Methods, apparatus, equipment and media for identifying cross-tank abnormalities in flotation processes
By acquiring labeled samples during the flotation process, encoding foam images and process numerical vectors, and fusing two-level attention residuals, a benchmark cell model is trained and adapted to the target cell model. This solves the problems of differences in abnormal appearance features of different flotation cells and limited labeled samples, and achieves accurate identification of settling and running conditions.
Patent Information
- Application Number
- CN202611622486.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-20
- Publication Date
- 2026-11-17
AI Technical Summary
Given the differences in abnormal appearance characteristics among different flotation cells and the limited number of target cell labeled samples, identifying settling and running conditions has become an urgent problem to be solved.
By acquiring labeled samples of the reference slot and the target slot, the foam image and process numerical vector are encoded and two-level attention residual fusion is performed to train the reference slot model. The parameters of the reference slot model are used to initialize the target slot model. The target slot model is updated by combining classification loss, alignment loss between similar prototypes and intraclass compactness loss of the target slot, and outputting the working condition identification results of normal, sinking, or running slot.
When there are differences in the abnormal appearance characteristics of different flotation cells and the target cell has a limited number of labeled samples, it can accurately identify settling and running conditions, reduce the impact of inter-cell characteristic differences, and improve identification accuracy.
Smart Images

Figure CN123116183A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the fields of image data processing and computer vision technology, and in particular to a method, apparatus, equipment and medium for identifying cross-tank abnormal conditions in the flotation process. Background Technology
[0002] Flotation is a crucial step in mineral separation, and the state of the froth in the flotation cell is closely related to operating parameters such as liquid level, aeration rate, and valve opening. Abnormal conditions such as settling tank overflow and flotation cell runaway can affect the continuous operation of the flotation process. Therefore, it is necessary to identify abnormal conditions promptly based on the apparent state of the flotation froth and process data.
[0003] Currently, data is typically collected from individual flotation cells, either as images or process data, to extract features such as color, texture, morphology, and flow rate. Alternatively, classification models are used to identify normal, settling, and overflow conditions. However, the same abnormal condition can vary in foam residue morphology, surface exposure range, overflow location, and intensity depending on factors such as ore properties, slurry concentration and viscosity, reagent formulation, liquid level, and aeration rate. Furthermore, in the initial stages of anomalies, and during adjustment and recovery phases, only some abnormal features may appear, leading to overlap between normal, settling, and overflow conditions. Therefore, relying solely on a single foam image or fixed visual rules is insufficient to reliably reflect the current operating condition.
[0004] To simultaneously utilize both the apparent state of the foam and its internal operating state, existing methods can directly concatenate image features with process numerical features, or perform classification after a single attention weighting. However, direct concatenation makes it difficult to separately adjust the contributions of each feature dimension within the same input path and between different inputs; the model may primarily rely on partial image cues or local process numerical values. Furthermore, if only the attention weighting result is used to replace the original features, biased weight estimation may weaken some effective information.
[0005] However, the pulp properties, reagent conditions, equipment status, and image acquisition conditions vary among different flotation cells, and similar abnormal operating conditions may exhibit different foam appearance characteristics in different cells. When a model trained based on a single flotation cell sample is directly applied to other cells, the identification results are easily affected by differences in data distribution between cells. If a large number of abnormal annotation samples are re-collected for each target cell, it is constrained by the limited number of abnormal operating condition samples and the high annotation cost. Therefore, how to identify settling and running cell conditions when the abnormal appearance characteristics of different flotation cells differ and the number of target cell annotation samples is limited becomes an urgent problem to be solved. Summary of the Invention
[0006] The purpose of this application is to provide a method, apparatus, equipment and medium for identifying abnormal operating conditions in the flotation process, which aims to solve the technical problem of how to identify settling and running conditions when there are differences in the abnormal appearance characteristics of different flotation cells and the target cell labeling samples are limited.
[0007] To achieve the above objectives, this application proposes a method for identifying abnormal operating conditions across tanks in the flotation process, the method comprising: Acquire labeled samples of the reference slot and the target slot, wherein the labeled samples include time-corresponding foam images, process numerical vectors and operating condition labels; The foam image and the process numerical vector are encoded and fused using two levels of attention residuals to obtain fused features. A baseline slot model is then trained based on the fused features and the working condition label. The benchmark slot category prototype is calculated based on the similar fusion features output by the benchmark slot model, and the target slot model is initialized with the parameters of the benchmark slot model. The target slot class prototype for each training batch is calculated based on the fusion features output by the target slot model. The baseline slot class prototype remains unchanged. The target slot model is then updated by combining the classification loss, the alignment loss between similar prototypes, and the compactness loss within the target slot class. Input the foam image and process numerical vector corresponding to the current time of the target tank into the updated target tank model, and output the working condition identification result of normal, sinking or running out of the tank.
[0008] Furthermore, to achieve the above objectives, this application also proposes an abnormal operating condition cross-tank identification device for the flotation process, the device comprising: The sample acquisition module is used to acquire labeled samples of the reference slot and the target slot. The labeled samples include time-corresponding foam images, process numerical vectors, and operating condition labels. The benchmark training module is used to encode the foam image and the process numerical vector and perform two-level attention residual fusion to obtain fusion features, and train the benchmark slot model based on the fusion features and the working condition label. The adaptation preparation module is used to calculate the benchmark slot category prototype based on the same type of fusion features output by the benchmark slot model, and initialize the target slot model with the parameters of the benchmark slot model. The cross-slot adaptation module is used to calculate the target slot category prototype for each training batch based on the fusion features output by the target slot model, keep the baseline slot category prototype unchanged, and update the target slot model by combining classification loss, alignment loss between prototypes of the same class and compactness loss within the target slot class. The working condition identification module is used to input the foam image and process numerical vector corresponding to the current time of the target tank into the updated target tank model, and output the working condition identification results of normal, sinking or running tank.
[0009] Furthermore, to achieve the above objectives, this application also proposes an abnormal operating condition cross-cell identification device for the flotation process, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the abnormal operating condition cross-cell identification method for the flotation process as described above.
[0010] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the abnormal operating condition cross-slot identification method for the flotation process described above.
[0011] One or more technical solutions proposed in this application have at least the following technical effects: Labeled samples from the baseline and target slots are acquired. Time-corresponding foam images, process numerical vectors, and operating condition labels are used to establish the correspondence between foam appearance information, process operation information, and operating condition categories, providing a data foundation for model training. The foam images and process numerical vectors are encoded separately and fused using a two-level attention residual fusion method to form fusion features. This allows operating condition-related information from both types of data to be expressed in the same feature space. The baseline slot model is then trained using the operating condition labels to obtain the classification criteria for normal, sinking, and running conditions in the baseline slot. Based on the fusion features output by the baseline slot model, a baseline slot category prototype is calculated to represent the common characteristics of various operating conditions in the baseline slot. The parameters of the baseline slot model are then used to initialize the target slot model, enabling the target slot model to continue... Building upon existing feature extraction and classification capabilities, the system calculates target cell category prototypes for each training batch based on the fusion features output by the target cell model. While maintaining the baseline cell category prototypes unchanged, it uses classification loss to constrain the classification of operating conditions, utilizes alignment loss between prototypes of the same type to reduce the feature differences between cells under the same operating conditions, and uses intra-class compactness loss to constrain the feature distribution of samples of the same type. Based on this, the target cell model is updated to adapt the model to the data features of the target cell. The foam image and process numerical vector corresponding to the current time of the target cell are input into the updated target cell model, and the system outputs the operating condition identification results of normal, settling, or running cell based on the adapted classification relationship. This enables the identification of settling and running cell operating conditions when there are differences in the abnormal appearance features of different flotation cells and the target cell labeled samples are limited. Attached Figure Description
[0012] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.
