A deep cone thickener state classification method in filling scenarios

Abstract

The present invention discloses a method for classifying the state of a deep-cone thickener in a filling scenario, comprising: obtaining historical operating parameters of a thickener system under different operating states; wherein the historical operating parameters include operating parameters without operating state labels and operating parameters with operating state labels; constructing a thickener representation learning network based on contrastive learning, and training it using the operating parameters without operating state labels; constructing a thickener state classifier, inputting the operating parameters with operating state labels into the trained thickener representation learning network to obtain corresponding feature representations, and training the thickener state classifier using the obtained feature representations; obtaining the operating parameters of the thickener to be classified, and classifying the current operating state of the thickener to be classified. The present invention utilizes contrastive learning to train the thickener classifier and obtain accurate classification results using only a small amount of labeled data, enabling detection and alarm of the thickener's real-time operating state.

Description

Technical Field

[0001] The present invention relates to the technical field of mine tailings disposal, and in particular to a deep cone thickener state classification method in a filling scenario. Background Art

[0002] The deep cone thickener is a solid-liquid separation device widely used in complex process industrial scenarios such as mining and metallurgy. It uses the concentration principle, that is, the gravity of the solid particles themselves, to separate the slurry into two parts: a precipitate with a higher concentration and a clarified liquid. It can be used in production processes such as dehydration, drug removal, and concentration.

[0003] Thickener operation is cumbersome. In most production scenarios, engineers must manually adjust the thickener. For example, in the common tailings filling process in the mining industry, the thickener is required to provide a concentrated slurry of acceptable concentration for subsequent processes. However, the required slurry concentration varies depending on the production mode, such as tailings filling or discharge to a tailings pond. Consequently, the thickener typically operates under different operating conditions. Engineers often need to determine the current thickener operating state based on the thickener's operating parameters and manually adjust valves and certain process settings to adjust the thickener's operating state to meet the required slurry concentration. However, the complex operation of the thickener makes it difficult to easily determine its status, which complicates equipment adjustment and control. Furthermore, the thickener may experience sensor failure or incorrect operating conditions due to operator error. These abnormalities can go undetected due to the difficulty in determining the thickener's status. Abnormal thickener operation can lead to reduced slurry quality, resulting in financial losses for the company and even production accidents such as rake pressure. Therefore, it is of great significance to study a high-accuracy thickener status classification method.

[0004] To address the problem of thickener status classification, existing approaches typically rely on developing precise mechanism analysis models for the thickener. This first requires exploring the internal operating mechanisms of the thickener system, identifying factors and parameters related to the thickener's operating status, and analyzing their inherent relationships. However, due to the complex structure of thickeners, which varies greatly from one type to another, establishing a precise mechanism analysis model is difficult without prior knowledge of the thickener itself. Therefore, a new approach is needed to address the problem of thickener status classification during operation. Summary of the Invention

[0005] The present invention provides a method for classifying the states of a deep cone thickener in a filling scenario, so as to solve the technical problem that the prior art cannot effectively classify the states of the thickener during use.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In one aspect, the present invention provides a method for classifying a deep cone thickener state in a filling scenario, comprising:

[0008] Acquire historical operating parameters of the thickener system under different operating states; wherein the historical operating parameters include: operating parameters without operating state markings and operating parameters with operating state markings;

[0009] Based on contrastive learning, a thickener representation learning network is constructed, and the thickener representation learning network is trained using working parameters without working state labels to obtain a trained thickener representation learning network;

[0010] Based on a neural network, a thickener state classifier is constructed, and the working parameters with working state labels are input into a trained thickener representation learning network to obtain corresponding feature representations. The obtained feature representations are used to train the thickener state classifier to obtain a trained thickener state classifier;

[0011] The working parameters of the thickener to be classified are obtained, and the feature representation of the working parameters of the thickener to be classified is extracted using the trained thickener representation learning network. The extracted feature representation is then sent to the trained thickener state classifier to classify the current working state of the thickener to obtain the classification result.

