Vibrating screen abnormity early warning system and method based on unbalanced data enhancement

Abstract

The invention discloses a vibrating screen abnormity early warning system and method based on unbalanced data enhancement. The early warning method comprises the following steps: processing an original vibration data set through a self-supervised dual-code adversarial generative network and enhanced transformation processing to obtain a generated vibration data set; through the original vibration data set and the generated vibration data set, a vibration screen operation state identification model constructed based on a multi-kernel CNN network and a classifier is trained; and inputting the predicted vibration data into the trained vibrating screen operation state recognition model, carrying out digital coding on the output to obtain a predicted operation state code, and combining the predicted operation state code with the abnormal score to cooperatively judge the abnormal operation state of the vibrating screen. According to the method, high-precision early warning of the vibration screen abnormity can be realized through technologies of adaptive data generation, semi-supervised learning, metric learning and the like.

Description

Technical Field

[0001] The present invention belongs to the technical field of mineral processing equipment, and particularly relates to a vibrating screen abnormal warning system and method based on unbalanced data augmentation. Background Art

[0002] As an important equipment in coal washing and processing, the operating ability of the vibrating screen has a significant impact on the screening efficiency of coal preparation and the sustainable development of the coal industry. Therefore, it is very necessary to timely warn and handle the problems of easy wear and faults caused by the long-term high-load operation of the vibrating screen in a harsh environment to avoid equipment damage and production accidents.

[0003] Currently, the identification of the operating state of the vibrating screen usually relies on manual inspection. This method is difficult to effectively capture potential abnormalities of the equipment and will lead to unnecessary shutdowns and repairs. Therefore, how to automatically identify the operating state of the vibrating screen using intelligent methods has become a key research content in this field. Among them, research on collecting equipment data through sensors and combining traditional machine learning methods for state identification has achieved many results. However, traditional machine learning methods often rely on complex data feature engineering and expert experience and have limitations in dealing with nonlinear and complex dynamic systems.

[0004] In recent years, the rapid development of deep learning has provided new ideas for solving the above problems. Deep learning has powerful feature mining and complex modeling capabilities and can effectively learn the complex relationship between equipment monitoring data and states. However, the success of deep learning usually depends on a large amount of labeled data. In the industrial field, due to the complexity and variability of the industrial environment, it is difficult to obtain sufficient and balanced distributed labeled data, which limits the application of deep learning in solving the above problems. Summary of the Invention

[0005] To solve the problems of the prior art, the purpose of the present invention is to propose a vibrating screen abnormal warning system and method based on unbalanced data augmentation. The system or method can realize timely warning and positioning of abnormal operation of the vibrating screen through the efficient combination of technologies such as adaptive data augmentation, semi-supervised learning, metric learning, coding correction mechanism, and abnormal scores, reducing equipment downtime and wasting of manpower and material resources.

[0006] The technical solution of the present invention is as follows:

[0007] A vibrating screen abnormal warning method based on unbalanced data augmentation, comprising:

[0008] S1 Based on the vibration data at the key positions of the vibrating screen measured and their corresponding operating state labels, i.e., the original vibration data set, a generated vibration data set with diversity and balanced distribution is obtained through a self-supervised dual-encoding adversarial generation network and enhanced transformation processing; wherein, the vibration data includes vibration signals and / or their corresponding time-frequency diagrams that are time-sequentially aligned.

[0009] S2 The operating state recognition model of the vibrating screen is trained with the original vibration data set and the generated vibration data set, and the operating state of the vibrating screen is recognized by the trained operating state recognition model of the vibrating screen; wherein, the operating state recognition model of the vibrating screen is constructed based on a multi-core CNN network and a classifier.

[0010] S3 According to the vibration data at the key positions of the vibrating screen measured and their historical data, the predicted vibration data of the vibrating screen is obtained through the state prediction model, the predicted vibration data is input into the trained operating state recognition model of the vibrating screen, the predicted operating state is obtained, digital encoding is performed on it to obtain the predicted operating state encoding, and the predicted operating state encoding or the predicted operating state encoding corrected by the encoding correction mechanism is combined with the anomaly score obtained from the similarity measure between the predicted vibration data and the measured vibration data to jointly determine the abnormal operating state of the vibrating screen.

[0011] Preferably, the operating state label is obtained according to the working condition diary.

[0012] Preferably, the operating state labels include normal, partial load, no load, and full load.

[0013] Preferably, the key positions of the vibrating screen include one or more of the following: at the fuel tank of the vibrating screen, at the two side plates of the vibrating screen, at the cross beam of the vibrating screen, at the exciter of the vibrating screen, at the inlet and outlet of the vibrating screen, and at the four spring bases of the vibrating screen.

[0014] Preferably, the processing process of the self-supervised dual-encoding adversarial generation network includes:

[0015] S104 Perform initial information measurement on the data in the original vibration data set, and assign initial weights to the data of different categories based on the obtained measurement values to obtain the original vibration data set with weight assignment.

[0016] S105 Train the self-supervised dual-encoding adversarial generation network according to the original vibration data set with weight assignment, and dynamically adjust the weights during the training according to the generated vibration data of the self-supervised dual-encoding adversarial generation network until the weight fluctuations of the generated vibration data of all categories obtained do not exceed 5%, and obtain the generated vibration data with balanced distribution.