[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0014] Figure 1 This is a flowchart illustrating the first embodiment of the abnormal operating condition cross-cell identification method for the flotation process in this application. Figure 2 This is a schematic diagram of time matching and sample construction of foam images and process values provided in the first embodiment of the abnormal cross-tank identification method for flotation process of this application; Figure 3 This is a schematic diagram of the dual-branch fusion recognition model of foam images and process values provided in the first embodiment of the abnormal working condition cross-tank recognition method for flotation process of this application; Figure 4 This is a schematic diagram of the two-level attention residual fusion process provided in the first embodiment of the abnormal working condition cross-cell identification method for the flotation process of this application; Figure 5 This is a flowchart illustrating the second embodiment of the abnormal operating condition cross-cell identification method for the flotation process in this application. Figure 6 This is a schematic diagram of the target cell model adaptation process based on category prototype constraints provided in the second embodiment of the abnormal operating condition cross-cell identification method for flotation process of this application. Figure 7 A schematic diagram of the reference cell model construction, target cell model adaptation and online identification process of the abnormal operating condition cross-cell identification method for flotation process provided in the second embodiment of this application; Figure 8 This is a schematic diagram of the module structure of the abnormal operating condition cross-cell identification device for the flotation process according to an embodiment of this application; Figure 9 This is a schematic diagram of the equipment structure of the hardware operating environment involved in the abnormal operating condition cross-cell identification method for the flotation process in the embodiments of this application. Detailed Implementation
[0015] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.
[0016] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.
[0017] It should be noted that the method for identifying cross-tank abnormalities in the flotation process can be executed by an industrial computer, server, edge computing device, or abnormal cross-tank abnormality identification system with data processing and program execution capabilities. The following explanation will focus on the abnormal cross-tank abnormality identification system as the executing entity.
[0018] Based on this, the first embodiment of this application provides a method for identifying cross-tank abnormalities in the flotation process, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the abnormal operating condition cross-cell identification method for the flotation process according to this application.
[0019] In this embodiment, the abnormal operating condition cross-cell identification method for the flotation process includes steps S10 to S50: Step S10: Obtain labeled samples of the reference slot and the target slot. The labeled samples include time-corresponding foam images, process numerical vectors, and operating condition labels. In this embodiment, the reference cell refers to a flotation cell in the same flotation process with relatively sufficient historical operating data and labeled samples, used to train the initial identification model; the target cell refers to other flotation cells in the same flotation process that require model adaptation using the existing identification capabilities of the reference cell. The feed composition, pulp concentration, foam load, and operating range of liquid level and aeration may differ between different cells, and the camera installation position and shooting angle may also vary. Therefore, the foam coverage, texture, and overflow pattern of the same operating condition need not be the same in different cells.
[0020] In this embodiment, a foam image refers to an image that reflects the state of the foam layer and discharge boundary in the flotation cell; a process value vector refers to a set of operating values arranged in a fixed variable order and corresponding to the time of the foam image; and a condition label refers to the category marker for the sample, such as normal, settling, or overflowing. Under normal operating conditions, foam coverage and product discharge remain in the normal operating state of the corresponding cell position; settling typically manifests as reduced foam coverage, exposure of local liquid levels or overflow weirs, and insufficient foam product discharge; overflowing typically manifests as foam or slurry exceeding the normal discharge range and continuously overflowing.
[0021] Samples for the baseline and target slots are stored separately, and the correspondence between samples is maintained through slot identifiers, acquisition time, and category labels. The baseline slot samples are used to establish initial recognition capabilities, while the target slot samples are used to adjust the model's recognition relationship for that slot. Model training uses condition labels as supervision information; during online recognition, the current foam image and process numerical vector are input, but the condition labels are not.
[0022] Step S20: Encode the foam image and the process numerical vector and perform two-level attention residual fusion to obtain fusion features, and train the baseline slot model based on the fusion features and the working condition label; In this embodiment, encoding refers to converting image pixels or process values into feature vectors that can participate in subsequent fusion calculations. Image features reflect information such as foam color, texture, coverage shape, and discharge boundary, while process value features reflect the corresponding operating conditions. Two-level attention residual fusion refers to adjusting the contributions of different feature dimensions within each feature path and the relative contributions between the two feature paths, while retaining the features before weighting to participate in the fusion calculation.
[0023] When training the baseline tank model, only the baseline tank training samples are used. The foam image and process numerical vector of each sample are encoded separately to form a fusion feature. Then, a working condition classification score is generated based on the fusion feature. The classification score is compared with the working condition label, and the model parameters are updated based on the comparison result. Here, the fusion feature is taken from before the classifier, and the working condition classification score is the numerical value of the three categories output by the classifier: normal, sinking, and running.
[0024] Step S30: Calculate the benchmark slot category prototype based on the similar fusion features output by the benchmark slot model, and initialize the target slot model with the parameters of the benchmark slot model. In this embodiment, the category prototype refers to a representative feature vector obtained by summarizing the fused features of multiple samples of the same working condition category, used to characterize the common features of the category in the fused feature space. The baseline trench category prototype corresponds to normal, sinking, and running trenches, respectively, and serves as the category reference when adapting the target trench model.
[0025] After training the baseline slot model, the fusion features of the baseline slot samples are extracted and combined with the working condition labels to form prototypes for each category. The target slot model adopts the network structure of the baseline slot model and uses the trained parameters as initial parameters to ensure that the fusion feature dimension and classification output order of the two are consistent. For multiple target slots, corresponding models are built separately, and each target slot model is initialized from the same baseline slot model and then adapted independently.
[0026] Step S40: Calculate the target slot category prototype for each training batch based on the fusion features output by the target slot model, keep the baseline slot category prototype unchanged, and update the target slot model by combining classification loss, alignment loss between similar prototypes and intra-class compactness loss of the target slot. In this embodiment, the training batch refers to a set of training samples used for a single parameter update; the classification loss is used to reflect the difference between the working condition classification score and the working condition label; the alignment loss between similar prototypes is used to constrain the feature differences of the same working condition category between the baseline slot and the target slot; the intraclass compaction loss of the target slot is used to constrain the feature dispersion of similar samples in the target slot relative to the prototypes of that category.
[0027] The target slot training samples are fed into the target slot model in batches. Based on the currently output fusion features and sample labels, a target slot category prototype is calculated. Each batch uses a fixed baseline slot category prototype as a reference, and the target slot model is adjusted using a combination of classification constraints and two prototype constraints. The baseline slot category prototype provides existing category information, while the labeled target slot samples enable the model to adjust the recognition relationship by combining the image of the slot with the runtime data.
[0028] Step S50: Input the foam image and process numerical vector corresponding to the current time of the target tank into the updated target tank model, and output the working condition identification result of normal, sinking or running out of the tank.