[0012] Furthermore, the obtaining of historical operating parameters of the thickener system under different operating conditions includes:

[0013] Collect the original working parameters of the thickener system in different working states during its historical operation; wherein the working states include: filling state, tail state and static state; the working parameters include: inlet and outlet flow rate, inlet and outlet concentration and mud layer pressure;

[0014] The collected original operating parameters are converted into uniform time series by linear interpolation to obtain historical operating parameters of the thickener system under different operating conditions.

[0015] Furthermore, the thickener representation learning network includes: a sequence encoder module, a trend feature extraction module and a period feature extraction module; wherein,

[0016] The sequence encoder module is used to encode the input sequence data to obtain a high-dimensional feature representation that covers more semantic information;

[0017] The trend feature extraction module is used to extract the trend features represented by the high-dimensional features output by the sequence encoder module, that is, the long-term change features of the sequence data;

[0018] The periodic feature extraction module is used to extract the periodic features represented by the high-dimensional features output by the sequence encoder module, that is, the variation features of the periodic fluctuations of the sequence data.

[0019] Furthermore, the sequence encoder module is a temporal convolutional network; the input sequence data is first expanded by a linear layer, and then encoded into a high-dimensional feature representation by a multi-layer temporal convolutional network.

[0020] Furthermore, the trend feature extraction module uses a hybrid autoregressive filter with different window sizes to extract trend features; each autoregressive filter is a causal convolutional network, which performs autoregressive filtering through one-dimensional causal convolution, and the convolution kernel size is the autoregressive window size. By performing mean pooling on the results of all autoregressive filters, the trend feature extraction results of the sequence data are obtained.

[0021] Furthermore, the periodic feature extraction module uses fast Fourier transform to convert the sequence from the time domain to the frequency domain, then extracts the frequency domain features through a linear transformation layer, and finally uses inverse Fourier transform to convert the frequency domain features back to the time domain.

[0022] Furthermore, the thickener representation learning network is trained, including:

[0023] The thickener representation learning network is trained using the MOCO contrastive learning architecture. During training, the original training data is first subjected to two different data augmentations to construct two sets of data with different views of the original training data. The two sets of data are then passed through the sequence encoder module to obtain corresponding high-dimensional feature representations, and the obtained high-dimensional feature representations are used to train the trend feature extraction module and the period feature extraction module respectively.

[0024] Furthermore, the data enhancement methods include data offset, data scaling and adding Gaussian noise.

[0025] Furthermore, the loss function of the trend feature extraction module is as follows:

[0026]

[0027] Among them, Loss t represents the loss function of the trend feature extraction module, N represents the total number of samples, K represents the length of the negative sample queue, and q i is the trend benchmark representation of the i-th sample, k i is the positive sample representation of the i-th sample, k i represents the jth sample in the negative sample queue, τ is a hyperparameter;

[0028] The periodic feature extraction module constructs loss functions for amplitude and phase respectively as follows:

[0029]

[0030]

[0031] Among them, Loss amp Represents the loss function corresponding to the amplitude, Loss phase Represents the loss function corresponding to the phase, N represents the total number of samples, and F represents the number of frequencies obtained by Fourier decomposition. is the result of Fourier decomposition of the jth sample at the i-th frequency, is the result of the Fourier decomposition of the positive sample of the j-th sample at the i-th frequency, is the amplitude of the jth sample at the i-th frequency after Fourier decomposition, is the amplitude of the j-th sample at the i-th frequency after Fourier decomposition of the positive sample, is the amplitude of the kth sample in the negative sample queue at the i-th frequency after Fourier decomposition, is the phase of the j-th sample at the i-th frequency after Fourier decomposition, is the phase of the j-th sample at the i-th frequency after Fourier decomposition of the positive sample, is the phase of the kth sample in the negative sample queue at the i-th frequency after Fourier decomposition, represents the dot product operation, which is used to measure the similarity of two operands;

[0032] The final densifier represents the overall loss function of the learning network as follows:

[0033]

[0034] Among them, Loss represents the overall loss function of the learning network, and α is a hyperparameter.