[0017] Preferably, the enhanced transformation processing includes:

[0018] S106 performs enhanced transformations on the vibration data in the original vibration dataset in various ways, calculates the symmetric KL divergence and distance for each enhanced transformation method, and obtains the transformation degree of each enhanced transformation.

[0019] S107 assigns adaptive weights to different enhanced transformation methods according to the transformation degree, adds the adaptive weights to the loss function of the self-supervised dual-encoding adversarial generation network, and obtains a self-supervised dual-encoding adversarial generation network with adaptive diversity expansion ability.

[0020] S108 inputs the generated vibration data with balanced distribution into the self-supervised dual-encoding adversarial generation network with adaptive diversity expansion ability to obtain generated vibration data with diversity and balanced distribution, and forms the generated vibration dataset with diversity and balanced distribution from it.

[0021] Preferably, the initial information metric adopts the information entropy and the multi-kernel maximum mean discrepancy method.

[0022] Preferably, the methods of enhanced transformation include: performing two or more of the following on the time-series aligned vibration signal: adding noise, rotating, masking, inverting, scaling, randomly intercepting and transposing; performing two or more of the following on the labeled time-frequency diagram: rotating, masking, adding noise, local linear interpolation, non-linear mixing.

[0023] Preferably, the self-supervised dual-encoding adversarial generation network includes: a self-supervised prediction subnet that performs vibration signal prediction based on the time-series aligned vibration signal to obtain a predicted time-series vibration signal, a time-series data encoder that encodes the predicted time-series vibration signal obtained by the self-supervised prediction subnet to obtain an encoded time-series signal, a time-frequency diagram encoder that encodes the time-frequency diagram to obtain an encoded time-frequency diagram, a signal feature extraction layer and a time-frequency diagram feature extraction layer that respectively extract features from the encoded time-series signal and the encoded time-frequency diagram to obtain a time-series signal feature and a time-frequency diagram feature, a fusion layer that fuses the time-series signal feature and the time-frequency diagram feature to obtain fusion data, a decoder that performs decoding processing on the fusion data to obtain decoded data, a generator that uses the decoded data with added noise as input, and a discriminator that discriminates the authenticity of the generated data output by the generator; wherein, the self-supervised prediction subnet is formed by a Bi-LSTM network.

[0024] Preferably, in step S2, the training includes:

[0025] S201 performs Fourier transform on the time-aligned vibration signals in the original vibration data set to obtain their corresponding frequency components and amplitude values, selects the 5 most frequently occurring frequency components among the 10 frequency components with the largest amplitude values, converts them into a time scale sequence based on the relationship between period and frequency, segments the time-aligned vibration signals using each time scale value in the time scale sequence as a segmentation window size to obtain a number of non-overlapping segmented signal segments, and splices the segmented signal segments up and down into a two-dimensional matrix according to the time relationship to obtain labeled reconstructed data; performs the same processing on the vibration signals in the generated vibration data set to obtain unlabeled reconstructed data;

[0026] S202 performs a first-stage training on the vibration screen operation state recognition model using the labeled reconstructed data and its corresponding time-frequency graph to obtain a partially optimized vibration screen operation state recognition model;

[0027] S203 performs a second-stage training on the partially optimized vibration screen operation state recognition model using the labeled reconstructed data and its corresponding time-frequency diagram and the unlabeled reconstructed data and its corresponding time-frequency diagram. During the training, a metric learning method and its corresponding loss function are used to shorten the distribution distance of the labeled reconstructed data and the unlabeled reconstructed data in the feature space until the identification network cannot identify whether the feature comes from the labeled reconstructed data or the unlabeled reconstructed data, thereby obtaining the trained vibration screen operation state recognition model.

[0028] Preferably, S2 further includes training, verifying and testing a vibration screen operation state recognition model using the original vibration data set and the generated vibration data set, and performing vibration screen operation state recognition using the vibration screen operation state recognition model that passes the test.

[0029] Preferably, the vibration screen operation state recognition model includes a feature extraction model and a classifier, wherein the feature extraction model is composed of multiple feature blocks connected in a residual manner, and each feature block is composed of two multi-core CNN networks and one fusion network, wherein one multi-core CNN network is used to extract reconstructed data features, and one multi-core CNN network is used to extract time-frequency graph features; the fusion network is used to fuse the extracted reconstructed data features and time-frequency graph features.

[0030] Preferably, the state prediction model is constructed based on an LSTM gate mechanism and a Transformer with a multi-head attention mechanism.

[0031] More preferably, the state prediction model includes: 3 encoders with gated multi-head self-attention mechanisms, each encoder including a multi-head attention layer connected by a residual connection, a first normalization layer, a feed-forward network layer, a second normalization layer, and a first fully-connected layer, where the multi-head attention layer includes 4 different self-attention heads; a Sigmoid activation layer connected to the multi-head attention layer and the first normalization layer; a third normalization layer connected to the first fully-connected layer; a second fully-connected layer connected to the third normalization layer; a number of feature extraction blocks connected to the second fully-connected layer; each feature extraction block contains a convolutional layer, a fully-connected layer, and a Relu activation layer.

[0032] Preferably, the encoding correction mechanism includes: inputting the vibration data obtained in real time into the trained vibration sieve operating state recognition model, encoding the output operating state recognition result to obtain a real-time recognition code, and judging and optimizing the jump code in the predicted operating state code according to the real-time recognition code and its historical data, i.e., the historical recognition code.