[0029] After adaptation is complete, the target slot model and corresponding data preprocessing parameters are saved. During online recognition, the corresponding model is loaded according to the slot identifier, and the same image processing and process numerical processing as in the training phase are performed on the current input to obtain three working condition classification scores. The category with the highest score is taken as the current recognition result, and it is saved in association with the slot identifier and image acquisition time, or output through the result interface of the on-site monitoring system.
[0030] Model training, target slot fitting, and online recognition are performed in stages. Online recognition only performs encoding, fusion, and classification calculations; it does not build category prototypes or calculate classification or prototype losses. This process utilizes the existing classification capabilities of the baseline slots and labeled samples of the target slots to complete slot fitting, which can reduce the impact of inter-slot feature differences on the identification of abnormal operating conditions.
[0031] As an example, the steps of obtaining labeled samples of the reference tank and the target tank include: reading the foam image, process variables, foam surface velocity, and operating condition label of the reference tank and the target tank respectively, wherein the process variables include the liquid level, liquid level regulating valve opening, aeration volume, and aeration regulating valve opening of the preceding tank and the current tank respectively; using the acquisition timestamp of the foam image as a reference, matching the process variables of the current tank and the preceding tank corresponding to the foam image within a preset time deviation, and matching the foam surface velocity within the time window corresponding to the foam image; removing records with missing data or time matching failures, arranging the matched process variables and the foam surface velocity according to the preset variable order to obtain a process value vector; associating the foam image, the process value vector, and the corresponding operating condition label to obtain labeled samples.
[0032] Please refer to Figure 2 , Figure 2 This diagram illustrates the time matching and sample construction of foam images and process values, as provided in the first embodiment of the abnormal cross-tank identification method for flotation processes according to this application. The diagram shows the foam images and their acquisition timestamps. Based on this, the process of time-matching and constructing samples is performed on the process data of the preceding tank, the process data of this tank, and the foam surface velocity related to the current tank position: Within a preset allowable time deviation, the liquid level, liquid level regulating valve opening, aeration volume, and aeration regulating valve opening of the preceding tank and this tank are matched, totaling 8 distributed control system process variables. At the same time, one foam surface velocity value within the corresponding time window of the foam image is matched and associated with the corresponding operating condition label. Records with missing data, failed time matching, or that do not meet data quality requirements are filtered out. The 8 process variables and 1 foam surface velocity value that pass the screening are arranged in the order of fixed variables to form a nine-dimensional process numerical vector of the i-th sample. And compare the vector with the corresponding bubble image. and working condition labels By associating data, labeled time-corresponding samples are obtained for model training; during online recognition, only samples from the current foam image are constructed. and nine-dimensional process numerical vector The input samples are composed without inputting operating condition labels. .
[0033] In this example, "this cell" refers to the flotation cell currently constructing the sample, and "the upstream cell" refers to the upstream flotation cell that supplies material to this cell in the flotation process. Process variables are read from the Distributed Control System (DCS), including the upstream cell level, upstream cell level control valve opening, upstream cell aeration rate, upstream cell aeration control valve opening, this cell level, this cell level control valve opening, this cell aeration rate, and this cell aeration control valve opening. The level and aeration rate reflect the operating status of the corresponding cell, and the valve opening reflects the degree of opening of the corresponding regulating component. All variables maintain the same units and numerical values during sample construction, training, and online identification.
[0034] The foam surface velocity refers to the movement rate of the foam surface within the image monitoring area. It is output by the existing dynamic image processing module and used as a numerical value in sample construction. The image window used by the dynamic image processing module should contain the current time corresponding to the foam image. For online recognition, a continuous image window ending at the current foam image acquisition time can be used, with the window endpoint serving as the time marker for the velocity value, avoiding the use of images after the current time. The window length follows the output window configured on-site for this module and remains consistent in training data construction and online recognition; for example, if the module supports it, the velocity can be calculated using images from 1 second before the current time to the current time.
[0035] Using the foam image timestamp as a reference, eight process variables (both in the preceding and current sections) are matched to the current foam image, along with the corresponding foam surface velocity. When multiple candidate records exist, the record with the smallest deviation from the image timestamp is selected; if the time deviations are the same, records no later than the image acquisition time are prioritized. The preset time deviation can be set to 90 seconds, based on factors including the process variable sampling period, time synchronization error, and the rate of change in operating conditions. 90 seconds is the upper limit for matching tolerance and does not mean that any record within this range can be used; when an operating condition change occurs within the time interval, the matching deviation should be reduced or the set of records should be removed. For sites with short sampling periods or rapid changes in operating conditions, a smaller allowable deviation should be used.
[0036] A 9-dimensional process numerical vector is constructed in the following order: upstream tank level, upstream tank level control valve opening, upstream tank aeration rate, upstream tank aeration control valve opening, current tank level, current tank level control valve opening, current tank aeration rate, current tank aeration control valve opening, and foam surface velocity. Each of the nine positions in the vector corresponds to a different variable, not nine consecutive time points, and is not used as a time series input. For upstream tank variables, the data should be read based on the relationship between the current tank and the upstream tank to avoid mixing in data from other tanks.
[0037] Records are discarded if there are missing images, missing any necessary variables, invalid values, or time matching failures. Continuous duplicate records can be identified based on slot identifiers, image source record identifiers, process data source record identifiers, and acquisition time; for consecutively read records of the same source group, only one group is retained. Variable values that are identical at different acquisition times are not used as a sole criterion for duplicate records, in order to retain valid samples from stable operation periods.
[0038] Operating condition labels are read from confirmed historical annotation records and associated with the corresponding image time or operating condition duration interval. Label confirmation is based on foam images, process operation records, and changes in operating conditions at adjacent time points. A "normal" label corresponds to a state where foam coverage and discharge are within normal ranges; a "settling tank" label corresponds to a state where foam discharge is insufficient and accompanied by reduced coverage, exposure of the liquid surface or weir edge, etc.; a "running tank" label corresponds to a state of continuous overflow exceeding the normal discharge range. Samples in the switching phase where existing records are insufficient to clearly define the category are temporarily excluded from labeled training samples to avoid using uncertain labels for model training.
[0039] The sample construction method maintains the time correspondence between foam appearance information, operating values, and operating condition labels, and maintains the consistency of the meaning of each value position by fixing the order of variables.
[0040] As an example, the steps of encoding the foam image and the process numerical vector and performing two-level attention residual fusion to obtain fused features, and training a baseline slot model based on the fused features and the working condition label include: inputting the foam image of the baseline slot into an image encoder to obtain image features; inputting the process numerical vector of the baseline slot into a process numerical encoder to obtain process numerical features; determining feature dimension weights and modal weights based on the image features and the process numerical features, and performing two-level attention weighting and residual fusion according to the feature dimension weights and the modal weights to obtain fused features; inputting the fused features into a classifier to obtain a working condition classification score; calculating a classification loss based on the working condition classification score and the working condition label, backpropagating the classification loss, and updating the parameters of the image encoder, the process numerical encoder, the attention fusion module, and the classifier to obtain the baseline slot model.
[0041] The baseline slot samples are divided into training, validation, and test sets based on the collection date to avoid the same consecutive collection segments being scattered across different data subsets. For example, 6300 samples can be divided into 4500 training samples, 900 validation samples, and 900 test samples. The training set is used for parameter updates, the validation set is used for model selection, and the test set is used for independent evaluation after training.