[0035] Furthermore, the thickener state classifier outputs a multidimensional vector through a fully connected layer and a Relu activation layer for the feature representation of the working parameters output by the thickener representation learning network, representing the number of pre-set thickener state categories; finally, the result is input into the softmax layer for normalization to obtain the probability of each thickener state category, and the category with the largest value is the final thickener state judgment result.

[0036] On the other hand, the present invention further provides an electronic device, comprising a processor and a memory; wherein the memory stores at least one instruction, and the instruction is loaded and executed by the processor to implement the above method.

[0037] In yet another aspect, the present invention further provides a computer-readable storage medium, wherein the storage medium stores at least one instruction, and the instruction is loaded and executed by a processor to implement the above method.

[0038] The beneficial effects brought about by the technical solution provided by the present invention include at least:

[0039] 1. To address the problems that end-to-end time series classification methods require a large amount of labeled data and thickener status labeling requires a lot of manpower, the present invention proposes a thickener status classification method based on representation learning. First, a comparative learning architecture is used to unsupervisedly learn good feature representations for different categories of thickener data. In this way, the thickener status classifier trained with a small amount of labeled data can achieve good classification results.

[0040] 2. To address the characteristic of thickeners having different means and periods under different operating conditions, this invention divides the representation learning network into two parts: a trend feature extraction module and a period feature extraction module. It uses a hybrid autoregressive filtering algorithm to extract trend features and a learnable Fourier layer to model periodic features. This method effectively improves feature extraction and enhances the interpretability of the network. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0042] Figure 1 is a flow chart of a deep cone thickener state classification method in a filling scenario provided by an embodiment of the present invention;

[0043] Figure 2 1 is a framework diagram of a trend feature extraction module provided by an embodiment of the present invention;

[0044] Figure 3 This is a framework diagram of a periodic feature extraction module provided by an embodiment of the present invention;

[0045] Figure 4 This is a framework diagram of a deep cone thickener state classification method in a filling scenario provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0046] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0047] First embodiment

[0048] This embodiment provides a deep cone thickener state classification method in a filling scenario, which is a data-driven thickener state identification method. This method can accurately classify the current working state of the thickener using the thickener working parameters. The execution process of this method is as follows: Figure 1 As shown, the following steps are included:

[0049] S1, obtaining historical operating parameters of the thickener system under different operating conditions;

[0050] The working status of the thickener system includes: filling state, tail state and static state; historical working parameters include: a large number of working parameters without working status mark and a small number of working parameters with working status mark; working parameters include: inlet and outlet flow rate, inlet and outlet concentration and mud layer pressure;

[0051] Specifically, in this embodiment, the implementation process of the above S1 is as follows:

[0052] S11, collecting the original working parameters of the thickener system under different working states during the historical working process;

[0053] Specifically, this embodiment uses the OPC technology to read data from the industrial DCS system, obtain historical operating parameters of the thickener system monitored by various sensors, and store the data in a local database.

[0054] S12, converting the collected original operating parameters into a uniform time series by linear interpolation to obtain historical operating parameters of the thickener system under different operating conditions;

[0055] Specifically, this example exports key thickener production data from the database. Data for five points, including inlet and outlet concentration, inlet and outlet flow rates, and mud layer pressure, are entered into a CSV file. Because the time intervals of the monitored production data are uneven, the raw data represents a non-uniform time series. Therefore, linear interpolation is first performed on the data at all points to convert it into a uniform time series.

[0056] S2, based on contrastive learning, construct a thickener representation learning network, and use the working parameters without working state annotations to train the thickener representation learning network to obtain a trained thickener representation learning network;

[0057] Specifically, in this embodiment, the thickener representation learning network based on contrastive learning is used to construct a feature representation of thickener working parameters that is easier to classify based on contrastive learning. The network includes: a sequence encoder module, a trend feature extraction module, and a period feature extraction module; wherein,

[0058] The sequence encoder module is used to encode the input sequence data to expand the dimension of the original data, thereby obtaining a high-dimensional feature representation of the input sequence that covers more semantic information;

