Pseudo-visual anomaly detection method based on control quantity-response quantity

Abstract

The invention discloses a control quantity-response quantity-based pseudo-visual anomaly detection method, which comprises the following steps of: determining a target measurement point capable of responding to a service scene, and analyzing and searching a control quantity related to the target measurement point; collecting and cleaning monitoring data containing a control quantity and a response quantity, and assembling the monitoring data into a data structure suitable for model training; processing the data by using an improved 1D U-Net network architecture, and carrying out model training on a supervised learning task with a fault sample by using a cross entropy loss function; for a learning task of an unsupervised sample, performing model training by using a reconstruction error; self-adaptive anomaly judgment and dynamic anomaly density labeling are carried out, finally, a detection model is trained and evaluated, and detection and early warning are carried out; the problems of low precision and high false alarm rate caused by difficulty in setting an early warning threshold value when a regression algorithm is used for prediction in the prior art are solved, and fault and normal classification judgment is realized through the classification model.

Description

Technical Field

[0001] The present invention belongs to the technical field of data-driven intelligent early warning, and particularly relates to a pseudo-visual anomaly detection method based on control quantity-response quantity. Background Art

[0002] This method belongs to data-driven intelligent early warning technology. This type of technology mainly collects real-time monitoring data of hydropower plants, such as water level, flow rate, vibration, temperature, etc., and uses data mining and machine learning algorithms to process and analyze the data, so as to achieve early warning of equipment failures.

[0003] The prior art such as patent application number CN202211535534.5 discloses a method, device and terminal for fault early warning of a hydropower unit, which obtains a normal vibration signal sample set under the normal state of the hydropower unit, extracts vibration signal features according to the vibration signal sample set; trains a sparse autoencoder according to the normal vibration signal features, and uses the trained sparse autoencoder as the unit health model; determines the fault early warning threshold of the hydropower unit according to the normal vibration signal sample set under the normal state of the hydropower unit and the unit health model; receives the vibration signal sample of the hydropower unit to be early warned, inputs the vibration signal sample into the unit health model, determines the corresponding characteristic quantity, and determines the working state of the hydropower unit to be early warned according to the characteristic quantity and the fault early warning threshold; starting from the characteristics of the normal state vibration signal of the hydropower unit, it excavates the early fault signs of the unit and gives an early warning in time at the stage of fault germination, which can improve the accuracy of fault early warning of the hydropower unit; Another example is patent application number CN202210520155.2, which discloses a method and system for trend analysis and early warning of operation data of a hydropower unit based on big data, including obtaining the first real-time operating position of a preset main monitoring component and the shutdown form of the main component during shutdown, as well as the second real-time operating position of a preset auxiliary monitoring component and the shutdown form of the auxiliary component during shutdown; generating a real-time relative cooperation position between the preset main monitoring component and the preset auxiliary monitoring component; obtaining the standard working cooperation position between the preset main monitoring component and the preset auxiliary monitoring component based on a preset big data communication module, and generating current relative position difference data; if it is determined that the current relative position difference data is greater than a preset normal working range reference value, generating a first early warning indication and giving an alarm for the hydropower unit based on the first early warning indication; this invention greatly improves the accuracy and reliability of early warning, and at the same time greatly improves the production efficiency.

[0004] Currently, most of these methods are data-driven and use prediction or regression algorithms to determine whether there are anomalies in real-time data. Due to difficulties in setting warning thresholds and low accuracy of the regression algorithms currently in use, the false alarm rate is too high. Therefore, it is necessary to propose a pseudo-visual anomaly detection method based on control variables - response variables to solve the above problems. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a pseudo-visual anomaly detection method based on control variables - response variables, aiming to solve the problems of low accuracy and high false alarm rate caused by difficulties in setting warning thresholds when using regression algorithms for prediction in the prior art, and to achieve the classification determination of faults and normality through a classification model.