[0042] The bubble images were uniformly adjusted to 100×100 pixel RGB three-channel images, and each channel was normalized according to a fixed pixel scale, for example, converting the pixel values of the 8-bit image to 0 to 1. The process numerical vectors were standardized using the mean and standard deviation calculated variable-wise from the baseline slot training set. The standardized variables were dimensionless values, and the same statistics were used for the validation and test sets. For variables with a standard deviation of 0 or close to 0, the standardization denominator was restricted to a positive number, such as 0.000001, to avoid numerical anomalies, and the actual standardization parameters used were saved.
[0043] The image encoder employs a two-dimensional convolutional neural network (2D-CNN), with four 2D convolutional layers arranged sequentially. The first 2D convolutional layer has 3 input channels, and the kernel size of all four 2D convolutional layers is 3×3, resulting in 16, 32, 64, and 128 output channels respectively. Each convolutional layer is followed by a batch normalization layer and a rectified linear unit (ReLU) activation layer. After the first three activation layers, a max pooling layer is connected. Finally, an adaptive average pooling layer summarizes the spatial features of each channel into a single value, yielding 128-dimensional image features.
[0044] One possible size configuration is: a 2D convolution stride of 1 and edge padding of 1, a max pooling window of 2×2 and a stride of 2, and an end average pooling output spatial size of 1×1. This configuration allows the convolutional layers to extract local image information, the pooling layers to reduce the spatial size, and the end output dimension to be determined by 128 channels.
[0045] The process numerical encoder employs a one-dimensional convolutional neural network (1D-CNN). The 9-dimensional process numerical vector is processed into a single-channel input of length 9, which is then passed sequentially through two 1D convolutional layers with kernel size 3. The number of channels increases from 1 to 16, then to 32. Each 1D convolutional layer is followed by a batch normalization layer and a ReLU activation layer. The features are then flattened and mapped to 96-dimensional process numerical features through a fully connected layer.
[0046] In one possible configuration, the one-dimensional convolution has a stride of 1 and padding of 1 at both ends, maintaining an output length of 9. After flattening, this yields a 288-dimensional vector, which is then mapped to 96 dimensions. The one-dimensional convolution processes the input along the direction of the fixed variable arrangement; adjacent positions represent the combination relationship between different process variables, not their temporal order. Eight process variables and the foam surface flow velocity are input to the encoder without being separately split into encoding branches.
[0047] In this example, feature dimension weights are used to adjust the contribution of each component in the same feature path, while modality weights are used to adjust the relative contribution of image information and process numerical information; modality here represents the type of input information. The attention fusion module receives 128-dimensional image features and 96-dimensional process numerical features, and outputs 224-dimensional fused features after two levels of weighting and residual fusion. The specific calculation will be elaborated in subsequent examples.
[0048] The classifier consists of a fully connected layer with an input dimension of 224 and an output dimension of 128, a ReLU activation layer, a random deactivation layer with a deactivation probability of 0.30, and a classification output layer with an input dimension of 128 and an output dimension of 3. Random deactivation refers to the process of randomly setting some feature units to zero during the training phase. The classification output layer outputs three unnormalized classification scores arranged in a fixed order of normal, sinking, and running. During the training phase, the Softmax cross-entropy loss is calculated based on the three classification scores and the working condition label, and the two encoders, attention fusion module, and classifier are updated through backpropagation. In the verification and online recognition phases, the category with the highest classification score is taken as the working condition recognition result.
[0049] Please refer to Figure 3 , Figure 3This diagram illustrates the structure of a dual-branch fusion recognition model for foam images and process numerical values, provided in the first embodiment of the abnormal flotation process cross-tank identification method of this application. The diagram shows the process by which the foam image and the nine-dimensional process numerical vector are processed through two feature encoding paths to complete the condition classification: the foam image is input into a two-dimensional convolutional neural network image encoder, and then processed through two-dimensional convolution, batch normalization, nonlinear activation, pooling, and adaptive average pooling to obtain image features. The nine-dimensional process numerical vector is expanded into a single-channel one-dimensional input and then fed into a one-dimensional convolutional neural network process numerical encoder. After one-dimensional convolution, batch normalization, nonlinear activation, flattening, and fully connected mapping, the process numerical features are obtained. During the training phase, some numerical feature units are randomly deactivated, and this process is disabled during the verification and online recognition phases; subsequently, image features are... and process numerical characteristics The input modality-balanced dual-normalized residual attention fusion module is used to adjust the different feature dimensions within each feature path and the relative contributions between the two feature paths in turn. The original features before weighting are preserved through residual connection to obtain the fused feature Z. Finally, the fused feature Z is input into the fully connected classifier to obtain the classification scores corresponding to normal, sinking and running, and the category with the highest score is taken as the working condition recognition result.
[0050] Training can employ an Adaptive Moment Estimator (AdamW) with decoupled weight decay, a learning rate of 0.00025, weight decay of 0.0001, a batch size of 32, and a maximum training epoch count of 50. The macro-average F1 score on the validation set is used as the model selection criterion; this macro-average F1 score is the average of the F1 scores calculated for each class, with each class having equal weights. Training stops when the validation metric fails to improve for seven consecutive epochs, and the model parameters at which the validation metric reached their highest value are saved. Similarly, when the maximum number of training epochs is reached, the model with the highest validation metric is retained.
[0051] The training parameters mentioned above can be used as initial values for on-site configuration and adjusted based on validation set performance, without using a test set to select parameters. This network preserves the data structure of both types of inputs by separately encoding images and process numerical values, and establishes a classification relationship from fused features to working condition categories.
[0052] As an example, the step of determining feature dimension weights and modal weights based on the image features and the process numerical features, and performing two-level attention weighting and residual fusion according to the feature dimension weights and the modal weights to obtain fused features includes: performing linear mapping on the image features and the process numerical features respectively, normalizing the mapping scores along their respective feature dimensions to obtain feature dimension weights; multiplying the feature dimension weights by the corresponding feature dimensions, and then multiplying them element-wise with the corresponding image features or the process numerical features to obtain image recalibration features and process numerical recalibration features; respectively... Modal scores are calculated based on the image recalibration features and the process numerical recalibration features. The two modal scores are normalized to obtain image modal weights and process numerical modal weights. Based on the image modal weights and the process numerical modal weights, the image recalibration features and the process numerical recalibration features are weighted respectively and concatenated according to a preset concatenation order to obtain attention enhancement features. The image features and the process numerical features are concatenated according to the preset concatenation order to obtain residual features. The attention enhancement features and the residual features are added element-wise to obtain fusion features.
[0053] During the training phase, before attention fusion, random deactivation with a probability of 0.25 is performed on the 96-dimensional process numerical features. The process numerical features involved in the fusion calculation below refer uniformly to the output after this random deactivation process. In the same forward computation, the attention branch and the residual branch use the same output and do not generate separate random deactivation results. Random deactivation is disabled during the validation, baseline slot prototype extraction, testing, and online recognition phases, and the complete process numerical features are used.