[0059] Specifically, this example uses a temporal convolutional network (TCN) to encode historical data sequences. The original data first undergoes a linear layer for dimension expansion, and then passes through an N-layer (6 in this example) TCN network to encode the original sequence X into h(N) of length L (32 in this example). The network model is expressed as:

[0060]

[0061]

[0062] h(i)=σ(f(h(i-1),θ i ))

[0063] in, represents the sequence after linear layer expansion, X represents the original sequence, W represents the linear layer weight, b represents the linear layer bias, σ represents the activation function, f represents the time convolution network, θ i represents the parameters of the i-th layer dilated causal convolutional network, and h(i) represents the output result of the i-th layer dilated causal convolutional network.

[0064] The architecture of the trend feature extraction module is as follows: Figure 2 As shown, it is used to extract the trend characteristics of the high-dimensional feature representation output by the sequence encoder module, that is, the long-term change characteristics of the sequence data;

[0065] Specifically, in this embodiment, the trend feature extraction module uses a hybrid autoregressive filter with different window sizes to extract trend features. Through training, the model can adaptively learn the appropriate window size. The module uses a total of M (set to 6 in this embodiment) autoregressive filters, which perform autoregressive filtering through one-dimensional causal convolution. The convolution kernel size is the autoregressive window size. After comparative experiments, this embodiment sets 6 convolution kernel sizes of 1, 2, 4, 8, 16, and 32 respectively. After the data is filtered by each autoregressive filter, the result is average pooled to obtain the final extracted trend feature. The relationship between the input and output y of this module is as follows:

[0066]

[0067] Among them, x is the output result of the sequence encoder module, C i is the i-th autoregressive filter.

[0068] The architecture of the period feature extraction module is as follows: Figure 3As shown, it is used to extract the periodic features represented by the high-dimensional features output by the sequence encoder module, that is, the changing features of the periodic fluctuations of the sequence data;

[0069] Specifically, in this embodiment, the periodic feature extraction module uses fast Fourier transform to convert the sequence from the time domain to the frequency domain, then extracts the frequency domain features through the linear transformation layer, and finally uses inverse Fourier transform to convert the frequency domain features back to the time domain, thereby obtaining the periodic fluctuation variation characteristics of the sequence. It should be noted that for a sequence of length S, the Fourier transform can be decomposed into According to Euler's formula, the sine and cosine waves are converted into the form of complex exponential functions. This is the Fourier transform in the complex frequency domain. The Fourier transform layer uses fast Fourier transform to convert the input data from the time domain to the frequency domain. The transformation result is complex numbers, each of which reflects the phase and amplitude information of the wave at a certain frequency when the input data is decomposed. If the complex number form is a+bi, the corresponding amplitude and phase are as follows:

[0070]

[0071]

[0072] The output of the Fourier transform layer passes through a linear transformation layer to extract frequency domain features. For each frequency result in the decomposed frequency domain, a parameterized learnable linear transformation matrix is set to learn the feature information at that frequency. The output of the linear transformation layer is a feature representation of the sequence from a frequency domain perspective. In order to combine it with the trend features extracted in the time domain for use in downstream tasks, the frequency domain feature representation is converted to the time domain through an inverse Fourier transform layer. The relationship between the input and output V of this module is as follows:

[0073] V=F -1 (AF(x)+B)

[0074] Among them, x is the output result of the encoder module, F is the Fourier transform, A and B are the linear transformation layer parameters, and F -1 is the inverse Fourier transform.

[0075] The training process of the thickener representation learning network is as follows: the thickener representation learning network is trained using the MOCO contrastive learning architecture; during training, the original training data first undergoes two different data augmentations to construct two sets of data with different views of the original training data. The two sets of data are then passed through the sequence encoder module to obtain corresponding high-dimensional feature representations, and the obtained high-dimensional feature representations are used to train the trend feature extraction module and the period feature extraction module respectively.