[0006] To solve the above technical problems, the technical solution adopted by the present invention is: A pseudo-visual anomaly detection method based on control variables - response variables, comprising the following steps: S1. Determine the target measurement points that can respond to the business scenario, and the target measurement points are the response variables during the operation of the hydropower plant; S2. Analyze and find the control variables related to the target measurement points; S3. Collect and clean the monitoring data containing control variables and response variables, and assemble the monitoring data into a data structure suitable for model training; S4. Use an improved 1D U-Net network architecture to process the data. The network architecture includes an encoder and a decoder, and if it is used for a classification task, it also includes a classifier output head; use 1D temporal convolution for time series data; S5. For the supervised learning task with faulty samples, use the cross-entropy loss function for model training; for the unsupervised sample learning task, use the reconstruction error for model training; on the basis of the reconstruction error, use two reconstruction heads with the same structure, the D1 reconstructor and the D2 generator, to improve the task effect; S6. Design an adaptive anomaly determination threshold to distinguish normal and abnormal data based on the mean of the reconstruction errors of the training set multiplied by a preset coefficient; S7. Design a labeling method based on dynamic anomaly density for fine-grained labeling of the abnormal regions of a segment of data; S8. Train a detection model based on the 1D U-Net network architecture to capture the data relationship between control variables and response variables; S9. Test and evaluate the detection model; S10. Use the trained detection model to perform fault warning on new monitoring data and output the warning result.

[0007] Preferably, the response quantity is the water turbine speed, and the control quantity is the regulating valve opening. When analyzing and finding the control quantity related to the target measurement point, the measurement point early warning is carried out through the physical relationship between the control quantity and the response quantity.

[0008] Preferably, the network architecture is an architecture based on a convolutional neural network for capturing the relationship between measurement points and performing early warning; the network architecture uses a 1D convolutional neural network, and uses a 1D convolutional kernel to capture both the dynamic characteristics between measurement points and the information in the time dimension to achieve early warning.

[0009] Preferably, in step S4, the encoder in the improved 1D U-Net network architecture is as follows: Input: The input time series data X ∈ R LxC , where L is the time step length and C is the number of signal channels; Convolutional layer: The encoder contains multiple consecutive one-dimensional convolutional layers Conv1D, and cooperates with the max pooling operation MaxPooling to gradually reduce the time step length layer by layer to compress the time dimension and increase the number of channels, so as to extract high-level time series features; each layer of convolutional operation is expressed as: ); In the formula, and are the weights and biases of the i-th layer of convolution, and ReLu is the activation function; Skip connection: In each layer of the encoder, the feature map is passed to the corresponding layer of the decoder for fusing features in the decoding stage, helping to restore information and improve accuracy.

[0010] Preferably, in step S4, the decoder in the improved 1D U-Net network architecture is as follows: Transposed convolution and feature fusion: The decoder contains multiple transposed convolution TransposeConvolution layers, gradually restoring the time step dimension, i.e., upsampling; the output of each layer is concatenated with the skip connection features of the corresponding encoder layer, and the fused information is: ); In the formula, is the feature map of the decoder, is the feature map from the encoder, and the understanding of the time features is deepened through feature concatenation.

[0011] Preferably, in step S4, the output layer in the improved 1D U-Net network architecture is as follows: The final output of the decoder generates a classification result through a fully connected layer or a 1D convolutional layer; for a binary classification task, the Sigmoid activation function is used to map the output to a probability value between [0,1]: ; where y represents the probability value that the input sequence belongs to the positive class, represents the Sigmoid activation function.

[0012] Preferably, in step S4, 1D temporal convolution is used for the temporal data. In 1D convolution, the filter, i.e., the convolution kernel, slides along the time step dimension, performs operations on each local region of the data, and gradually extracts features at different levels; assuming the input data X ∈ R LxC , where L represents the length of the time step and C is the number of input channels, the convolution operation is expressed as: ; where Y(t) is the eigenvalue of the convolution output at time step t, W is the weight matrix of the convolution kernel, its size is K x C, where K represents the length of the kernel and C represents the number of input channels, and X(t + k - 1, c) represents the value of the input signal at channel c and time step t + k - 1.

[0013] Furthermore, each convolution kernel contains several weight parameters. By learning these parameters, the convolution layer can automatically extract important features from the input signal. After each convolution operation, a feature map is generated to represent different local features.

[0014] Preferably, for the supervised learning of faulty samples, cross - entropy loss is used for learning: ; where, is the true label, is the probability value predicted by the model, and N is the number of samples.