[0054] Please refer to Figure 4 , Figure 4 This diagram illustrates the two-stage attention residual fusion process provided in the first embodiment of the abnormal cross-cell identification method for the flotation process according to this application. The diagram shows the image features. and process numerical characteristics The process of fusing the attention enhancement branch and the residual branch: In the attention enhancement branch, the image features are first processed separately. and process numerical characteristics A linear mapping is performed, and the first normalization is carried out along each feature dimension to obtain the corresponding feature dimension weights. Then, the weights are compensated using the feature dimensions of each path. The compensated weights are multiplied element-wise with the corresponding features to obtain the image recalibration features. and process numerical recalibration features .
[0055] Subsequently, modal scores are calculated based on the two recalibrated features, and a second normalization is performed on the two modal scores to obtain the image modal weights. and process numerical mode weights The recalibrated features are weighted using corresponding modal weights and concatenated in the order of image features first, followed by process numerical features, to form attention-enhanced features. In the residual branch, the original image features and process numerical features are directly concatenated in the same order to obtain the original feature concatenation result. Finally, the elements corresponding to the positions in the concatenation result of the attention-enhanced features and the original features are added together to obtain the fused feature Z.
[0056] In this example, recalibrated features refer to the feature vector after adjusting the weights of each component according to the feature dimensions. First, the scores of each feature dimension are obtained through a trainable linear mapping, and then Softmax normalization is performed on the 128 dimensions of image features and the 96 dimensions of process numerical features. The sum of the feature dimension weights within each path is 1, and then multiplied by the feature dimension of that path to compensate for the overall amplitude reduction caused by normalization.
[0057] The calculation formula is as follows: Layer normalization is performed on the two recalibrated features respectively, and then two modality scores are obtained through trainable linear scoring. Softmax normalization is performed on these two scores to obtain image modality weights and process numerical modality weights.
[0058] The calculation formula is as follows: The weighted results are concatenated in the order of image features first, followed by process numerical features, to form the attention-enhanced features; the two unweighted features are concatenated in the same order to form the residual features. The two are then added element-wise to obtain the fused features.
[0059] The calculation formula is as follows: In the above formula, m represents the mode, and its value is either the image mode v or the process numerical mode p; The modal features involved in the fusion are represented and processed as column vectors; Indicates the corresponding feature dimension. , ; For size The feature scoring matrix, for dimensional bias vector; The feature dimension score vector, The feature dimension weight vector, This is for relabeling features; ⊙ indicates element-wise multiplication; Layer Normalization is represented in this example and is calculated along the feature dimension of the current sample. for A 3D modal scoring vector, where T denotes the transpose. For scalar bias, Modal score; and These represent the image modal weights and the process numerical modal weights, respectively, and their sum is 1; the semicolon indicates vector concatenation; Z represents the 224-dimensional fusion feature. Each scoring parameter is updated during model training.
[0060] Layer normalization is used for modality scoring, and the weighting of augmented branches still targets the recalibrated features. The numerical stability term in the layer normalization calculation can be set to 0.00001. Feature dimension score normalization is performed within each modality, while modality score normalization is performed only between two modality scores; neither is calculated along the training batch dimension.
[0061] The first 128 positions of both branches correspond to image features, and the last 96 positions correspond to process numerical features, thus maintaining positional correspondence during addition. The residual branch retains the features before attention weighting, which can reduce the suppression of effective features by weight estimation bias.
[0062] As an example, the step of calculating the benchmark slot category prototype based on the same type of fusion features output by the benchmark slot model, and initializing the target slot model with the parameters of the benchmark slot model includes: selecting prototype samples from the labeled samples of the benchmark slot, the prototype samples covering normal, sinking, and running conditions; switching the benchmark slot model to inference mode, inputting the prototype samples, and extracting the fusion features before the classifier; grouping the fusion features according to the condition labels of the prototype samples, calculating the mean of each group of fusion features and performing Euclidean norm normalization to obtain the benchmark slot category prototype; saving the benchmark slot category prototype, and assigning the trainable parameters of the encoding, weighting, and classification of the benchmark slot model to the target slot model with the same structure to complete the initialization of the target slot model.
[0063] In this example, prototype samples are selected from the baseline tank training data according to the collection date and runtime, covering three operating conditions: normal, sinking, and running. Prototype samples are labeled samples specifically used to calculate prototypes for the baseline tank category; they can come from existing training sets, without using test set samples to build prototypes or select models.
[0064] For example, 2534 prototype samples can be selected, including 1208 normal samples, 866 sedimentation samples, and 460 running samples. Samples of each category are summarized separately, and it is not required that the three categories have the same number of samples; however, each category should have valid samples and should cover the performance of that type of working condition at different collection time periods as much as possible.
[0065] Switch the trained baseline slot model to inference mode, disable random deactivation, and use the statistics saved during training for the batch normalization layer without further updates. Input prototype samples, extract the 224-dimensional fused features before the classifier, group them according to the work condition label, calculate the mean of the features for each group, and then perform Euclidean norm normalization. Euclidean norm normalization is also known as... Normalization is the process of adjusting a non-zero vector to a vector of length 1 while preserving its direction.
[0066] It should be noted that, , and The subscript p in the text indicates the process numerical mode. and The letter 'p' in the symbol represents the category prototype, and the two are distinguished by the symbol position and context.
[0067] Z represents the general fusion feature output by the model. and These represent the fusion features corresponding to the reference slot sample i and the target slot sample i, respectively, and are the specific values of Z on the corresponding samples.
[0068] The formula for calculating the category prototype is as follows: In the formula, This is the prototype of reference slot category c; the superscript s indicates a reference slot; This is the set of prototype samples for this category. is the number of samples; i is the sample index; These are the fusion features of the corresponding samples; express Normalization. Vectors with values close to zero are processed according to the numerical stability rules in the following examples.
[0069] Three baseline slot class prototypes are associated with a fixed class order and stored. During target slot fitting, these prototypes are not recalculated, nor are they used as trainable parameters in gradient updates.
[0070] A target slot model with the same structure as the baseline slot model is established. The trainable parameters of the image encoder, process numerical encoder, attention fusion module, and classifier are copied to their corresponding positions. Simultaneously, the runtime statistics saved in the batch normalization layer are copied to form a complete initialization state. The target slot model maintains the 224-dimensional fusion features and the output order of normal, sinking, and running slots to avoid dimensional or category mismatches during prototype comparison or classification output. The process numerical normalization parameters of the target slot are calculated separately based on the target slot training data and are not mixed with the copied model parameters.
[0071] This process provides a fixed baseline slot category reference and enables the target slot model to be trained starting from the existing feature extraction and classification parameters.
[0072] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to the first embodiment described above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 5 , Figure 5 This is a flowchart illustrating a second embodiment of the abnormal operating condition cross-tank identification method for the flotation process according to this application. Step S40 of the abnormal operating condition cross-tank identification method for the flotation process includes steps S41 to S44: Step S41: Input the labeled samples of the target slot into the target slot model according to the training batch to obtain the fusion features and working condition classification scores; Step S42: Group the fusion features of the current training batch according to the working condition label, calculate the mean of each group of fusion features, and obtain the target slot category prototype. Step S43: Calculate the classification loss, alignment loss between similar prototypes, and intra-class compactness loss of the target slot based on the working condition classification score, the working condition label, the fusion feature, the baseline slot category prototype, and the target slot category prototype. Step S44: Keeping the baseline slot category prototype unchanged, backpropagate the weighted sum of the various losses and update the target slot model.