[0076] The high-dimensional representation enters the trend feature extraction module and the period feature extraction module respectively to obtain the corresponding feature representation. Different views of the same data are represented as positive sample pairs, and different data are represented as negative sample pairs. The positive sample pairs and negative sample pairs are used as sub-elements in the contrast loss function to calculate the contrast loss function. The contrast loss function is used to improve the similarity of the feature sequence corresponding to the same instance, improve the similarity of the positive sample pairs corresponding to the same instance, and reduce the similarity between the positive samples and the negative samples.

[0077] Among them, data enhancement methods include data offset, data scaling and adding Gaussian noise.

[0078] Data offset refers to adding or subtracting a random scalar value from the time series. The random scalar value is sampled from a Gaussian distribution, that is, ε~N(0,0.5). Where N(0,0.5) is a Gaussian distribution with a mean of 0 and a variance of 0.5, ε is a random scalar value, and x t is the sequence value with timestamp t, The result of data enhancement.

[0079] Data scaling scales the time series by a single random scalar value sampled from a Gaussian distribution, i.e., ε~N(0,0.5). Where N(0,0.5) is a Gaussian distribution with a mean of 0 and a variance of 0.5, ε is a random scalar value, and x t is the sequence value with timestamp t, The result of data enhancement.

[0080] Gaussian noise addition means adding Gaussian noise to each time step. t ~N(0,0.5), Among them, N(0,0.5) is a Gaussian distribution with a mean of 0 and a variance of 0.5, ε t is the Gaussian noise generated for the sequence value with timestamp t, x t is the sequence value with timestamp t, The result of data enhancement.

[0081] The loss function of the trend feature extraction module is as follows:

[0082]

[0083] Among them, Loss t represents the loss function of the trend feature extraction module, N represents the total number of samples, K represents the length of the negative sample queue, and q i is the trend benchmark representation of the i-th sample, k i is the positive sample representation of the i-th sample, k irepresents the jth sample in the negative sample queue, τ is a hyperparameter;

[0084] The periodic feature extraction module constructs loss functions for amplitude and phase respectively, as follows:

[0085]

[0086]

[0087] Among them, Loss amp Represents the loss function corresponding to the amplitude, Loss phase Represents the loss function corresponding to the phase, N represents the total number of samples, and F represents the number of frequencies obtained by Fourier decomposition. is the result of Fourier decomposition of the jth sample at the i-th frequency, is the result of Fourier decomposition of the positive sample of the j-th sample at the i-th frequency, is the amplitude of the jth sample at the i-th frequency after Fourier decomposition, is the amplitude of the j-th sample at the i-th frequency after Fourier decomposition of the positive sample, is the amplitude of the kth sample in the negative sample queue at the i-th frequency after Fourier decomposition, is the phase of the j-th sample at the i-th frequency after Fourier decomposition, is the phase of the j-th sample at the i-th frequency after Fourier decomposition of the positive sample, is the phase of the kth sample in the negative sample queue at the i-th frequency after Fourier decomposition, represents the dot product operation, which is used to measure the similarity of two operands;

[0088] The final densifier represents the overall loss function of the learning network as follows

[0089]

[0090] Loss represents the overall loss function of the dense learning network. α is a hyperparameter used to control the weight ratio of trend contrast learning loss and season contrast learning loss. Because the trend representation loss and the season representation loss may not be on the same order of magnitude, after comparative experiments, the α value is finally set to 0.01.

[0091] S3, based on the neural network, construct a thickener state classifier, input the working parameters with working state labels into the trained thickener representation learning network to obtain the corresponding feature representation, and use the obtained feature representation to train the thickener state classifier to obtain a trained thickener state classifier;

[0092] Specifically, in this embodiment, the thickener state classifier outputs a multi-dimensional vector through a fully connected layer and a Relu activation layer for the feature representation output by the thickener representation learning network, representing the number of pre-set thickener state categories; finally, the result is input into the softmax layer for normalization to obtain the probability of each thickener state category, and the category with the largest value is the final thickener state judgment result.

[0093] S4, obtaining the working parameters of the thickener to be classified, using the trained thickener representation learning network to extract the feature representation of the working parameters of the thickener to be classified, and then sending the extracted feature representation to the trained thickener state classifier to classify the current working state of the thickener to be classified to obtain a classification result.