[0033] The present invention further provides a vibration sieve abnormal warning system, which includes a vibration sieve adaptive data enhancement module and a vibration sieve abnormal warning module. Among them, the vibration sieve adaptive data enhancement module is used to obtain a generated vibration data set with diversity and balanced distribution through a self-supervised dual-encoding adversarial generation network and enhanced transformation processing based on the vibration data of the key positions of the vibration sieve measured in real time and its corresponding operating state labels, i.e., the original vibration data set. The vibration sieve abnormal warning module is used to train the vibration sieve operating state recognition model through the original vibration data set and the generated vibration data set, identify the operating state of the vibration sieve by the trained vibration sieve operating state recognition model, and obtain the predicted vibration data of the vibration sieve through the state prediction model according to the vibration data of the key positions of the vibration sieve measured in real time and its historical data. Input the predicted vibration data into the trained vibration sieve operating state recognition model to obtain the predicted operating state, encode it digitally to obtain the predicted operating state code, and combine the predicted operating state code or the predicted operating state code corrected by the encoding correction mechanism with the abnormal score obtained by the similarity metric between the predicted vibration data and the measured vibration data to jointly determine the abnormal operating state of the vibration sieve.

[0034] Preferably, the vibrating screen adaptive data augmentation module includes a data generation sub-module and an adaptive diversification augmentation sub-module. Among them, the data generation sub-module is used to obtain generated vibration data with a balanced distribution through a self-supervised dual-encoding adversarial generation network based on the vibration data at the key positions of the vibrating screen measured and its corresponding operating state label, that is, the original vibration data set; the adaptive diversification augmentation sub-module is used to perform adaptive enhancement transformation on the vibration data in the original vibration data set and the generated vibration data to obtain vibration data with diversity and balanced distribution.

[0035] Preferably, the vibrating screen abnormal warning module includes an operating state recognition sub-module and an abnormal warning sub-module. Among them, the operating state recognition sub-module is used to train the vibrating screen operating state recognition model through the original vibration data set and the generated vibration data set, and the trained vibrating screen operating state recognition model is used to recognize the operating state of the vibrating screen; the abnormal warning sub-module is used to obtain the predicted vibration data of the vibrating screen through the state prediction model according to the vibration data at the key positions of the vibrating screen measured and its historical data, input the predicted vibration data into the trained vibrating screen operating state recognition model to obtain the predicted operating state, perform digital encoding on it to obtain the predicted operating state code, and combine the predicted operating state code or the predicted operating state code corrected by the encoding correction mechanism with the abnormal score obtained by the similarity measure between the predicted vibration data and the measured vibration data to jointly determine the abnormal operating state of the vibrating screen.

[0036] The present invention has the following beneficial effects:

[0037] (1) By integrating the methods of unbalanced data generation and adaptive diversification augmentation, the present invention effectively solves the problem of unbalanced data of vibrating screens in industrial sites, and at the same time combines real-time operating state recognition and abnormal warning, which can overcome the difficulties of difficult extraction and prediction of state signal features. Compared with traditional machine learning methods, the present invention can achieve more accurate abnormal warning and positioning of vibrating screens;

[0038] (2) In some specific embodiments, the present invention can skillfully handle the imbalance of various operating state data of vibrating screens, and through the comprehensive application of information entropy, metric learning, self-supervised learning, and adversarial generation networks, etc., realize the efficient augmentation of data and provide a high-quality data basis for downstream tasks;

[0039] (3) In some specific embodiments, the present invention uses Fourier transform to perform multi-scale two-dimensional reconstruction on the original vibration signal, and constructs a vibration sieve operating state recognition model based on multi-core CNN trained in two stages. The recognition model integrates semi-supervised learning and metric learning, makes full use of the information of labeled and unlabeled data, significantly reduces the dependence of the model on labeled data, and constructs a discriminative network to make it unable to accurately identify the source of features, so as to further enhance the generalization and robustness of the model, and demonstrates excellent performance and flexibility in the recognition of the vibration sieve operating state.

[0040] (4) In some specific embodiments, in the abnormal warning part of the vibration sieve, the present invention combines the LSTM gate mechanism and the Transformer architecture to fully capture the key information of longer historical data, realize high-precision prediction of future operation data, and combine the trained state recognition model, coding correction mechanism and abnormal score to jointly capture the early abnormal state information of the vibration sieve, and improve the accuracy of abnormal warning of the vibration sieve. Description of the Drawings

[0041] Figure 1 It is a flowchart of the warning method in the embodiment of the present invention.

[0042] Figure 2 It is a schematic diagram of the adaptive data augmentation unit in the embodiment of the present invention.

[0043] Figure 3 It is a schematic diagram of the abnormal warning unit in the embodiment of the present invention. Detailed Embodiments

[0044] The present invention will be described in detail below with reference to the embodiments and the drawings. However, it should be understood that the embodiments and the drawings are only used for exemplary description of the present invention, and do not constitute any limitation to the protection scope of the present invention. All reasonable transformations and combinations within the scope of the inventive concept of the present invention fall within the protection scope of the present invention.