[0073] Please refer to Figure 6 , Figure 6 This diagram illustrates the target tank model adaptation process based on category prototype constraints, provided in the second embodiment of the abnormal operating condition cross-tank identification method for the flotation process of this application. The diagram shows the process of training the target tank model under the joint constraints of the benchmark tank category prototype and the target tank category prototype: The benchmark tank prototype samples are input into the benchmark tank model after training and switching to inference mode; the fusion features before the classifier are extracted; the fusion features are grouped according to the operating condition labels of normal, settling, and running tanks; the mean of each group of features is calculated and normalized to obtain the benchmark tank category prototype. This type of prototype is calculated before the target slot adaptation begins and remains fixed during the adaptation process, without participating in parameter updates. Figure 6 The “classification result” refers to the working condition classification score corresponding to normal, sinking and running in the target tank model classifier, and is not the final working condition identification result determined based on the maximum classification score. Figure 6 The “target slot label sample” refers to the target slot labeled sample.
[0074] A target slot model with the same structure is initialized using the parameters of the baseline slot model. Labeled samples from the target slot are then input into the target slot model in training batches to obtain the working condition classification score and the fusion features of each sample. The fused features are grouped and averaged according to the actual working condition labels of the categories that actually appear in the current batch to obtain the corresponding target slot category prototypes. Pair the target slot category prototype with a fixed reference slot category prototype according to the working condition category, and calculate the alignment loss between prototypes of the same type. The classification loss is calculated based on the working condition classification score and the working condition label. The intra-class compactness loss of the target slot is calculated based on the distance between the fusion features of each target slot sample and its class prototype. The three losses are weighted to form the total fitting loss L, and the total fitting loss is backpropagated to update the trainable parameters of the target slot model. The fixed baseline slot class prototype is not updated with backpropagation.
[0075] The labeled samples of the target slots are divided into training, validation, and test sets according to the collection date. For example, each target slot can use 920 training samples, 170 validation samples, and 170 test samples respectively. The images are processed using the same size and pixel method as the reference slots. The process numerical vectors are standardized using the mean and standard deviation calculated variable-wise from the corresponding target slot training set, and these statistics are saved for use in the validation, testing, and online recognition of the target slot.
[0076] The target slot training samples are input into the initialized target slot model in batches to obtain the fused features before the classifier and the condition classification score output by the classifier. During the training phase, the aforementioned random deactivation is enabled, and the image encoder, process numerical encoder, attention fusion module, and classifier all participate in the training.
[0077] Based on the actual working condition labels, the fusion features of the current batch are grouped, and the mean value of each group of features is calculated to obtain the target slot category prototype.
[0078] The calculation formula is as follows: In the formula, the superscript t represents the target slot, not time; The target slot category prototype for the current batch category c; This is the set of samples belonging to category c in the current batch; This represents the corresponding number of samples; Let i be the fusion feature of the target slot sample i.
[0079] Only prototypes for the actual categories appearing in the current batch are calculated. For example, if the current batch only contains normal and sinking samples, only prototypes for these two target sink categories are generated. Alternative prototypes for sinking are not constructed, and sinking is not included in the alignment loss of this batch. Prototypes for different batches are recalculated based on the current samples and the current model output, without using the prototypes from the previous batch to replace the prototypes of the current batch.
[0080] The mean calculation of the target slot class prototype is retained in the gradient calculation path without truncating the gradient. After calculating the classification loss, the alignment loss between prototypes of the same class, and the compaction loss within the target slot class, the weighted total loss is backpropagated. The classification loss updates all trainable parameters, including the classifier; the two prototype losses are backpropagated from the fused features before the classifier to the image encoder, procedural numerical encoder, and attention fusion module, without passing through the classifier update path; the fixed baseline slot prototype does not participate in gradient updates.
[0081] Target slot training can reuse the optimizer, learning rate, weight decay, batch size, maximum number of training epochs, and stopping condition from the baseline slot training, and select the model based on the macro-average F1 score of the corresponding target slot's validation set. Random deactivation is disabled during validation, and model parameters and batch-normalized runtime statistics are not updated. Each target slot retains the model with the highest validation metric and its standardized parameters; the test set is used only for final evaluation.
[0082] This adaptation process retains the reference to the baseline slot category while adjusting the feature extraction and classification relationship based on the actual samples of the target slot, and constrains the feature distribution of the target slot by using similar samples in the current batch.
[0083] As an example, the step of calculating the classification loss, the alignment loss between similar prototypes, and the intra-class compaction loss of the target slot based on the working condition classification score, the working condition label, the fusion feature, the baseline slot category prototype, and the target slot category prototype includes: calculating the cross-entropy based on the working condition classification score and the working condition label to obtain the classification loss; pairing the target slot category prototype with the baseline slot category prototype according to the working condition category and calculating the cosine distance of each pair of category prototypes; averaging the cosine distances corresponding to the actual working condition categories in the current training batch to obtain the alignment loss between similar prototypes; and pairing the fusion feature of each sample with the corresponding target slot category prototype according to the working condition label and averaging the cosine distances corresponding to each sample to obtain the intra-class compaction loss of the target slot.
[0084] In this example, the cosine distance is 1 minus the cosine similarity, used to reflect the directional difference between two non-zero feature vectors. When calculating the prototype alignment loss, pairing is performed according to category, comparing only normal with normal, sinking with sinking, and running with running; prototype pairing between different categories is not performed.
[0085] The calculation formula is as follows: When calculating the compact loss within the target slot class, the fused features of each sample are paired with the target slot class prototype corresponding to its real working condition label, and then the average distance of all samples in the current batch is calculated.
[0086] The calculation formula is as follows: The three losses are weighted and summed to obtain the total loss for target slot adaptation.
[0087] The calculation formula is as follows: In the above formula, The alignment loss is between similar prototypes. For compact loss within the target slot class, Let L be the classification loss, L be the total fitting loss, and B be the sample set of the current target slot training batch. This refers to the batch sample size. This is the set of actual operating condition categories that occur in the current batch. Number of categories; The actual working condition label for sample i. This is the prototype of the target slot category corresponding to this label; Indicates cosine similarity; and These are the non-negative weights of the two prototype losses, and the remaining symbols have the same meaning as in the previous text.
[0088] The prototype alignment loss is averaged over the actual categories, giving each category equal weight in the loss; the intra-class compaction loss is averaged over the samples, ensuring that every sample in the current batch participates in intra-class constraints. The averaging objects differ; the intra-class compaction loss is not changed by first averaging by category and then averaging again.
[0089] The classification loss coefficient is 1, the prototype alignment loss weight can be set to 0.08, and the intra-class compactness loss weight can be set to 0.02. These two weights are selected based on the classification performance on the target slot validation set, and their impact is checked in conjunction with the magnitude of each loss term during training. If the prototype constraint causes a continuous decline in the validation set classification performance, the corresponding weights can be reduced. After the parameters are selected, fixed weights are used for training, without adjustment based on test results.
[0090] To handle vectors with or close to the zero norm, the denominator is restricted to a preset positive number, such as 0.00001, during normalization. Since zero vectors lack comparable directions, the norms of the two participating vectors are checked before calculating the cosine distance; if either norm is not greater than this threshold, the distance term is set to 0 while retaining the original average count, thus preventing the generation of directional constraint gradients. The classification loss is still calculated normally, and batch training is not halted due to an invalid prototype term. The above loss formula applies to vectors with valid non-zero norms; numerical outliers are handled according to this rule.
