Memory enhancement-based sludge dewatering state cascade identification method and device and terminal

Abstract

The application discloses a sludge dewatering state cascade identification method and device based on memory enhancement and a terminal, and belongs to the technical field of image detection. The method comprises the following steps: collecting a sludge liquid level ROI image and extracting multi-dimensional CV features; calculating the feature separation degree to determine the weight, and obtaining a liquid state tendency score through weighted fusion. According to the threshold partition routing, the high confidence interval directly outputs the identification result; the image is taken as a sample to be judged in the fuzzy interval, and a large model is used for identification; based on the CV features and the liquid state tendency score, positive and negative similar cases are retrieved from a memory enhancement library, case features and historical judgment information are extracted; the CV features, the case features and the historical judgment information are injected into prompt words, a large model is driven by few-sample learning to identify the state of the sample to be judged, and the output result is obtained. Through the cascade identification architecture of the CV feature adaptive weighted quantization grading, the high-low confidence partition routing and the positive and negative case few-sample prompt enhancement of the memory library, the sludge dewatering state identification is realized.

Description

Technical Field

[0001] This invention relates to the field of image detection technology, and in particular to a method, apparatus and terminal for cascaded recognition of sludge dewatering status based on memory enhancement. Background Technology

[0002] Sludge dewatering is a core process in wastewater treatment. The flocculation state of sludge directly determines the dewatering efficiency and the final sludge moisture content. Accurate detection of flocculation effect is a key prerequisite for achieving real-time optimization of chemical dosage and intelligent closed-loop operation of the plant.

[0003] Currently, sludge dewatering and flocculation status detection mainly includes: 1) manual on-site visual observation and manual chemical adjustment based on experience; 2) online acquisition of physical parameters such as turbidity and solids content by sensors; 3) direct input of images into a large-scale visual model for end-to-end recognition; and 4) extraction of manual features using traditional computer vision and determination based on threshold values.

[0004] However, existing technologies suffer from significant subjective errors in manual inspection, making 24-hour monitoring impossible. Sensors are expensive, susceptible to corrosion, and can only acquire single parameters. Pure large-scale model inference is slow, API calls are costly, and they are unsuitable for high-frequency monitoring. Traditional computer vision (CV) is fast but has poor generalization and is prone to misclassification under complex lighting conditions. Consequently, existing detection methods cannot simultaneously achieve high recognition accuracy, real-time response speed, and long-term operating costs, and recognition accuracy is difficult to improve with continuous operation of the factory. Summary of the Invention

[0005] This invention provides a memory-enhanced cascade identification method, device, and terminal for sludge dewatering status, which solves the technical problems of existing technologies that are difficult to balance the identification accuracy, real-time response speed, and deployment cost of sludge flocculation status detection, and that the identification accuracy cannot be continuously iterated and optimized with operation.

[0006] In a first aspect, embodiments of the present invention provide a cascade identification method for sludge dewatering state based on memory enhancement, comprising: Collect images of the sludge liquid surface and crop the ROI region. Extract multidimensional CV feature indicators from the ROI region. Calculate the separation degree of each CV feature indicator. Determine the weight based on the separation degree of each CV feature indicator. Use the weight to weight and fuse the multidimensional CV feature indicators to obtain a liquid tendency score. The liquid tendency score is partitioned and routed according to a preset threshold. When the liquid tendency score meets the high confidence interval, the sludge dewatering status identification result is directly output. When the liquid tendency score meets the fuzzy interval, the sludge liquid surface image is sent as a sample to be judged into the visual large model. Based on the multidimensional CV feature index and liquid tendency score of the sample to be judged, a two-layer screening mechanism is used to retrieve similar sample cases from the memory enhancement library. The case features and historical judgment information of the similar sample cases are extracted. The similar sample cases include two categories: positive examples and negative examples. The multidimensional CV feature indicators are injected into the prompt words, and the case features and historical judgment information of the similar sample cases are concatenated into the prompt words. Combined with few-shot learning to drive the large visual model, the state recognition of the sample to be judged is performed, and the sludge dewatering state recognition result is output.