[0045] Embodiment 1

[0046] Referring to the attached Figure 1 , the vibration sieve abnormal warning method based on unbalanced data augmentation of the present invention includes the following steps:

[0047] S1 Based on the vibration data measured at the key positions of the vibration sieve and their corresponding operating state labels, that is, the original vibration data set, a generated vibration data set with diversity and balanced distribution is obtained through a self-supervised dual-encoding adversarial generation network and enhancement transformation processing; wherein, the vibration data includes vibration signals that have been time-aligned and / or their corresponding time-frequency diagrams.

[0048] In some specific embodiments, the operating state markers include normal and abnormal.

[0049] In other specific embodiments, the operating state markers include normal and faulty, wherein the faults are further marked differently according to the fault types.

[0050] In some specific embodiments, the operating state markers can be obtained according to the working condition diary.

[0051] Preferably, the key positions of the vibrating screen include one or more of the following: the fuel tank of the vibrating screen, the two side plates of the vibrating screen, the cross beam of the vibrating screen, the vibrator of the vibrating screen, the inlet and outlet of the vibrating screen, and the four spring bases of the vibrating screen.

[0052] Preferably, the operating state markers include normal, off-load, no-load, and full-load.

[0053] Preferably, the acquisition of the original vibration data set includes:

[0054] S101 Measure or extract the vibration signals of the vibrating screen at the key positions from the vibrating screen database of the coal preparation plant, and perform time series alignment to obtain the vibration signals with time series alignment;

[0055] S102 Based on the records of the operating state of the vibrating screen in the working condition log, perform operating state marking on the vibration signals with time series alignment to obtain marked vibration signals;

[0056] S103 Perform time-frequency diagram conversion on the marked vibration signals to obtain their corresponding marked time-frequency diagrams, and the original vibration data set is composed of the marked vibration signals and their corresponding marked time-frequency diagrams.

[0057] Preferably, the processing process of the self-supervised dual-encoding adversarial generation network includes:

[0058] S104 Perform initial information measurement on the data in the original vibration data set, and assign initial weights to the data of different categories based on the obtained measurement values to obtain the original vibration data set with weight assignment;

[0059] S105 Train the self-supervised dual-encoding adversarial generation network according to the original vibration data set with weight assignment, and dynamically adjust the weights according to the generated vibration data of the self-supervised dual-encoding adversarial generation network during training until the weight fluctuations of the generated vibration data of all categories obtained do not exceed 5%, and obtain the generated vibration data with balanced distribution.

[0060] Among them, more preferably, the initial information measurement adopts the information entropy and multi-kernel maximum mean discrepancy method.

[0061] Preferably, the enhancement transformation processing includes:

[0062] S106 Perform enhancement transformations on the vibration data in the original vibration dataset in various ways, calculate the symmetric KL divergence and distance of each enhancement transformation method, and obtain the transformation degree of each enhancement transformation;

[0063] S107 Assign adaptive weights to different enhancement transformation methods according to the transformation degree, add the adaptive weights to the loss function of the self-supervised dual-encoding adversarial generation network, and obtain a self-supervised dual-encoding adversarial generation network with adaptive diversity expansion ability;

[0064] S108 Input the generated vibration data with balanced distribution into the self-supervised dual-encoding adversarial generation network with adaptive diversity expansion ability, obtain generated vibration data with diversity and balanced distribution, and form the generated vibration dataset with diversity and balanced distribution from it.

[0065] Among them, more preferably, the methods of the enhancement transformation include: performing two or more of the following on the time-series aligned vibration signals: adding noise, rotating, masking, inverting, scaling, randomly intercepting and transposing; performing two or more of the following on the labeled time-frequency diagram: rotating, masking, adding noise, local linear interpolation, non-linear mixing.

[0066] Through the above step S1, the present invention can achieve diversified expansion of the original and generated data, strengthen data diversity, provide balanced and rich training data for the recognition model, and improve the generalization performance of the model.

[0067] Preferably, referring to the appendix Figure 2 , the self-supervised dual-encoding adversarial generation network includes: a self-supervised prediction subnet that performs vibration signal prediction based on the time-series aligned vibration signals to obtain predicted time-series vibration signals, a time-series data encoder that encodes the predicted time-series vibration signals obtained by the self-supervised prediction subnet to obtain encoded time-series signals, a time-frequency diagram encoder that encodes the time-frequency diagram to obtain an encoded time-frequency diagram, a signal feature extraction layer and a time-frequency diagram feature extraction layer that respectively extract features from the encoded time-series signals and the encoded time-frequency diagram to obtain time-series signal features and time-frequency diagram features, a fusion layer that fuses the time-series signal features and the time-frequency diagram features to obtain fusion data, a decoder that performs decoding processing on the fusion data to obtain decoded data, a generator that uses the decoded data with added noise as input, and a discriminator that discriminates the authenticity of the generated data output by the generator; among them, the self-supervised prediction subnet is formed by a Bi-LSTM network.

[0068] Among them, more preferably, the time series data encoder includes an input layer with a dimension of (Batch, 512, 10, 20) and 8 convolutional layers, and the size of its convolutional kernel is randomly selected between 1 and 4. The time-frequency map encoder has the same structure as the time series data encoder.

[0069] Among them, more preferably, the time series signal features and the time-frequency map features are vertically spliced and fused through a fully connected layer.