[0015] Preferably, for the task of unsupervised samples, reconstruction error is used for learning: ; Based on the reconstruction error, two reconstruction heads with the same structure, the reconstructor D1 and the generator D2, are used to improve the task effect, including: The training part of the reconstructor D1 - Decoder (Discriminator): During the training of the reconstructor D1, two loss terms are designed: The reconstruction error L1: ; The reconstruction error L1 calculates the reconstruction error of the reconstructor D1 for the real data. Through this error, the reconstructor D1 tries to minimize the reconstruction error on the normal data, so as to learn the feature distribution of the normal data; The adversarial error L2: ; Reconstruction error The reconstruction error calculated is the reconstruction error of the reconstructor D1 for the generated data, i.e., abnormal data. Through this error, the reconstruction error of the reconstructor D1 for the generated data is increased, so as to distinguish normal data from abnormal data; Total loss : ; Among them, the weight coefficient 0.25 controls the balance between the two losses; this loss term hopes to improve the sensitivity of the reconstructor D1 to abnormal data by minimizing the reconstruction error for normal data and maximizing the reconstruction error for abnormal data.

[0016] Training part of the generator D2 - Generator: The training of the generator D2 includes two main loss terms: Reconstruction error L3: ; The reconstruction error L3 of the generator D2 represents the gap between the generated data and the real data; it enables the Generator to learn the characteristics of normal data during training.

[0017] Adversarial error L4: ; By introducing the detach function to avoid backpropagation, the reconstruction error of the reconstructor D1 for the generated data is reduced. Through this loss, the sensitivity of the decoder to the generated data is enhanced.

[0018] Preferably, for the reconstruction error, an adaptive anomaly detection threshold is adopted, and the mean of the reconstruction error of the training set * 1.5 is used for distinction: Threshold setting and detection: In every 10 epochs, the model enters the evaluation mode to calculate the value of the reconstruction error and perform anomaly detection. The threshold is calculated as follows: ; Calculate the reconstruction error of each batch, multiply the mean of the reconstruction error by 1.5 as the detection threshold to distinguish normal data from abnormal data; Anomaly detection in the test stage: For each sample of the test data, calculate its reconstruction error. If its reconstruction error is greater than the threshold, it is determined to be abnormal, otherwise it is normal; Accuracy calculation: The detection accuracy is statistically calculated by comparing the predicted labels and the true labels.

[0019] The beneficial effects of the present invention are as follows: 1. Innovation in the analysis method, innovatively proposing a measuring point early warning mode by controlling the physical relationship between the quantity and the response quantity; 2. Innovation in the network architecture, proposing to capture and give early warnings about the relationships between measuring points based on a convolutional neural network; 3. The detailed innovation points of this network are as follows: The 1D convolutional neural network is used. The 1D convolutional kernel is used to capture both the dynamic characteristics between measuring points and the information in the time dimension to achieve more accurate early warnings. Description of the Drawings

[0020] Figure 1 Schematic diagram of the flow chart of the present invention; Figure 2 Schematic diagram of the unsupervised early warning model in the embodiment of the present invention; Figure 3 Schematic diagram of the model training curve in the embodiment of the present invention. Detailed Implementation Modes

[0021] Embodiment 1: As Figure 1 shown, a pseudo-visual anomaly detection method based on the control quantity - response quantity includes the following steps: S1. Determine the target measuring points that can respond to the business scenario, and the target measuring points are the response quantities during the operation of the hydropower plant; S2. Analyze and find the control quantities related to the target measuring points; S3. Collect and clean the monitoring data containing the control quantity and the response quantity, and assemble the monitoring data into a data structure suitable for model training; S4. Use the improved 1-DU-Net network architecture to process the data. The network architecture includes an encoder and a decoder. If it is used for a classification task, it also includes a classifier output head; Use 1D time convolution for time series data; S5. For the supervised learning task of faulty samples, use the cross-entropy loss function for model training; For the learning task of unsupervised samples, use the reconstruction error for model training; On the basis of the reconstruction error, use two reconstruction heads D1 reconstructor and D2 generator with the same structure to improve the task effect; S6. Design an adaptive anomaly determination threshold, and distinguish normal and abnormal data based on the mean of the reconstruction error of the training set multiplied by a preset coefficient; S7. Design a set of annotation methods based on dynamic anomaly density for fine-grained annotation of the abnormal regions of a piece of data; S8. Train a detection model based on the 1D U-Net network architecture to capture the data relationship between the control quantity and the response quantity; S9. Test and evaluate the detection model; S10. Use the trained detection model to perform fault warning on new monitoring data and output the warning result.