[0094] Specifically, the implementation process of the above S4 is: obtaining the working parameters of the thickener to be classified in the most recent time period, first inputting the obtained working parameters into the trained thickener representation learning network to extract the feature representation corresponding to the working parameters, and then sending the extracted feature representation into the trained thickener state classifier, and using the thickener state classifier to output the final classification result of the current working state of the classified thickener.

[0095] In summary, this embodiment provides a deep cone thickener state classification method in a filling scenario, and its system framework is as follows: Figure 4 As shown, this deep-cone thickener state classification method can learn the characteristic differences between thickener operating data under different states and train a model. This model can then be used to classify thickener states given new data. Compared to existing methods, this method reduces reliance on expert experience and, with sufficient data, can be applied to thickener production processes in various industrial scenarios.

[0096] Second embodiment

[0097] This embodiment provides an electronic device, which includes a processor and a memory; wherein the memory stores at least one instruction, and the instruction is loaded and executed by the processor to implement the method of the first embodiment.

[0098] The electronic device may have relatively large differences due to different configurations or performances, and may include one or more processors (central processing units, CPU) and one or more memories, wherein the memory stores at least one instruction, which is loaded by the processor to execute the above method.

[0099] Third embodiment

[0100] This embodiment provides a computer-readable storage medium storing at least one instruction, which is loaded and executed by a processor to implement the method of the first embodiment described above. The computer-readable storage medium may be a ROM, random access memory, CD-ROM, magnetic tape, floppy disk, or optical data storage device. The instructions stored therein can be loaded by a processor in a terminal to execute the method described above.

[0101] Furthermore, it should be noted that the present invention may be provided as a method, apparatus, or computer program product. Thus, embodiments of the present invention may take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of the present invention may take the form of a computer program product embodied on one or more computer-usable storage media containing computer-usable program code.

[0102] The embodiments of the present invention are described with reference to the flowcharts and / or block diagrams of the methods, terminal devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, an embedded processor, or other programmable data processing terminal device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal device generate instructions for implementing the process in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0103] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing terminal device to operate in a specific manner, so that the instructions stored in the computer readable memory produce a manufactured product including an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device, so that a series of operation steps are executed on the computer or other programmable terminal device to produce a computer-implemented process, thereby providing instructions for implementing the process in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0104] It should also be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or terminal device comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or terminal device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or terminal device comprising the element.

[0105] Finally, it should be noted that the above is a preferred embodiment of the present invention. It should be noted that although the preferred embodiment of the present invention has been described, it is clear that those skilled in the art, once they understand the basic inventive concept of the present invention, can make various improvements and modifications without departing from the principles of the present invention. Such improvements and modifications should also be considered as within the scope of protection of the present invention. Therefore, the appended claims are intended to be interpreted as including the preferred embodiment and all changes and modifications that fall within the scope of the embodiments of the present invention.