[0070] Among them, more preferably, the decoder includes an input layer, 5 hidden layers based on regular normalization processing, and an output layer.

[0071] Among them, more preferably, the generator includes an input layer, 5 hidden layers and an output layer. Among them, the input layer has 53 nodes, and the 5 hidden layers respectively contain 200, 400, 600, 1200, and 1850 nodes. Each hidden layer uses the LeakyReLU activation function. The Dropout value of the middle hidden layer is 0.2, and the Dropout values of the remaining hidden layers are 0.3. The output layer contains 512 nodes, and the activation function is tanh.

[0072] Among them, more preferably, the discriminator is composed of an 8-layer neural network, and the number of nodes is 125, 512, 1024, 2048, 1024, 512, 128, and 2 respectively. The middle three hidden layers use the LeakyReLU activation function, and the Dropout value of the third layer is 0.3.

[0073] Among them, more preferably, the self-supervised dual-encoding adversarial generation network is trained by alternately optimizing the generator and the discriminator, and a total of 200 epochs are performed. The amount of information of various vibrating screen operating state data is updated every 20 epochs.

[0074] S2 trains the vibrating screen operating state recognition model through the original vibration data set and the generated vibration data set, and the trained vibrating screen operating state recognition model is used to recognize the vibrating screen operating state; among them, the vibrating screen operating state recognition model is constructed based on a multi-core CNN network and a classifier.

[0075] Preferably, S2 also includes training, validating and testing the vibrating screen operating state recognition model through the original vibration data set and the generated vibration data set, and the vibrating screen operating state recognition model that passes the test is used to recognize the vibrating screen operating state.

[0076] Preferably, the training includes:

[0077] S201 performs Fourier transform on the time-aligned vibration signals in the original vibration data set to obtain their corresponding frequency components and amplitude values, sorts all frequency components by amplitude value, selects the 5 most frequently occurring frequency components among the 10 frequency components with the largest amplitude values, and converts these 5 frequencies into corresponding time scale sequences (such as [1s, 2s, 5s, 8s, 10s]) based on the relationship between period and frequency. Subsequently, based on these time scales, the original one-dimensional vibration signal is divided into multiple non-overlapping signal segments according to different scales, and spliced into a two-dimensional matrix in chronological order. The number of rows of the matrix is equal to the result of dividing the total duration of the signal by the time scale, and the number of columns is determined by the specific time scale and sampling rate, thereby obtaining labeled reconstructed data; the vibration signals in the generated vibration data set are subjected to the same processing to obtain unlabeled reconstructed data;

[0078] S202 performs a first-stage training on the vibration screen operation state recognition model using the labeled reconstructed data and its corresponding time-frequency graph to obtain a partially optimized vibration screen operation state recognition model;

[0079] S203 performs a second-stage training on the partially optimized vibration screen operation state recognition model using the labeled reconstructed data and its corresponding time-frequency diagram and the unlabeled reconstructed data and its corresponding time-frequency diagram. During the training, a metric learning method and its corresponding loss function are used to shorten the distribution distance of the labeled reconstructed data and the unlabeled reconstructed data in the feature space until the identification network cannot identify whether the feature comes from the labeled reconstructed data or the unlabeled reconstructed data, thereby obtaining the trained vibration screen operation state recognition model.

[0080] In some more specific implementations, the metric learning method is a multi-kernel maximum mean difference method.

[0081] Preferably, the vibration screen operation state recognition model includes a feature extraction model and a classifier, wherein the feature extraction model is composed of a plurality of feature blocks connected in a residual manner, and each feature block is composed of several multi-core CNN networks and a fusion network.

[0082] Preferably, refer to the attached Figure 3 The feature extraction model includes four feature blocks connected in a residual manner. Each feature block consists of two multi-core CNN networks and one fusion network. One multi-core CNN network is used to extract reconstructed data features, and one multi-core CNN network is used to extract time-frequency graph features. The fusion network is used to fuse the extracted reconstructed data features and time-frequency graph features.

[0083] More preferably, each multi-core CNN network includes 3 convolutional blocks. Among them, the first two convolutional blocks have the same structure, each including 1 convolutional layer, 1 up-connection layer, 1 batch normalization layer, and 1 activation layer with the activation function Relu. Among them, the convolutional layer contains convolutional kernels with a prime number between 1 and 19, and the third convolutional block contains a convolutional layer with a convolutional kernel size of 2 and 3.

[0084] With the above structure, the first two convolutional blocks can cover the receptive field sizes of all even numbers, and the third convolutional block can cover the receptive field sizes of all integers within a certain range, comprehensively improving the ability to extract effective features in the state recognition network.

[0085] More preferably, the fusion network includes 4 fully connected layers, 1 activation layer with the activation function Relu, and 1 activation layer with the activation function Sigmoid. Among them, the number of nodes in the fully connected layers is 512, 256, 125, and 125 respectively.

[0086] With the above structure, the fusion network can adaptively fuse and restore the features extracted from the two types of data.

[0087] Preferably, the classifier includes an input layer, 3 convolutional blocks, a fully connected layer, an activation layer with the activation function Relu, a Dropout layer with a Dropout value of 0.5, and an output layer. Among them, each convolutional block contains 1 convolutional layer with a convolutional kernel size of 3×3, 1 activation layer with the activation function Relu, and 1 max-pooling layer with a pooling window size of 2×2.