[0007] Secondly, embodiments of the present invention also provide a memory-enhanced sludge dewatering state cascade identification device, comprising: The acquisition and fusion module is used to acquire sludge liquid surface images and crop ROI regions, extract multidimensional CV feature indicators from the ROI regions, calculate the separation degree of each CV feature indicator, determine the weight based on the separation degree of each CV feature indicator, and use the weight to weight and fuse the multidimensional CV feature indicators to obtain a liquid tendency score. The partitioned routing module is used to partition and route the liquid tendency score according to a preset threshold. When the liquid tendency score meets the high confidence interval, the sludge dewatering status identification result is directly output. When the liquid tendency score meets the fuzzy interval, the sludge liquid surface image is sent as a sample to be judged into the visual large model. The extraction and filtering module is used to retrieve similar sample cases from the memory enhancement library based on the multidimensional CV feature index and liquid tendency score of the sample to be judged, and to extract the case features and historical judgment information of the similar sample cases. The similar sample cases include two categories: positive examples and negative examples. The result output module is used to inject the multidimensional CV feature index into the prompt words, and to concatenate the case features and historical judgment information of the similar sample cases into the prompt words. It combines few-shot learning to drive the large visual model to perform state recognition on the sample to be judged, and outputs the sludge dewatering state recognition result.

[0008] Thirdly, embodiments of the present invention also provide a terminal, including: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the memory-enhanced sludge dewatering state cascade identification method provided in the above embodiments.

[0009] The present invention provides a memory-enhanced cascaded identification method, apparatus, and terminal for sludge dewatering state. The method involves acquiring sludge surface images and extracting Regions of Interest (ROIs). Multidimensional characterization (CV) features are extracted from the ROIs, and the separation degree of each CV feature is calculated. Weights are determined based on the separation degree of each CV feature, and the multidimensional CV feature indicators are weighted and fused using these weights to obtain a liquid state tendency score. The liquid state tendency score is then routed according to a preset threshold. When the liquid state tendency score meets the high confidence interval, the sludge dewatering state identification result is directly output; otherwise, when the liquid state tendency score meets the fuzzy interval... The sludge surface image is fed into a large-scale visual model as the sample to be judged. Based on the multidimensional CV feature index and liquid tendency score of the sample to be judged, a two-layer screening mechanism is used to retrieve similar sample cases from the memory enhancement library. The case features and historical judgment information of the similar sample cases are extracted. The similar sample cases include positive examples and negative examples. The multidimensional CV feature index is injected into the prompt words, and the case features and historical judgment information of the similar sample cases are concatenated into the prompt words. Combined with few-shot learning to drive the large-scale visual model, the state recognition of the sample to be judged is performed, and the sludge dewatering state recognition result is output. Through the cascaded architecture of traditional CV feature weighted hierarchical filtering, memory library positive and negative similar case retrieval and fusion into the prompt words of the large model, the sludge dewatering image recognition effect is achieved. Attached Figure Description

[0010] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0011] Figure 1 This is a flowchart of the cascaded identification method for sludge dewatering state based on memory enhancement provided in Embodiment 1 of the present invention; Figure 2 This is a flowchart of the cascaded identification method for sludge dewatering state based on memory enhancement provided in Embodiment 2 of the present invention; Figure 3 This is a structural diagram of the cascaded sludge dewatering status identification device based on memory enhancement provided in Embodiment 3 of the present invention; Figure 4 This is a structural diagram of the terminal provided in Embodiment 4 of the present invention. Detailed Implementation

[0012] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0013] Example 1 Figure 1 This is a flowchart of the cascaded identification method for sludge dewatering status based on memory enhancement provided in Embodiment 1 of the present invention. This embodiment is applicable to scenarios of online intelligent identification of sludge flocculation conditions and real-time determination of sludge dewatering status in wastewater treatment plants, and specifically includes the following steps: Step 110: Acquire sludge liquid surface image and crop ROI region, extract multidimensional CV feature indexes from ROI region, calculate the separation degree of each CV feature index, determine the weight according to the separation degree of each CV feature index, and use the weight to weighted fuse the multidimensional CV feature indexes to obtain liquid tendency score.

[0014] Images from the sludge dewatering workshop contain numerous irrelevant background objects such as railings and pipes. Directly processing the entire original image would introduce invalid interference pixels, significantly reducing the accuracy of subsequent feature extraction. By using image acquisition and ROI (Region of Interest) cropping, interference from on-site equipment can be filtered out, focusing on the effective observation area of ​​the sludge surface, ensuring the reliability of subsequent multidimensional CV feature extraction from the source. Optionally, original images of the sludge surface are acquired using industrial cameras installed in the sludge dewatering workshop at a preset acquisition frequency. Then, the central area of ​​the image is cropped as the ROI according to a ratio of 10%-90% on the x-axis and 15%-85% on the y-axis, removing edge interference objects such as concrete railings and pipes, retaining only the effective image of the sludge surface, eliminating the negative impact of irrelevant objects on feature calculation, and ensuring the accuracy and effectiveness of subsequent multidimensional CV feature extraction results.