[0022] Preferably, the response quantity is the turbine speed, and the control quantity is the opening of the regulating valve. When analyzing and finding the control quantity related to the target measuring point, the measuring point warning is carried out through the physical relationship between the control quantity and the response quantity.

[0023] Preferably, the network architecture is an architecture based on a convolutional neural network for capturing the relationship between measuring points and performing warning; the network architecture uses a 1D convolutional neural network, and uses a 1D convolutional kernel to capture both the dynamic characteristics between measuring points and the information in the time dimension to achieve warning.

[0024] Preferably, in step S4, the encoder in the improved 1D U-Net network architecture is as follows: Input: Input time series data X ∈ R LxC , where L is the time step length and C is the number of signal channels; Convolutional layer: The encoder contains multiple consecutive one-dimensional convolutional layers Conv1D, and cooperates with the max pooling operation MaxPooling to gradually reduce the time step length to compress the time dimension and increase the number of channels, so as to extract high-level time series features; each layer of convolutional operation is expressed as: ); In the formula, and are the weights and biases of the i-th layer of convolution, and ReLu is the activation function; Skip connection: In each layer of the encoder, the feature map is passed to the corresponding layer of the decoder for fusing features in the decoding stage, helping to restore information and improve accuracy.

[0025] Preferably, in step S4, the decoder in the improved 1D U-Net network architecture is as follows: Transposed convolution and feature fusion: The decoder contains multiple transposed convolution TransposeConvolution layers, gradually restoring the time step dimension, i.e., upsampling; the output of each layer is concatenated with the skip connection features of the corresponding encoder layer, and the fused information is: ); In the formula, is the feature map of the decoder, is the feature map from the encoder, and the time features are better understood through feature concatenation.

[0026] Preferably, in step S4, the output layer in the improved 1D U-Net network architecture is as follows: The final output of the decoder generates classification results through a fully connected layer or a 1D convolutional layer; for binary classification tasks, the Sigmoid activation function is used to map the output to a probability value between [0,1]: ; where y represents the probability value that the input sequence belongs to the positive class, represents the Sigmoid activation function.

[0027] Preferably, in step S4, 1D temporal convolution is used for temporal data. In 1D convolution, the filter, i.e., the convolution kernel, slides along the time step dimension, performs operations on each local region of the data, and gradually extracts features at different levels; assuming the input data X ∈ R LxC , where L represents the time step length and C is the number of input channels, the convolution operation is expressed as: ; where Y(t) is the eigenvalue of the convolution output at time step t, W is the weight matrix of the convolution kernel, whose size is K x C, where K represents the length of the kernel and C represents the number of input channels, and X(t + k - 1, c) represents the value of the input signal at channel c and time step t + k - 1.

[0028] Furthermore, each convolution kernel contains several weight parameters. By learning these parameters, the convolutional layer can automatically extract important features from the input signal. After each convolution operation, a feature map is generated to represent different local features.

[0029] Preferably, for the supervised learning of faulty samples, cross-entropy loss is used for learning: ; where, is the true label, is the probability value predicted by the model, and N is the number of samples.

[0030] Preferably, for the tasks of unsupervised samples, reconstruction error is used for learning: ; Based on the reconstruction error, two reconstruction heads with the same structure, the reconstructor D1 and the generator D2, are used to improve the task effect, including: The training part of the reconstructor D1 - Decoder (Discriminator): During the training of the reconstructor D1, two loss terms are designed: The reconstruction error L1: ; The reconstruction error L1 calculates the reconstruction error of the reconstructor D1 for the real data. By minimizing this error, the reconstructor D1 can learn the feature distribution of normal data with a minimized reconstruction error on normal data. The adversarial error L2: ; The reconstruction error calculates the reconstruction error of the reconstructor D1 for the generated data, i.e., abnormal data. By maximizing this error, the reconstructor D1 can distinguish normal data from abnormal data. The total loss : ; Among them, the weight coefficient 0.25 controls the balance between the two losses. This loss term aims to improve the sensitivity of the reconstructor D1 to abnormal data by minimizing the reconstruction error for normal data and maximizing the reconstruction error for abnormal data.