[0088] Preferably, the discriminant network includes a neural network layer with 6 layers and the number of nodes being 512, 1024, 2048, 1024, 512, 128, and 2 respectively. Among them, the middle three layers use the Relu activation function, and the third layer sets the Dropout value to 0.3.

[0089] S3 According to the measured vibration data of the key positions of the vibrating screen and its historical data, obtain the predicted vibration data of the vibrating screen through the state prediction model. Input the predicted vibration data into the trained vibrating screen operation state recognition model to obtain the predicted operation state, digitally encode it to obtain the predicted operation state code. Combine the predicted operation state code or the predicted operation state code corrected by the code correction mechanism with the anomaly score obtained from the similarity measure between the predicted vibration data and the measured vibration data to jointly determine the abnormal operation state of the vibrating screen.

[0090] Preferably, the state prediction model is constructed based on the LSTM gate mechanism and the Transformer with a multi-head attention mechanism.

[0091] More preferably, the state prediction model includes: 3 encoders with gated multi-head self-attention, each encoder including a multi-head attention layer, a first normalization layer, a feed-forward network layer, a second normalization layer, and a first fully connected layer connected by residual connections, where the multi-head attention layer includes 4 different self-attention heads; a Sigmoid activation layer connected to the multi-head attention layer and the first normalization layer as a gating mechanism; a third normalization layer connected to the first fully connected layer; a second fully connected layer connected to the third normalization layer; and a number of feature extraction blocks connected to the second fully connected layer; each feature extraction block contains a convolutional layer, a fully connected layer, and a Relu activation layer.

[0092] In a specific embodiment, the processing process of the state prediction model includes:

[0093] S301 Input the vector x (window * 53) of the vibration signals aligned in time series into the state prediction model, and allocate a set of query vectors q, key vectors k, and value vectors v for this vector through each self-attention head of the multi-head attention layer;

[0094] S302 Introduce a gating mechanism, that is, use the value of each group of q * k processed by the activation function Sigmoid as an attention score of the vector x, normalize all the attention scores of the vector x obtained according to each self-attention head, and perform Softmax activation function activation processing to obtain the attention score V of each self-attention head;

[0095] S303 For each self-attention head, perform weighted summation of the attention score V and the original input vector x to obtain the output result z of this head, splice the z values of the 4 self-attention heads column by column to form a feature matrix, and after passing through the fully connected layer, obtain the output feature vector Z;

[0096] S304 Add the vector x and the output feature vector Z and perform normalization processing through the third normalization layer to obtain a normalized vector;

[0097] S305 Input the normalized vector into the second fully connected layer including two convolutional layers, and then input it into the feature extraction block. After being processed by a number of feature extraction blocks, the predicted vibration data is obtained; where the first layer in the second fully connected layer uses the activation Relu, and the second layer does not use an activation function.

[0098] Preferably, the methods adopted for the similarity metric include dynamic time warping and / or multi-kernel maximum mean discrepancy, etc.

[0099] Preferably, the encoding correction mechanism includes: inputting the vibration data obtained in real time into the trained vibration sieve operating state recognition model, encoding the output operating state recognition result to obtain a real-time recognition code, and judging and optimizing the jump code in the predicted operating state code according to the real-time recognition code and its historical data, i.e., the historical recognition code.

[0100] In a specific embodiment, the encoding correction mechanism is as follows: when the code for the normal state is 0, the code for the partial load state is 1, the code for the no-load state is 2, and the code for the full-load state is 3, for the first-occurring jump code in the real-time recognition result, retrieve 50 historical recognition codes before and after the jump code, and combine the next 10 consecutive real-time recognition code results. Using the mode of these codes as a reference, correct the jump code. For example, correct the jump code "1" in the recognition sequence "000000101000..." to "0".

[0101] In a specific embodiment, the method for jointly determining the abnormal operating state of the vibrating screen is as follows: when the abnormal score shows an increasing trend over time in the code state 0, the system needs to promptly remind to update the state recognition model. If the abnormal score shows an increasing trend over time in other code states, then based on the operating state of the vibrating screen and the dimension of the maximum abnormal score, give an early warning and locate the abnormal state of the vibrating screen.

[0102] Embodiment 2

[0103] The present invention further provides a system that can implement the early warning method described in Embodiment 1, which includes: a vibrating screen adaptive data enhancement module and a vibrating screen abnormal early warning module. Among them, the vibrating screen adaptive data enhancement module is used to obtain a generated vibration data set with diversity and balanced distribution through a self-supervised dual-encoding adversarial generation network and enhancement transformation processing based on the vibration data at the key positions of the measured vibrating screen and its corresponding operating state markers, i.e., the original vibration data set. The vibrating screen abnormal early warning module is used to train the vibrating screen operating state recognition model through the original vibration data set and the generated vibration data set. The trained vibrating screen operating state recognition model is used to recognize the operating state of the vibrating screen and obtain the predicted vibration data of the vibrating screen through the state prediction model according to the vibration data at the key positions of the measured vibrating screen and its historical data. Input the predicted vibration data into the trained vibrating screen operating state recognition model to obtain the predicted operating state, encode it to obtain the predicted operating state code, and combine the predicted operating state code or the predicted operating state code corrected by the encoding correction mechanism with the abnormal score obtained by the similarity measurement between the predicted vibration data and the measured vibration data to jointly determine the abnormal operating state of the vibrating screen.

