-precision adaptation to the target domain by fine-tuning local model parameters; S6 embeds the migrated and optimized WM-GRU prediction model into the CNC system, receives cutting parameters, machining signals, and milling force signals in real time, and outputs tool wear status and milling force prediction values, while continuously updating the target domain model parameters through online learning.

2. The milling force prediction method based on transfer learning enhanced WM-GRU neural network according to claim 1 is characterized in that: Step S1 further comprises: Build a test platform, select cutting parameters, and design a three-factor three-level orthogonal test table; Milling experiments were carried out according to the orthogonal test table to obtain a sample data set consisting of various cutting parameters, machining signals, tool wear data and milling force signals.

3. The milling force prediction method based on transfer learning enhanced WM-GRU neural network according to claim 1 is characterized in that: The cutting parameters include feed per tooth f z , spindle speed n and axial cutting depth a p .

4. The milling force prediction method based on transfer learning enhanced WM-GRU neural network according to claim 1 is characterized in that: Step S2 further comprises: S21 uses the truncation method to intercept the machining signals when the tool cuts in and out. The machining signals, including power, current, vibration, and cutting force, are filtered using a wavelet packet-Butterworth hybrid filter to obtain discrete wear data. The number of wavelet packet decomposition layers is adaptively adjusted according to the signal's main frequency, and the Butterworth filter cutoff frequency is dynamically set according to the signal's characteristic frequency. S23, implements physical constraint sliding average on discrete wear data to make the wear curve conform to the Taylor formula trend; S24, using the current signal phase as a timing reference, performing multi-signal alignment through dynamic time warping, and adjusting the wear tag timing according to the alignment result; S25, normalizing the sample data set after the preprocessing operation, and dividing it into a training set and a test set in proportion.

5. The milling force prediction method based on transfer learning enhanced WM-GRU neural network according to claim 1 is characterized in that: In step S3, the expression of the WM-GRU prediction model is: F (t) =GRU WM (x (t) ,GRU W (x (t) ,h W(t-1) ),h M(t-1) ); Where h M(t-1) and h W(t-1) Represents the milling force prediction model GRU M and tool wear prediction model GRU W In the hidden state of the previous moment, the initial hidden state is set to zero vector; x M(t) =[f z ,n,a p ,VB (t) ] T , x (t) =[f z ,n,a p ] T , f z Indicates the feed per tooth, n indicates the spindle speed, a p Indicates the axial cutting depth.

6. The milling force prediction method based on transfer learning enhanced WM-GRU neural network according to claim 1 is characterized in that: Step S4 further comprises: S41, retain the source domain pre-trained tool wear prediction model GRU W The weights of the middle and bottom GRU units are unfrozen. Only the last two layers of GRU units and the fully connected layer are used for target domain fine-tuning. Milling force prediction model GRU M Regularized parameter transfer is used to impose Fisher information matrix constraints on the source domain pre-training parameters, limiting the update amplitude of important parameters to protect the physical laws learned in the source domain from being destroyed; S42, in the tool wear prediction model GRU W and milling force prediction model GRU M Designing a shared-private feature space between two models and enforcing the tool wear prediction model GRU through orthogonal constraint loss W Hidden layer output and milling force prediction model GRU M The input features of the GRU are kept linearly correlated in the shared subspace, and the gated feature sharing unit is used to dynamically adjust the tool wear prediction model GRU according to the target domain data distribution. W and milling force prediction model GRU M The feature transfer weight between S43, using the maximum mean difference to calculate the MMD distance between the source domain and target domain data feature distributions, and minimize the MMD distance through optimization. The calculation formula of the MMD distance is: Among them, X s is the source domain dataset; X y is the target domain dataset; n s is the number of samples in the source domain dataset; n y is the number of samples in the target domain dataset; is a nonlinear feature mapping from the input space to the Hilbert space.

7. The milling force prediction method based on transfer learning enhanced WM-GRU neural network according to claim 1 is characterized in that: Step S5 further comprises: S51, fixed tool wear prediction model GRU W All parameters of the milling force prediction model GRU are trained using only target domain data M The last layer keeps the source domain knowledge intact; S52, unfreeze the parameters of the shared feature layer, adopt a curriculum learning strategy to gradually increase the complexity of the target domain samples, and optimize the MMD loss; S53, based on the dynamic weight averaging algorithm, adaptively adjusts the loss function weights of the source domain and the target domain.

Images