[0015] The separation ratio is quantified by Fisher's discriminant ratio to differentiate between normal and abnormal sludge conditions using a single CV feature. It is calculated based on the ratio of the difference between the mean values ​​of the two types of samples to the standard deviation within the samples. The higher the separation ratio, the more significant the difference in the distribution of the feature on the two types of sludge surface images. This results in a higher weighting ratio when used for weighted fusion, a greater contribution to the differentiation of liquid tendency scores, and the ability to filter out invalid features with weak differentiation capabilities, thus ensuring the effectiveness of the feature system classification.

[0016] For example, five categories of visual CV features—color, texture, edge, reflection, and frequency domain—are extracted from the extracted ROI region. These include seven sub-indicators: Canny edge density, GLCM contrast, GLCM entropy, HSV brightness uniformity, highlight area proportion, highlight concentration, and 2D-FFT high-frequency energy proportion. Optionally, based on a dataset of 176 measured sludge images (85 normal samples and 91 abnormal samples), the Fisher discriminant ratio formula is used to calculate the separation degree corresponding to each CV feature indicator. The specific formula is as follows: ,in, This represents the mean of the corresponding CV feature index on the normal sample. This represents the mean of the corresponding CV feature index on the outlier samples. This represents the standard deviation of the corresponding CV feature index on normal samples. To correspond to the standard deviation of CV feature indicators on abnormal samples, the separation degree values ​​of each feature are obtained. Based on the separation degree, the weighted fusion weights of each feature are determined. The top six high-separation indicators are selected for participation in the liquid tendency score calculation. Frequency domain features, due to their lower separation degree, are only used as auxiliary judgment criteria for fuzzy intervals. The values ​​of the six core CV feature indicators are uniformly normalized to the same numerical range, and weighted summation is performed according to the fixed weights matched by the separation degree. Optionally, the Canny edge density score is multiplied by a weight of 0.30, the GLCM contrast score by a weight of 0.20, the GLCM entropy score by a weight of 0.15, the HSV brightness uniformity score by a weight of 0.15, the highlight area proportion score by a weight of 0.12, and the highlight concentration score by a weight of 0.08. After weighted summation, the output liquid tendency score is in the range of 0-1.

[0017] By filtering out irrelevant interference, quantifying multi-dimensional visual features, and using adaptive weighting based on classification capabilities, complex visual information about sludge surfaces is transformed into a single quantifiable liquid tendency score. This provides a unified quantitative criterion for subsequent dual-threshold rapid partitioning and routing, enabling rapid identification of simple operating conditions and accurate diversion of difficult and ambiguous samples, while balancing overall identification efficiency and discrimination accuracy.

[0018] Step 120: The liquid tendency score is partitioned and routed according to a preset threshold. When the liquid tendency score meets the high confidence interval, the sludge dewatering status identification result is directly output. When the liquid tendency score meets the fuzzy interval, the sludge liquid surface image is sent as a sample to be judged into the large visual model.

[0019] If all samples are uniformly fed into a large-scale visual model for recognition, it will result in huge computational overhead, slow down the real-time detection response speed, and make it difficult to meet the continuous online monitoring needs of the workshop. In sludge dewatering operations, there are a large number of clear samples with extremely high feature discrimination, and only a small number of samples with blurred boundaries are prone to misjudgment. Hierarchical routing can allocate computing resources in a targeted manner.

[0020] For example, two threshold groups, low_thresh and high_thresh, are pre-configured, optionally with values ​​of 0.45 and 0.55 respectively. The liquid tendency score in the 0-1 range is divided according to the preset thresholds. When the liquid tendency score is less than 0.45, it enters the high-confidence normal range, and a "normal" recognition result is directly output. When the liquid tendency score is greater than 0.55, it enters the high-confidence abnormal range, and an "abnormal" recognition result is directly output. When the liquid tendency score is between 0.45 and 0.55, it is determined to be an ambiguous range, and the current sludge surface image is marked as a sample to be judged and sent to the large-scale visual model for refined judgment. Simultaneously, an auxiliary verification mechanism is added after the high-confidence range judgment to check whether auxiliary features such as edge density and highlight ratio match the liquid tendency score conclusion. If the features match, the confidence level of this judgment is increased to 0.95; if the features conflict, the corresponding confidence level is decreased. By employing a dual-threshold, three-level traffic splitting mechanism, lightweight and rapid identification is achieved for simple working conditions, significantly reducing the frequency of calls to the large visual model and lowering online identification computing costs and latency. Only easily confused boundary samples are fed into the large model for fine-tuning, balancing overall identification real-time performance with the accuracy of difficult samples, thus achieving a balance between detection efficiency and identification precision.