[0091] When a category contains only one sample in the current batch, the category prototype and the sample have the same fused features, and the corresponding intra-class compactness loss is 0. The sample still participates in the classification loss, and its category prototype still participates in the cross-slot alignment loss when the norm is valid. This processing avoids mistaking a single-sample category as a category with multiple sample discreteness, while maintaining the computational continuity of classification and cross-slot constraints.
[0092] In this example, the trained baseline slot model achieved a macro-average F1 score of 0.9800 on the baseline slot test set. For the three target slots, using the baseline slot model directly without updating parameters resulted in an average macro-average F1 score of 0.5991. When the model was trained and adjusted using only labeled samples from the target slots and the classification loss, the average macro-average F1 score was 0.8211. Further incorporating alignment loss between similar prototypes and intra-class compaction loss for the target slots, the average macro-average F1 score for the three target slots reached 0.8358, with average recalls of 0.8323 and 0.7736 for slot inclusion and slot exclusion, respectively. These results demonstrate that, under the data conditions of this example, the two prototype constraints further improved the cross-slot recognition results based on supervised training of the target slots.
[0093] To evaluate the on-site identification results, 100 samples were drawn daily from the on-site monitoring records of 14 effective operating days, resulting in a total of 1400 independent verification samples. These included 1003 normal samples, 226 settling tank samples, and 171 samples from running tank samples. These samples underwent offline manual verification and were not used for model training, parameter selection, or model selection. The adapted tank model correctly identified 1328 samples, achieving an overall accuracy rate of 94.86%. This result is only intended to illustrate the implementation effect in this example and does not guarantee the same results in other concentrators or under other operating conditions.
[0094] To better understand the relationship between baseline slot model construction, target slot model adaptation, and online identification, please refer to [reference needed]. Figure 7 , Figure 7 The diagram illustrates the process of constructing the baseline slot model, adapting the target slot model, and online identification. Specifically: The figure illustrates the overall process of constructing the baseline slot model, adapting the target slot model, and online recognition: First, a sample consisting of time-corresponding foam images, process numerical vectors, and operating condition labels is constructed. Dual-branch feature encoding is performed through an image encoder and a process numerical encoder. During the training phase, the process numerical features are randomly deactivated. Then, modal balance dual normalized residual attention is used to fuse the two features. The baseline slot model is trained based on the fused features and the operating condition labels.
[0095] Subsequently, the target trench model is initialized using the parameters of the baseline trench model. The baseline trench category prototype and the target trench category prototype are calculated based on the fusion features of similar samples in the baseline trench and the fusion features of similar samples in the current training batch of the target trench. By aligning similar prototypes and combining target trench classification constraints and intra-class compactness constraints, the adaptive target trench model is obtained. During online recognition, the adapted target trench model is loaded, and a sample composed of the current foam image and the corresponding process numerical vector is constructed. The model forward calculation outputs the working condition recognition results of normal, sinking, or running trench.
[0096] This application also provides an abnormal operating condition cross-tank identification device for the flotation process; please refer to [reference needed]. Figure 8 The abnormal operating condition cross-tank identification device for the flotation process includes: The sample acquisition module 10 is used to acquire labeled samples of the reference slot and the target slot. The labeled samples include time-corresponding foam images, process numerical vectors, and operating condition labels. The benchmark training module 20 is used to encode the foam image and the process numerical vector and perform two-level attention residual fusion to obtain fusion features, and train the benchmark slot model based on the fusion features and the working condition label. The adaptation preparation module 30 is used to calculate the benchmark slot category prototype based on the same type fusion features output by the benchmark slot model, and initialize the target slot model with the parameters of the benchmark slot model. The cross-slot adaptation module 40 is used to calculate the target slot category prototype for each training batch based on the fusion features output by the target slot model, keep the baseline slot category prototype unchanged, and update the target slot model by combining classification loss, alignment loss between similar prototypes and compactness loss within the target slot class. The working condition identification module 50 is used to input the foam image and process numerical vector corresponding to the current time of the target tank into the updated target tank model, and output the working condition identification result of normal, sinking or running tank.
[0097] The abnormal operating condition cross-cell identification device for the flotation process provided in this application adopts the abnormal operating condition cross-cell identification method for the flotation process in the above embodiments. It can solve the technical problem of how to identify settling and running conditions when there are differences in the abnormal appearance characteristics of different flotation cells and the target cell labeling samples are limited. It has the same technical effect as the aforementioned method embodiments.
[0098] This application provides an abnormal operating condition cross-cell identification device for the flotation process. The abnormal operating condition cross-cell identification device for the flotation process includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the abnormal operating condition cross-cell identification method for the flotation process in the first embodiment described above.
[0099] The following is for reference. Figure 9 The diagram illustrates a structural schematic of an abnormal operating condition cross-tank identification device suitable for implementing embodiments of the present application for the flotation process. The abnormal operating condition cross-tank identification device for the flotation process in the embodiments of the present application can be an industrial computer, server, edge computing device, or data processing device deployed in a flotation field monitoring system. Figure 9 The structure shown is only one implementation method, and each hardware component can be configured according to the data acquisition method and computing resources on site.
[0100] like Figure 9 As shown, the abnormal operating condition cross-slot identification device for the flotation process may include a processing unit 1001 (e.g., a central processing unit, a graphics processor, etc.), which can perform various appropriate actions and processes according to a program stored in ROM 1002 (Read Only Memory) or a program loaded from storage device 1003 into RAM 1004 (Random Access Memory). RAM 1004 also stores various programs and data required for the operation of the abnormal operating condition cross-slot identification device for the flotation process. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via bus 1005. I / O interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, LCDs (Liquid Crystal Displays), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the flotation process anomaly cross-tank identification device to wirelessly or wiredly communicate with other devices to exchange data. Although the figure shows an anomaly cross-tank identification device for the flotation process with various systems, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented alternatively.
[0101] The abnormal operating condition cross-cell identification device for the flotation process provided in this application adopts the abnormal operating condition cross-cell identification method for the flotation process in the above embodiments. It can solve the technical problem of how to identify settling and running conditions when there are differences in the abnormal appearance characteristics of different flotation cells and the target cell labeling samples are limited. It has the same technical effect as the aforementioned method embodiments.
[0102] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the abnormal operating condition cross-cell identification method for the flotation process described in the above embodiments.
[0103] The computer-readable storage medium provided in this application can be a tangible medium capable of storing a computer program, such as a hard disk, random access memory, read-only memory, flash memory, optical storage device, or magnetic storage device. When the computer program is executed by a processor, it implements the above-described method for identifying abnormal operating conditions across slots in the flotation process.
[0104] The aforementioned computer-readable storage medium may be included in an abnormal operating condition cross-tank identification device for the flotation process; or it may exist independently and not be assembled into an abnormal operating condition cross-tank identification device for the flotation process.
[0105] The computer-readable storage medium provided in this application stores computer-readable program instructions (i.e., computer program) for executing the above-described method for identifying cross-cell abnormal conditions in the flotation process. It can solve the technical problem of how to identify settling and running conditions when there are differences in the abnormal appearance characteristics of different flotation cells and the target cell labeled samples are limited. It has the same technical effect as the aforementioned method embodiments.
[0106] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.