[0104] Preferably, the vibrating screen adaptive data enhancement module includes a data generation sub-module and an adaptive diversification expansion sub-module. Among them, the data generation sub-module is used to obtain the generated vibration data with a balanced distribution through a self-supervised dual-encoding adversarial generation network based on the vibration data at the key positions of the vibrating screen measured and its corresponding operation status label, that is, the original vibration data set; the adaptive diversification expansion sub-module is used to perform adaptive enhancement transformation on the vibration data in the original vibration data set and the generated vibration data to obtain vibration data with diversity and a balanced distribution.

[0105] Preferably, the vibrating screen abnormal warning module includes an operation status recognition sub-module and an abnormal warning sub-module. Among them, the operation status recognition sub-module is used to train the vibrating screen operation status recognition model through the original vibration data set and the generated vibration data set, and the trained vibrating screen operation status recognition model is used to recognize the operation status of the vibrating screen; the abnormal warning sub-module is used to obtain the predicted vibration data of the vibrating screen through the state prediction model according to the vibration data at the key positions of the vibrating screen measured and its historical data, input the predicted vibration data into the trained vibrating screen operation status recognition model to obtain the predicted operation status, perform digital encoding on it to obtain the predicted operation status code, and combine the predicted operation status code or the predicted operation status code corrected by the encoding correction mechanism with the abnormal score obtained by the similarity measurement between the predicted vibration data and the measured vibration data to jointly determine the abnormal operation status of the vibrating screen.

[0106] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. An abnormal warning method for vibrating screens based on unbalanced data augmentation, characterized in that It includes: S1 Based on the vibration data at the key positions of the vibrating screen measured and its corresponding operation status markers, namely the original vibration data set, a generated vibration data set with diversity and balanced distribution is obtained through a self-supervised dual-encoding adversarial generation network and enhanced transformation processing; wherein, the vibration data includes vibration signals and / or their corresponding time-frequency diagrams that have been time-aligned. S2 The operation status recognition model of the vibrating screen is trained with the original vibration data set and the generated vibration data set, and the operation status of the vibrating screen is recognized by the trained operation status recognition model of the vibrating screen; wherein, the operation status recognition model of the vibrating screen is constructed based on a multi-core CNN network and a classifier. S3 According to the vibration data at the key positions of the vibrating screen measured and its historical data, the predicted vibration data of the vibrating screen is obtained through a state prediction model, the predicted vibration data is input into the trained operation status recognition model of the vibrating screen, the predicted operation status is obtained, digitally encoded, the predicted operation status code is obtained, and the predicted operation status code or the predicted operation status code corrected by an encoding correction mechanism is combined with the anomaly score obtained based on the similarity measure between the predicted vibration data and the measured vibration data to jointly determine the abnormal operation status of the vibrating screen.

2. The abnormal warning method for the vibrating screen according to claim 1, characterized in that In step S1, the operation status markers are obtained according to the working condition diary; and / or, the operation status markers include normal, partial load, no load, and full load; and / or, the key positions of the vibrating screen include one or more of the following: at the fuel tank of the vibrating screen, at the two side plates of the vibrating screen, at the cross beam of the vibrating screen, at the vibrator of the vibrating screen, at the inlet and outlet of the vibrating screen, and at the four spring bases of the vibrating screen.

3. The abnormal warning method of the vibrating screen according to claim 1, characterized in that, In step S1, the processing process of the self-supervised dual-encoding adversarial generation network includes: S104 Perform an initial information measure on the data in the original vibration data set, and assign initial weights to the data of different categories based on the obtained measure values to obtain an original vibration data set with weight assignment. S105 Train the self-supervised dual-encoding adversarial generation network according to the original vibration data set with weight assignment, and dynamically adjust the weights during the training according to the generated vibration data of the self-supervised dual-encoding adversarial generation network until the weight fluctuations of the generated vibration data of all categories obtained do not exceed 5%, and obtain the generated vibration data with a balanced distribution. and / or, the enhanced transformation processing includes: S106 Perform enhanced transformations on the vibration data in the original vibration data set in various ways, and calculate the symmetric KL divergence and distance of each enhanced transformation method to obtain the transformation degree of each enhanced transformation. S107 Assign adaptive weights to different enhanced transformation methods according to the transformation degree, and add the adaptive weights to the loss function of the self-supervised dual-encoding adversarial generation network to obtain a self-supervised dual-encoding adversarial generation network with an adaptive diversity expansion ability. S108 inputs the generated vibration data with balanced distribution into the self-supervised dual-coding adversarial generation network with adaptive diversity expansion capability to obtain generated vibration data with diverse and balanced distribution, thereby forming the generated vibration data set with diverse and balanced distribution.

4. The abnormal warning method of the vibrating screen according to claim 3, wherein in, The initial information metric adopts information entropy and multi-core maximum mean difference method; and / or the enhanced transformation method includes: performing on the time-aligned vibration signal: two or more of: noising, rotation, masking, inversion, scaling, and random truncation and transposition; performing on the marked time-frequency graph: two or more of: rotation, masking, noising, local linear interpolation, and nonlinear mixing.