[0021] Step 130: Based on the multidimensional CV feature index and liquid tendency score of the sample to be judged, a two-layer screening mechanism is used to retrieve similar sample cases from the memory enhancement library. The case features and historical judgment information of the similar sample cases are extracted. The similar sample cases include two categories: positive examples and negative examples.

[0022] Directly traversing the entire database to calculate sample similarity would incur massive computational overhead, resulting in excessively high retrieval latency in online real-time detection scenarios, failing to meet the rapid inference needs of the workshop. Relying solely on general vector similarity retrieval lacks constraints specific to the physical features of sludge processing conditions, making it difficult to accurately match visually confused samples and failing to provide simultaneous positive and negative historical examples.

[0023] The memory enhancement library internally stores standardized, structured case cards. Each case card contains three main categories of fixed fields. The first category is the search key field, which includes seven multi-dimensional CV features: liquid tendency score, Canny edge density, GLCM contrast, GLCM entropy, HSV brightness uniformity, highlight area proportion, highlight concentration, and 2D-FFT high-frequency energy proportion; thumbnail perceptual hash value; image acquisition time; on-site chemical dosage; sludge flow rate; and sludge solids content. The second category is the judgment result field, which includes the final normal / abnormal label of the sample, the original JSON of the large model inference, inference time, prompt word version, and small sample configuration number. The third category is the management information field, which includes the feedback source, confirmation status, case type label to distinguish positive / negative examples, search hit statistics, and quality score Q. Each card is labeled with a positive example (historically correctly judged) or a negative example (historically incorrectly judged) label as a criterion for search filtering. When performing a two-layer screening mechanism to retrieve similar cases, the first step is to conduct a single-dimensional coarse screening. Based on the liquid tendency score calculated for the current sample to be judged, a preset floating window of 0.05 is set above and below. Assuming that the liquid tendency score of the sample to be judged is 0.62, all case cards in the memory enhancement library with liquid tendency scores in the range of 0.57 to 0.67 are screened. A large number of irrelevant cards are quickly filtered by relying solely on a single numerical comparison, compressing tens of thousands of data points in the entire library into a set of dozens to hundreds of candidate cases, which greatly reduces the amount of subsequent calculations.

[0024] The second step performs multi-dimensional weighted ranking on the candidate case set output from the coarse screening. The Fisher discriminant ratio corresponding to each CV feature in step 110 is used as the weighting coefficient, and then substituted into the weighted Euclidean distance formula. In the formula The first sample to be judged The feature value y represents the th candidate case in the database. 1 eigenvalue, Using the Fisher discriminant ratio weight for the i-th feature, calculate the feature similarity distance between the sample to be judged and each candidate case. Sort all candidate cases according to the weighted Euclidean distance value from smallest to largest, and initially extract the top K cases. Then, perform a hard constraint of case type verification to check whether there are samples labeled as both positive and negative examples in the top K cases. If the natural ranking result only contains a single type of case, supplement the missing type sample closest to the sample to be judged from the remaining candidate cases to ensure that the final output of similar cases covers both positive and negative examples. Finally, extract all search key features, on-site process parameters, thumbnail hashes, historical judgment tags, original inference records of the large model, positive / negative example types, and other case features and historical judgment information of the filtered similar cases, and retain them for subsequent prompt word concatenation steps.

[0025] Additionally, it should be noted that to avoid the memory-enhanced library continuously storing valueless and misleading redundant case cards, the system assigns a dynamically updated quality score to each imported case card. Used for automated maintenance of the entire database, with a preset upper limit threshold M for the database capacity. When the total number of cases in the database reaches the upper limit threshold, a formula is used to determine when the database is fully automated. Calculate the quality score of all cards in batches, where This represents the hit accuracy rate of the large model after the sample cases are retrieved and cited. The cumulative number of citations for the sample cases. For time decay term, For conflict penalty items, , , and This is a weighting coefficient. When a card has similar characteristics to a newly added card but the judgment conclusion is opposite, the weighting coefficient is accumulated. The recommended initial configuration for the weighting coefficient is [value missing]. , , and Optimization can be performed by replaying historical datasets. After calculating the quality score of the entire database, elimination and cleanup are carried out in order of quality score from low to high. At the same time, boundary card protection rules are set. Cases with liquid tendency scores in the ambiguous interval between the low threshold (low_thresh) and the high threshold (high_thresh) are forcibly retained, accounting for no less than 30% of the total number, to prevent the database from containing only high-confidence, easily distinguishable samples and missing difficult boundary reference cases after long-term elimination. In addition, if the weighted Euclidean distance between a newly added card and an existing card in the database is less than the preset conflict threshold ε, but the dehydration status judgment conclusions of the two cards are completely opposite, a double-blind review process is triggered. The images of the two cards are independently sent to the large visual model for secondary judgment without historical reference, and the thumbnails and multi-dimensional CV features are simultaneously pushed to the human review. Based on the true conclusion determined by the review, the correct card is retained, and the incorrect card is deleted or corrected. At the same time, the reliability of the feedback channel for outputting the incorrect judgment conclusion is traced and verified, and the conflict penalty items of the two cards are updated accordingly. With quality Numerical value.