[0031] The training part of the generator D2 - Generator: The training of the generator D2 includes two main loss terms: The reconstruction error L3: ; The reconstruction error L3 of the generator D2 represents the gap between the generated data and the real data, enabling the Generator to learn the features of normal data during training.

[0032] The adversarial error L4: ; By introducing the detach function to avoid backpropagation, the reconstruction error of the reconstructor D1 on the generated data is reduced. Through this loss, the sensitivity of the decoder to the generated data is enhanced.

[0033] Preferably, for the reconstruction error, an adaptive anomaly detection threshold is adopted, which is set as the mean of the reconstruction error of the training set * 1.5 for discrimination: Threshold setting and detection: In every 10 epochs, the model enters the evaluation mode to calculate the value of the reconstruction error and perform anomaly detection. The threshold is calculated as follows: ; Calculate the reconstruction error of each batch, and multiply the mean of the reconstruction error by 1.5 as the detection threshold to distinguish normal data from abnormal data; Anomaly detection in the test phase: For each sample of the test data, calculate its reconstruction error. If the reconstruction error is greater than the threshold, it is determined as abnormal; otherwise, it is normal. Accuracy calculation: The detection accuracy is statistically calculated by comparing the predicted labels with the true labels.

[0034] Example 2: For the reconstruction task, not only can it determine whether a given data segment is abnormal, but also a set of annotation methods based on dynamic anomaly density is designed, and the effect is as Figure 2 shown. It can be seen from the figure that the provided dynamic density annotation method can finely annotate the abnormal regions of a segment of data.

[0035] Example 3: The model training curve is as Figure 3 shown. The abscissa is the number of training rounds, and the ordinate is the early warning accuracy rate. There are 0 missed alarms / 0 false alarms and 100% accuracy rate in the abnormal detection of the thrust oil sump oil smell.

Claims

1. A pseudo-visual anomaly detection method based on control quantity-response quantity, characterized in that, Including the following steps: S1. Determine the target measurement points that can respond to the business scenario, where the target measurement points are response quantities during the operation of the hydropower plant; S2. Analyze and find the control quantities related to the target measurement points; S3. Collect and clean the monitoring data containing control quantities and response quantities, and assemble the monitoring data into a data structure suitable for model training; S4. Use an improved 1D U-Net network architecture to process the data. The network architecture includes an encoder and a decoder. If it is used for a classification task, it also includes a classifier output head; for time series data, use 1D temporal convolution; S5. For the supervised learning task with faulty samples, use the cross-entropy loss function for model training; for the unsupervised sample learning task, use the reconstruction error for model training; based on the reconstruction error, use two reconstruction heads with the same structure, the D1 reconstructor and the D2 generator, to improve the task effect; S6. Design an adaptive anomaly determination threshold to distinguish normal and abnormal data based on the mean of the reconstruction errors of the training set multiplied by a preset coefficient; S7. Design a labeling method based on dynamic anomaly density for fine-grained labeling of the abnormal regions of a section of data; S8. Train a detection model based on the 1D U-Net network architecture to capture the data relationship between the control quantity and the response quantity; S9. Test and evaluate the detection model; S10. Use the trained detection model to conduct fault early warning on new monitoring data and output the early warning results.

2. The pseudo-visual anomaly detection method based on control quantity-response quantity according to claim 1, wherein The response quantity is the turbine speed, and the control quantity is the regulating valve opening. When analyzing and finding the control quantities related to the target measurement points, conduct measurement point early warning through the physical relationship between the control quantity and the response quantity.

3. A pseudo-visual anomaly detection method based on control quantity-response quantity according to claim 1, characterized in that In step S4, the network architecture is an architecture based on a convolutional neural network for capturing the relationship between measurement points and conducting early warning; the network architecture uses a 1D convolutional neural network, and uses a 1D convolutional kernel to capture both the dynamic characteristics between measurement points and the information in the time dimension to achieve early warning.