Claims

1. A deep cone thickener state classification method in a filling scenario, characterized in that: include: Acquire historical operating parameters of the thickener system under different operating states; wherein the historical operating parameters include: operating parameters without operating state markings and operating parameters with operating state markings; Based on contrastive learning, a thickener representation learning network is constructed, and the thickener representation learning network is trained using working parameters without working state labels to obtain a trained thickener representation learning network; Based on a neural network, a thickener state classifier is constructed, and the working parameters with working state labels are input into a trained thickener representation learning network to obtain corresponding feature representations. The obtained feature representations are used to train the thickener state classifier to obtain a trained thickener state classifier; Obtaining the working parameters of the thickener to be classified, extracting feature representations of the working parameters of the thickener to be classified using the trained thickener representation learning network, and then feeding the extracted feature representations into the trained thickener state classifier to classify the current working state of the thickener to obtain a classification result; The obtaining of historical operating parameters of the thickener system under different operating conditions includes: Collect the original working parameters of the thickener system in different working states during its historical operation; wherein the working states include: filling state, tail state and static state; the working parameters include: inlet and outlet flow rate, inlet and outlet concentration and mud layer pressure; The collected original operating parameters are converted into a uniform time series by using a linear interpolation method to obtain historical operating parameters of the thickener system under different operating conditions; The thickener representation learning network includes: a sequence encoder module, a trend feature extraction module and a period feature extraction module; wherein, The sequence encoder module is used to encode the input sequence data to obtain a high-dimensional feature representation that covers more semantic information; The trend feature extraction module is used to extract the trend features represented by the high-dimensional features output by the sequence encoder module, that is, the long-term change features of the sequence data; The periodic feature extraction module is used to extract the periodic features represented by the high-dimensional features output by the sequence encoder module, that is, the variation features of the periodic fluctuation of the sequence data; The sequence encoder module is a temporal convolutional network. The input sequence data is first expanded by a linear layer, and then encoded into a high-dimensional feature representation by a multi-layer temporal convolutional network. The trend feature extraction module uses a hybrid autoregressive filter with different window sizes to extract trend features. Each autoregressive filter is a causal convolutional network that performs autoregressive filtering through one-dimensional causal convolution. The convolution kernel size is the autoregressive window size. The trend feature extraction result of the sequence data is obtained by mean pooling the results of all autoregressive filters. The periodic feature extraction module uses fast Fourier transform to convert the sequence from the time domain to the frequency domain, then extracts the frequency domain features through a linear transformation layer, and finally uses inverse Fourier transform to convert the frequency domain features back to the time domain.

2. The deep cone thickener state classification method in a filling scenario according to claim 1, characterized in that: Training the thickener representation learning network includes: The thickener representation learning network is trained using the MOCO contrastive learning architecture. During training, the original training data is first subjected to two different data augmentations to construct two sets of data with different views of the original training data. The two sets of data are then passed through the sequence encoder module to obtain corresponding high-dimensional feature representations, and the obtained high-dimensional feature representations are used to train the trend feature extraction module and the period feature extraction module respectively.

3. The deep cone thickener state classification method in a filling scenario according to claim 2, characterized in that: The data enhancement methods include data offset, data scaling and adding Gaussian noise.

4. The deep cone thickener state classification method in a filling scenario according to claim 2, characterized in that: The loss function of the trend feature extraction module is as follows: Among them, Loss t represents the loss function of the trend feature extraction module, N represents the total number of samples, K represents the length of the negative sample queue, and q i is the trend benchmark representation of the i-th sample, k i is the positive sample representation of the i-th sample, k j represents the jth sample in the negative sample queue, τ is a hyperparameter; The periodic feature extraction module constructs loss functions for amplitude and phase respectively as follows: Among them, Loss amp Represents the loss function corresponding to the amplitude, Loss phase Represents the loss function corresponding to the phase, N represents the total number of samples, and F represents the number of frequencies obtained by Fourier decomposition. is the result of Fourier decomposition of the nth sample at the fth frequency, is the result of the Fourier decomposition of the positive sample of the nth sample at the fth frequency, is the amplitude of the nth sample at the fth frequency after Fourier decomposition, is the amplitude of the positive sample of the nth sample at the fth frequency after Fourier decomposition, is the amplitude of the kth sample in the negative sample queue at the fth frequency after Fourier decomposition, is the phase of the nth sample at the fth frequency after Fourier decomposition, is the phase of the positive sample of the nth sample at the fth frequency after Fourier decomposition, is the phase of the kth sample in the negative sample queue at the fth frequency after Fourier decomposition, represents the dot product operation, which is used to measure the similarity of two operands; The final densifier represents the overall loss function of the learning network as follows: Among them, Loss represents the overall loss function of the learning network, and α is a hyperparameter.

5. The deep cone thickener state classification method in a filling scenario according to claim 1, characterized in that: The thickener state classifier passes the feature representation of the thickener working parameters output by the thickener representation learning network through a fully connected layer and a Relu activation layer to output a multi-dimensional vector, representing the number of pre-set thickener state categories; finally, the result is input into the softmax layer for normalization to obtain the probability of each thickener state category, and the category with the largest value is the final thickener state judgment result.

Images