[0026] By leveraging a two-layer retrieval structure of coarse screening and weighted fine sorting, the computational load is significantly reduced, enabling rapid retrieval of similar cases while ensuring matching accuracy, thus meeting the real-time requirements of online sludge identification. Simultaneously, two types of contrasting cases with similar features are acquired, providing a two-way discriminative reference for the large visual model, fully utilizing historical operating experience, and reducing the probability of misjudgment of ambiguous boundary samples.

[0027] Step 140: Inject the multidimensional CV feature index into the prompt words, and concatenate the case features and historical judgment information of the similar sample cases into the prompt words. Combine the few-shot learning-driven visual large model to perform state recognition on the sample to be judged, and output the sludge dewatering state recognition result.

[0028] Relying solely on raw image information input to large models lacks quantitative visual priors and historical comparison samples, which can easily lead to blurred discrimination of boundary sludge samples, low accuracy, and an inability to fully utilize historical working condition identification experience.

[0029] Optionally, the V4 minimalist structured basic prompt words with iterative optimization are adopted. The core binary judgment criterion is "if the sludge surface is in a pure liquid state, the working condition is abnormal; if it is in a solid-liquid mixed state, the working condition is normal". The entire prompt word text is only about 200 words. Compared with the earlier 600-word full version, the total number of input tokens can be reduced by 60%, which takes into account both reasoning overhead and generalization ability. First, the multidimensional CV feature index injection operation is completed. All 7 CV quantitative features of the sample to be judged and the liquid tendency score are embedded in the front end of the prompt text in a standardized natural language format. The example text is "[CV pre-analysis auxiliary information]・Liquid tendency=0.512 (fuzzy interval sample), Canny edge density=0.0891, highlight area proportion=0.0956, HSV brightness uniformity=0.521, GLCM contrast=0.0098, GLCM entropy=0.0723, highlight concentration=0.0341, 2D-FFT high frequency energy proportion=0.0116". This allows the large visual model to directly read the quantitative visual numerical clues, reducing the uncertainty brought about by subjective visual judgment based solely on the image. Subsequently, the retrieved similar sample case information is layered and stitched together. First, based on the "case type" label built into the case card, the similar samples are divided into two major sections: positive examples and negative examples. The two sections are written independently within the prompt words to avoid information confusion. For similar samples marked as positive examples, the complete set of multi-dimensional CV feature values, liquid tendency score, thumbnail perception hash, on-site process parameters, and historical correct judgment results that have been verified through multiple rounds of feedback are stitched together as a positive discrimination reference template. The system selects 1-2 positive examples with the closest weighted Euclidean distance by default and writes them into the prompt words. For similar samples marked as negative examples, a special warning label is added to the prompt words, clearly stating that "the following sample is highly similar to the current sample to be judged in visual features, but the historical reasoning has been wrongly judged." The complete list of all CV feature indicators of the negative example, the original erroneous judgment conclusion of the large model, and the true standard result after manual / process feedback verification is fully listed. Only 1-2 negative examples that are closest to the sample to be judged and have the highest risk of discrimination confusion are selected to avoid excessive negative examples interfering with the model's discrimination logic. Finally, based on the above sample configuration, few-shot learning inference was performed. Using the default 2-shot example configuration, complete prompt words with injected CV features and concatenated positive and negative comparison cases were input into the Qwen series visual model for refined recognition. The model reads quantified feature priors, positive reference cases, and error-prone negative comparison cases for comprehensive inference, ultimately outputting standardized JSON format sludge dewatering state recognition results, along with supplementary fields such as inference confidence and judgment criteria. Through the few-shot enhanced prompt word mechanism of quantified CV feature pre-guidance and positive / negative boundary case comparison, the accuracy of binary classification of fuzzy sludge samples was significantly improved, while the prompt word token size was reduced, balancing recognition accuracy and online inference latency.