4. The pseudo-visual anomaly detection method based on control quantity-response quantity according to claim 3, characterized in that, In step S4, the encoder in the improved 1D U-Net network architecture is as follows: Input: Input time series data \(X\in\mathbb{R}\) LxC , where \(L\) is the length of the time step and \(C\) is the number of signal channels; Convolutional layer: The encoder contains multiple consecutive one-dimensional convolutional layers Conv1D, and cooperates with the max pooling operation MaxPooling to gradually reduce the time step length to compress the time dimension and increase the number of channels to extract high-level time series features; each convolutional operation is expressed as: )； Wherein, and are the weights and biases of the i-th layer of convolution, and ReLu is the activation function; Skip connection: In each layer of the encoder, the feature map is passed to the corresponding layer of the decoder to fuse features during the decoding phase, helping to recover information and improve accuracy.

5. The pseudo-visual anomaly detection method based on control quantity-response quantity according to claim 3, characterized in that, In step S4, the decoder in the improved 1D U-Net network architecture is as follows: Transposed convolution and feature fusion: The decoder contains multiple transposed convolution TransposeConvolution layers to gradually restore the time step dimension, that is, upsampling; the output of each layer is spliced with the skip connection features of the corresponding encoder layer to fuse the information as: )； In the formula, is the feature map of the decoder, is the feature map from the encoder, and the understanding of temporal features is deepened through feature concatenation.

6. The pseudo-visual anomaly detection method based on control quantity-response quantity according to claim 3, characterized in that, In step S4, the output layer in the improved 1D U-Net network architecture is as follows: The final output of the decoder generates a classification result through a fully connected layer or a 1D convolutional layer; for a binary classification task, use the Sigmoid activation function to map the output to a probability value between [0, 1]: ； where y represents the probability value that the input sequence belongs to the positive class, represents the Sigmoid activation function.

7. A pseudo-visual anomaly detection method based on control quantity-response quantity according to claim 1, characterized in that In step S4, 1D temporal convolution is used for the time series data. In 1D convolution, the filter, i.e., the convolution kernel, slides along the time step dimension and performs operations on each local region of the data, gradually extracting features at different levels. Assume the input data X ∈ R LxC , where L represents the length of the time step and C is the number of input channels. The convolution operation is expressed as: ； Among them, Y(t) is the eigenvalue of the convolution output at time step t, W is the weight matrix of the convolution kernel, with a size of K x C, where K represents the length of the kernel and C represents the number of input channels, and X(t + k - 1, c) represents the value of the input signal at channel c and time step t + k - 1.

8. A pseudo-visual anomaly detection method based on control quantity-response quantity according to claim 1, characterized in that, For supervised learning with faulty samples, cross-entropy loss is used for learning: ； Among them, is the true label, is the probability value predicted by the model, and N is the number of samples.

9. A pseudo-visual anomaly detection method based on control quantity-response quantity according to claim 8, characterized in that For the task of unsupervised samples, reconstruction error is used for learning: ； Based on the reconstruction error, two reconstruction head reconstructors D1 and generator D2 with the same structure are used to improve the task effect, including: Training part of the reconstructor D1: During the training of the reconstructor D1, two loss terms are designed: Reconstruction error L1: ； The reconstruction error L1 calculates the reconstruction error of the reconstructor D1 for the real data; Adversarial error L2: ； Reconstruction error It calculates the reconstruction error of the reconstructor D1 for the generated data, i.e., the abnormal data; Total loss : ； Among them, the weight coefficient 0.25 controls the balance between the two losses; Training part of the generator D2: The training of the generator D2 includes two main loss terms: Reconstruction error L3: ； The reconstruction error L3 of the generator D2 represents the gap between the generated data and the real data; Adversarial error L4: ； By introducing the detach function to avoid backpropagation, the reconstruction error of the reconstructor D1 on the generated data is reduced.

10. A pseudo-visual anomaly detection method based on control quantity-response quantity according to claim 9, characterized in that For the reconstruction error, an adaptive anomaly detection threshold is adopted, and the mean of the reconstruction error of the training set * 1.5 is used for discrimination: Threshold setting and detection: In every 10 epochs, the model enters the evaluation mode to calculate the value of the reconstruction error and perform anomaly detection. The threshold is calculated as follows: ； Calculate the reconstruction error of each batch, and multiply the mean of the reconstruction error by 1.5 as the detection threshold to distinguish normal data and abnormal data; Anomaly detection in the test stage: For each sample of the test data, calculate its reconstruction error. If its reconstruction error is greater than the threshold, it is determined as abnormal, otherwise it is normal; Accuracy calculation: The detection accuracy is statistically calculated by comparing the predicted labels and the true labels.

Images