5. The abnormal warning method of the vibrating screen according to claim 3, characterized in that, The self-supervised dual-coding adversarial generation network includes: a self-supervised prediction subnet that predicts a vibration signal based on a time-series aligned vibration signal to obtain a predicted time-series vibration signal, a time-series data encoder that encodes the predicted time-series vibration signal obtained by the self-supervised prediction subnet to obtain an encoded time-series signal, a time-frequency graph encoder that encodes a time-frequency graph to obtain an encoded time-frequency graph, a signal feature extraction layer and a time-frequency graph feature extraction layer that respectively extract features from the encoded time-series signal and the encoded time-frequency graph to obtain time-series signal features and time-frequency graph features, a fusion layer that fuses the time-series signal features and the time-frequency graph features to obtain fused data, a decoder that decodes the fused data to obtain decoded data, a generator that takes the decoded data with added noise as input, and a discriminator that distinguishes true from false on the generated data output by the generator; wherein the self-supervised prediction subnet is formed by a Bi-LSTM network.

6. The abnormal warning method of the vibrating screen according to claim 1, wherein In step S2, the training includes: S201 performs Fourier transform on the time-aligned vibration signals in the original vibration data set to obtain their corresponding frequency components and amplitude values, selects the 5 most frequently occurring frequency components among the 10 frequency components with the largest amplitude values, converts them into a time scale sequence based on the relationship between period and frequency, segments the time-aligned vibration signals using each time scale value in the time scale sequence as a segmentation window size to obtain a number of non-overlapping segmented signal segments, and splices the segmented signal segments up and down into a two-dimensional matrix according to the time relationship to obtain labeled reconstructed data; performs the same processing on the vibration signals in the generated vibration data set to obtain unlabeled reconstructed data; S202 performs a first-stage training on the vibration screen operation state recognition model using the labeled reconstructed data and its corresponding time-frequency graph to obtain a partially optimized vibration screen operation state recognition model; S203 performs a second-stage training on the partially optimized vibration screen operation state recognition model using the labeled reconstructed data and its corresponding time-frequency diagram and the unlabeled reconstructed data and its corresponding time-frequency diagram. During the training, a metric learning method and its corresponding loss function are used to shorten the distribution distance of the labeled reconstructed data and the unlabeled reconstructed data in the feature space until the identification network cannot identify whether the feature comes from the labeled reconstructed data or the unlabeled reconstructed data, thereby obtaining the trained vibration screen operation state recognition model.

7. The abnormal warning method of the vibrating screen according to claim 6, characterized in that, The vibration screen operation state recognition model includes a feature extraction model and a classifier. Among them, the feature extraction model is composed of multiple feature blocks connected in a residual manner. Each feature block consists of two multi-core CNN networks and one fusion network. Among them, one multi-core CNN network is used to extract the reconstructed data features, and one multi-core CNN network is used to extract the time-frequency diagram features; the fusion network is used to fuse the extracted reconstructed data features and time-frequency diagram features.

8. The abnormal warning method of the vibrating screen according to claim 1, characterized in that, Among them, the state prediction model is constructed based on the LSTM gate mechanism and the Transformer with a multi-head attention mechanism; and / or, the encoding correction mechanism includes: inputting the vibration data obtained in real time into the trained vibration screen operation state recognition model, encoding the output operation state recognition result to obtain a real-time recognition code, and judging and optimizing the jump code in the predicted operation state code according to the real-time recognition code and its historical data, that is, the historical recognition code.

9. The abnormal warning method of the vibrating screen according to claim 8, characterized in that, The state prediction model includes: 3 encoders with a multi-head self-attention mechanism with a gate mechanism. Each encoder includes a multi-head attention layer, a first normalization layer, a feed-forward network layer, a second normalization layer, and a first fully connected layer connected through a residual connection. Among them, the multi-head attention layer includes 4 different self-attention heads; a Sigmoid activation layer connected to the multi-head attention layer and the first normalization layer; a third normalization layer connected to the first fully connected layer; a second fully connected layer connected to the third normalization layer; several feature extraction blocks connected to the second fully connected layer; each feature extraction block contains a convolutional layer, a fully connected layer, and a Relu activation layer.

10. A vibrating screen abnormal warning system for implementing the vibrating screen abnormal warning method according to any one of claims 1-9, characterized in that, It includes a vibration screen adaptive data augmentation module and a vibration screen anomaly warning module. Among them, the vibration screen adaptive data augmentation module is used for the original vibration data set based on the vibration data at the key positions of the measured vibration screen and its corresponding operation state label, and obtains a generated vibration data set with diversity and balanced distribution through a self-supervised dual-coding adversarial generation network and enhancement transformation processing. The vibration screen anomaly warning module is used to train the vibration screen operation state recognition model through the original vibration data set and the generated vibration data set. The trained vibration screen operation state recognition model is used to recognize the operation state of the vibration screen. And according to the vibration data at the key positions of the measured vibration screen and its historical data, the predicted vibration data of the vibration screen is obtained through the state prediction model. The predicted vibration data is input into the trained vibration screen operation state recognition model to obtain the predicted operation state, which is digitally encoded to obtain the predicted operation state code. The predicted operation state code or the predicted operation state code corrected by the encoding correction mechanism is combined with the anomaly score obtained by the similarity metric between the predicted vibration data and the measured vibration data to jointly determine the abnormal operation state of the vibration screen.

Images