[0030] This embodiment acquires sludge surface images and extracts Regions of Interest (ROIs). Multidimensional characterization (CV) features are extracted from the ROIs, and the separation degree of each CV feature is calculated. Weights are determined based on the separation degree of each CV feature, and the multidimensional CV feature indicators are weighted and fused using these weights to obtain a liquid state tendency score. The liquid state tendency score is then routed to different regions based on a preset threshold. When the liquid state tendency score meets the high confidence interval, the sludge dewatering state identification result is directly output. When the liquid state tendency score meets the fuzzy interval, the sludge surface image is used as a candidate sample and fed into the large-scale visual model. Based on the multidimensional CV feature indicators and liquid state tendency score of the candidate sample, a two-layer screening mechanism is used to retrieve similar sample cases from the memory enhancement library. Case features and historical judgment information of the similar sample cases are extracted. The similar sample cases include positive and negative examples. The multidimensional CV feature indicators are injected into prompt words, and the case features and historical judgment information of the similar sample cases are concatenated into the prompt words. Combined with few-shot learning-driven large-scale visual model, the state identification of the candidate sample is performed, and the sludge dewatering state identification result is output. By employing a cascaded architecture that combines traditional CV feature weighted hierarchical filtering, memory-based positive and negative similar case retrieval, and integration with large model prompts, the image recognition effect for sludge dewatering is achieved.

[0031] Example 2 Figure 2 This is a flowchart of the cascaded identification method for sludge dewatering state based on memory enhancement provided in Embodiment 2 of the present invention. This embodiment is an optimization based on the above embodiment. After the step of combining few-shot learning to drive a large visual model to identify the state of the sample to be judged and outputting the sludge dewatering state identification result, it further includes: collecting three feedback signals for the sludge dewatering state identification result: time-series consistency feedback, posterior process verification feedback, and manual review feedback, and configuring corresponding confidence levels for each of the three feedback signals; integrating the three feedback signals to determine the truth or falsehood of the identification result; when the three feedback conclusions are consistent, marking the sample to be judged as a positive example and storing it in the memory enhancement library; when the three feedback conclusions conflict, selecting the feedback with the highest confidence value as the judgment basis, and marking the sample to be judged and storing it in the memory enhancement library; when the three feedback conclusions conflict and two feedback confidence levels are the same and the conclusions are contradictory, initiating a double-blind review process to determine the truth or falsehood of the sample to be judged, and marking the sample to be judged and storing it in the memory enhancement library.

[0032] See Figure 2 The memory-enhanced cascade identification method for sludge dewatering state includes: Step 210: Acquire sludge liquid surface image and crop ROI region, extract multidimensional CV feature indexes from ROI region, calculate the separation degree of each CV feature index, determine the weight according to the separation degree of each CV feature index, and use the weight to weight and fuse the multidimensional CV feature indexes to obtain liquid tendency score.

[0033] Step 220: The liquid tendency score is partitioned and routed according to a preset threshold. When the liquid tendency score meets the high confidence interval, the sludge dewatering status identification result is directly output. When the liquid tendency score meets the fuzzy interval, the sludge liquid surface image is sent as a sample to be judged into the large visual model.

[0034] Step 230: Based on the multidimensional CV feature index and liquid tendency score of the sample to be judged, a two-layer screening mechanism is used to retrieve similar sample cases from the memory enhancement library. The case features and historical judgment information of the similar sample cases are extracted. The similar sample cases include two categories: positive examples and negative examples.

[0035] Step 240: Inject the multidimensional CV feature index into the prompt words, and concatenate the case features and historical judgment information of the similar sample cases into the prompt words. Combine the few-shot learning-driven visual large model to perform state recognition on the sample to be judged, and output the sludge dewatering state recognition result.

[0036] Step 250: Collect three feedback signals for the sludge dewatering status identification result: time-series consistency feedback, post-process verification feedback, and manual review feedback, and assign corresponding confidence levels to each of the three feedback signals; integrate the three feedback signals to determine the authenticity of the identification result; when the three feedback conclusions are consistent, mark the sample to be judged as a positive example and store it in the memory enhancement library; when the three feedback conclusions conflict, select the feedback with the highest confidence value as the judgment basis and mark the sample to be judged and store it in the memory enhancement library; when the three feedback conclusions conflict and two feedbacks have the same confidence level and contradictory conclusions, initiate a double-blind review process to determine the authenticity of the sample to be judged and mark the sample to be judged and store it in the memory enhancement library.

[0037] Single-dimensional feedback cannot objectively verify the authenticity of sludge identification results. A multi-source hierarchical confidence feedback mechanism can achieve complementary three-level verification of fully automatic, semi-automatic and manual methods, accurately distinguishing positive and negative examples to complete the memory bank iteration.