Claims
1. A method for cross-tank identification of abnormal conditions for a flotation process, characterized in that, The method includes: Acquire labeled samples of the reference slot and the target slot, wherein the labeled samples include time-corresponding foam images, process numerical vectors and operating condition labels; The foam image and the process numerical vector are encoded and fused using two levels of attention residuals to obtain fused features. A baseline slot model is then trained based on the fused features and the working condition label. The benchmark slot category prototype is calculated based on the similar fusion features output by the benchmark slot model, and the target slot model is initialized with the parameters of the benchmark slot model. The target slot class prototype for each training batch is calculated based on the fusion features output by the target slot model. The baseline slot class prototype remains unchanged. The target slot model is then updated by combining the classification loss, the alignment loss between similar prototypes, and the compactness loss within the target slot class. Input the foam image and process numerical vector corresponding to the current time of the target tank into the updated target tank model, and output the working condition identification result of normal, sinking or running out of the tank.
2. The method of claim 1, wherein, The step of calculating the target slot category prototype for each training batch based on the fusion features output by the target slot model, keeping the baseline slot category prototype unchanged, and updating the target slot model by combining classification loss, alignment loss between similar prototypes, and intra-class compactness loss of the target slot includes: The labeled samples of the target slot are input into the target slot model in training batches to obtain fusion features and working condition classification scores; Based on the working condition labels, the fusion features of the current training batch are grouped, and the mean value of each group of fusion features is calculated to obtain the target slot category prototype. Based on the working condition classification score, the working condition label, the fusion feature, the baseline slot category prototype, and the target slot category prototype, calculate the classification loss, the alignment loss between similar prototypes, and the compactness loss within the target slot category; Keeping the baseline slot category prototype unchanged, backpropagate the weighted sum of the various losses to update the target slot model.
3. The method of claim 2, wherein, The step of calculating the classification loss, the alignment loss between similar prototypes, and the compactness loss within the target slot class based on the working condition classification score, the working condition label, the fusion feature, the baseline slot category prototype, and the target slot category prototype includes: The cross-entropy is calculated based on the working condition classification score and the working condition label to obtain the classification loss; According to the working condition category, the target slot category prototype is paired with the reference slot category prototype, and the cosine distance of each pair of category prototypes is calculated; The average cosine distance corresponding to the actual working condition categories in the current training batch is used to obtain the alignment loss between prototypes of the same type. Based on the operating condition label, the fusion features of each sample are paired with the corresponding target slot category prototype, and the mean of the cosine distances corresponding to each sample is calculated to obtain the intra-class compact loss of the target slot.
4. The method of claim 1, wherein, The steps of encoding the foam image and the process numerical vector and performing two-level attention residual fusion to obtain fused features, and training a baseline slot model based on the fused features and the working condition label include: The foam image of the reference groove is input into an image encoder to obtain image features; The process numerical vector of the reference slot is input into the process numerical encoder to obtain the process numerical features; Based on the image features and the process numerical features, feature dimension weights and modal weights are determined, and two-level attention weighting and residual fusion are performed according to the feature dimension weights and the modal weights to obtain fused features; The fused features are input into the classifier to obtain the working condition classification score; The classification loss is calculated based on the working condition classification score and the working condition label. The classification loss is then backpropagated to update the parameters of the image encoder, the process numerical encoder, the attention fusion module, and the classifier to obtain the baseline slot model.
5. The method of claim 4, wherein, The step of determining feature dimension weights and modality weights based on the image features and the process numerical features, and performing two-level attention weighting and residual fusion according to the feature dimension weights and the modality weights to obtain fused features includes: Linear mapping is performed on the image features and the process numerical features respectively, and the mapping scores are normalized along their respective feature dimensions to obtain the feature dimension weights; Multiply the feature dimension weight by the corresponding feature dimension, and then multiply it element by element with the corresponding image feature or process numerical feature to obtain the image recalibration feature and the process numerical recalibration feature. Modal scores are calculated based on the image recalibration features and the process numerical recalibration features, respectively. The two modal scores are then normalized to obtain the image modal weights and the process numerical modal weights. Based on the image modal weights and the process numerical modal weights, the image recalibration features and the process numerical recalibration features are weighted respectively, and then spliced according to a preset splicing order to obtain attention enhancement features; The image features and the process numerical features are spliced together according to the preset splicing order to obtain residual features. The attention enhancement features are then added element by element to the residual features to obtain fused features.
6. The method of claim 1, wherein, The step of calculating the benchmark slot category prototype based on the similar fusion features output by the benchmark slot model, and initializing the target slot model with the parameters of the benchmark slot model includes: A prototype sample is selected from the tagged samples of the reference tank, the prototype sample covering normal, settling, and running conditions; Switch the baseline slot model to inference mode, input the prototype sample, and extract the fusion features before the classifier; The fusion features are grouped according to the working condition labels of the prototype samples. The mean of each group of fusion features is calculated and Euclidean norm normalization is performed to obtain the benchmark slot category prototype. Save the prototype of the baseline slot category, and assign the trainable parameters of the encoding, weighting, and classification of the baseline slot model to the target slot model with the same structure to complete the initialization of the target slot model.
7. The method of any one of claims 1 to 6, wherein, The steps for obtaining labeled samples of the reference and target slots include: Read the foam images, process variables, foam surface flow rate and operating condition labels of the reference tank and the target tank respectively. The process variables include the liquid level, liquid level regulating valve opening, air volume and air regulating valve opening of the front tank and the target tank respectively. Based on the acquisition timestamp of the foam image, the process variables of the current tank and its preceding tank corresponding to the foam image are matched within a preset time deviation, and the foam surface flow rate within the time window corresponding to the foam image is matched. Records with missing data or failed time matching are removed, and the matched process variables and the foam surface flow rate are arranged in a preset variable order to obtain a process numerical vector. The foam image, the process numerical vector, and the corresponding working condition label are associated to obtain a labeled sample.
8. A device for cross-tank identification of abnormal conditions in a flotation process, characterized in that, The apparatus is used to perform the abnormal operating condition cross-tank identification method for the flotation process as described in any one of claims 1 to 7, the apparatus comprising: The sample acquisition module is used to acquire labeled samples of the reference slot and the target slot. The labeled samples include time-corresponding foam images, process numerical vectors, and operating condition labels. The benchmark training module is used to encode the foam image and the process numerical vector and perform two-level attention residual fusion to obtain fusion features, and train the benchmark slot model based on the fusion features and the working condition label. The adaptation preparation module is used to calculate the benchmark slot category prototype based on the same type of fusion features output by the benchmark slot model, and initialize the target slot model with the parameters of the benchmark slot model. The cross-slot adaptation module is used to calculate the target slot category prototype for each training batch based on the fusion features output by the target slot model, keep the baseline slot category prototype unchanged, and update the target slot model by combining classification loss, alignment loss between prototypes of the same class and compactness loss within the target slot class. The working condition identification module is used to input the foam image and process numerical vector corresponding to the current time of the target tank into the updated target tank model, and output the working condition identification results of normal, sinking or running tank.
9. A flotation process oriented abnormal condition cross-cell identification device, characterized by, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the method for identifying cross-cell abnormal conditions in a flotation process as described in any one of claims 1 to 7.
10. A storage medium, characterized by The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the abnormal operating condition cross-slot identification method for the flotation process as described in any one of claims 1 to 7.