[0038] Optionally, a three-tiered feedback channel with different confidence weights is designed to verify the authenticity of the sludge dewatering identification conclusions output by the large model. The first channel is a fully automatic time-series consistency feedback with a fixed confidence level of 0.6. Relying on the physical characteristic that sludge operating conditions do not change drastically in a short period of time, the recognition results of five consecutive frames are used as a reference. If an isolated abnormal frame appears in the middle, and the preceding and following frames are judged to be normal, then the frame is judged as a false alarm. The second channel is a semi-automatic post-process verification feedback with a confidence level of 0.8. When the system outputs an abnormal judgment, the control system adds flocculant and adjusts the sludge inlet flow rate. Downstream process indicators such as sludge moisture content, equipment torque, and flow stability are read at fixed intervals. If the indicators significantly improve, it proves that the judgment is accurate; if the indicators do not change or deteriorate, it is judged as a false alarm. The third channel is a manual review feedback with a maximum confidence level of 1.0. The operation and maintenance engineer only actively intervenes to correct the identification conclusions for samples with low-confidence judgments and rapid drops in the quality score Q. After completing the three-way feedback collection, a comprehensive judgment logic is executed. If the normal / abnormal conclusions of the three feedback outputs are completely consistent, the current sample to be judged is directly marked as a positive example and written into the standardized case card of the memory enhancement library. If the three feedback conclusions are inconsistent and conflicting, the feedback with the highest confidence value is selected as the final judgment criterion, and the corresponding result is marked as a positive example or a negative example and entered into the library. If a special conflict scenario occurs where the confidence values ​​of the time-series feedback and the posterior process feedback are both equal values ​​other than 0.6 and 0.8, and the two conclusions are completely opposite, a double-blind review process is initiated. The image of the sample to be judged without any historical prompt words is sent to the visual big model for a second independent judgment. At the same time, the sample thumbnail and a full set of multi-dimensional CV features are pushed to the human reviewer. The true and false of the sample are determined by combining the second output of the no-reference big model and the human review results. After completing the positive / negative example marking, the sample is stored in the memory enhancement library. Meanwhile, during the cold start phase of the system's launch, 176 fully labeled real-world images will be pre-imported as seed cases to fill the memory bank. During the transition phase, if there are fewer than two positive and negative comparison cases, memory retrieval will be temporarily suspended, and reasoning will rely solely on basic prompt words. Once the storage capacity in the database reaches more than half of the set capacity, it will enter a steady-state loop and continuously execute the entire process of self-iteration, including recognition, feedback, database entry, and periodic elimination.

[0039] By relying on hierarchical confidence multi-source feedback and conflict double-blind verification mechanism, the correctness of samples is accurately labeled, and reliable positive and negative cases are continuously accumulated for memory enhancement library, so as to realize long-term self-optimization of recognition model.

[0040] This embodiment acquires sludge surface images and extracts Regions of Interest (ROIs). Multidimensional characterization (CV) features are extracted from the ROIs, and the separation degree of each CV feature is calculated. Weights are determined based on the separation degree of each CV feature, and the multidimensional CV feature images are weighted and fused using these weights to obtain a liquid tendency score. The liquid tendency score is then routed according to a preset threshold. When the liquid tendency score meets the high confidence interval, the sludge dewatering state identification result is directly output. When the liquid tendency score meets the fuzzy interval, the sludge surface image is used as a candidate sample and fed into the large-scale visual model. Based on the multidimensional CV feature indicators and liquid tendency score of the candidate sample, a two-layer screening mechanism is used to retrieve similar sample cases from the memory enhancement library. Case features and historical judgment information of the similar sample cases are extracted. The similar sample cases include positive and negative examples. The multidimensional CV feature indicators are then... A prompt word is injected, and the case features and historical judgment information of similar sample cases are concatenated into the prompt word. Combined with few-shot learning-driven visual large model, the state of the sample to be judged is identified, and the sludge dewatering state identification result is output. Three feedback signals are collected from the sludge dewatering state identification result: time-series consistency feedback, posterior process verification feedback, and manual review feedback. Corresponding confidence levels are configured for each of the three feedback signals. The three feedback signals are integrated to determine the truth or falsehood of the identification result. When the three feedback conclusions are consistent, the sample to be judged is marked as a positive example and stored in the memory enhancement library. When the three feedback conclusions conflict, the feedback with the highest confidence value is selected as the judgment basis, and the sample to be judged is marked and stored in the memory enhancement library. When the three feedback conclusions conflict and two feedback signals have the same confidence level and contradictory conclusions, a double-blind review process is initiated to determine the truth or falsehood of the sample to be judged, and the sample to be judged is marked and stored in the memory enhancement library. Using the above methods, the authenticity of the identification results is accurately verified through multi-source differentiated confidence feedback and conflict review mechanisms, and the positive and negative labels of samples are standardized for storage. This enables the dynamic updating and continuous iteration of memory enhancement library case resources, allowing the model's recognition capabilities to be continuously optimized and upgraded based on actual working condition data.

[0041] Example 3 Figure 3 This is a schematic diagram of the cascaded identification device for sludge dewatering status based on memory enhancement provided in Embodiment 3 of the present invention, as shown below. Figure 3 As shown, the device includes: The acquisition and fusion module 310 is used to acquire sludge liquid surface images and crop ROI regions, extract multidimensional CV feature indicators from the ROI regions, calculate the separation degree of each CV feature indicator, determine the weight based on the separation degree of each CV feature indicator, and use the weight to weight and fuse the multidimensional CV feature indicators to obtain a liquid tendency score. The partition routing module 320 is used to partition and route the liquid tendency score according to a preset threshold. When the liquid tendency score meets the high confidence interval, it directly outputs the sludge dewatering status identification result. When the liquid tendency score meets the fuzzy interval, it sends the sludge liquid surface image as a sample to be judged into the visual large model. The extraction and filtering module 330 is used to retrieve similar sample cases from the memory enhancement library based on the multidimensional CV feature index and liquid tendency score of the sample to be judged, and to extract the case features and historical judgment information of the similar sample cases. The similar sample cases include two categories: positive examples and negative examples. The result output module 340 is used to inject the multidimensional CV feature index into the prompt words, and to concatenate the case features and historical judgment information of the similar sample cases into the prompt words. It combines the few-shot learning-driven visual large model to perform state recognition on the sample to be judged and outputs the sludge dewatering state recognition result.

[0042] The memory-enhanced sludge dewatering state cascade recognition device provided in this embodiment acquires sludge surface images and extracts Regions of Interest (ROIs). Multidimensional characterization (CV) features are extracted from the ROIs, and the separation degree of each CV feature is calculated. Weights are determined based on the separation degree of each CV feature, and the multidimensional CV feature indicators are weighted and fused using these weights to obtain a liquid state tendency score. The liquid state tendency score is then routed according to a preset threshold. When the liquid state tendency score meets the high confidence interval, the sludge dewatering state recognition result is directly output; when the liquid state tendency score meets the fuzzy interval, the sludge is... The liquid surface image is fed into the large-scale visual model as the sample to be judged. Based on the multidimensional CV feature index and liquid tendency score of the sample to be judged, a two-layer screening mechanism is used to retrieve similar sample cases from the memory enhancement library. The case features and historical judgment information of the similar sample cases are extracted. The similar sample cases include positive examples and negative examples. The multidimensional CV feature index is injected into the prompt words, and the case features and historical judgment information of the similar sample cases are concatenated into the prompt words. Combined with few-shot learning to drive the large-scale visual model, the state recognition of the sample to be judged is performed, and the sludge dewatering state recognition result is output. Through the cascade architecture of traditional CV feature weighted hierarchical filtering, memory library positive and negative similar case retrieval and fusion into the prompt words of the large-scale model, the sludge dewatering image recognition effect is achieved.

[0043] Based on the above embodiments, the acquisition and fusion module includes: The indicator extraction unit is used to extract multi-dimensional CV feature indicators from the ROI region. The multi-dimensional CV feature indicators include Canny edge density, GLCM contrast, GLCM entropy, HSV brightness uniformity, highlight area proportion, highlight concentration, and 2D-FFT high-frequency energy proportion. The separation degree calculation unit is used to calculate the separation degree of each CV feature index.

[0044] Based on the above embodiments, the partition routing module includes: The direct result output unit is used to set a low threshold and a high threshold. When the liquid tendency score is less than the low threshold or greater than the high threshold, the liquid tendency score satisfies the high confidence interval and directly outputs the normal sludge dewatering state identification result or the abnormal sludge dewatering state identification result. The large model precision judgment unit is used to send the sludge liquid surface image as a sample to be judged into the large visual model when the liquid tendency score is greater than or equal to the low threshold and less than or equal to the high threshold, and the liquid tendency score satisfies the fuzzy interval.

[0045] Based on the above embodiments, the extraction and filtering module includes: The candidate sample construction unit is used to select sample cases from the memory enhancement library whose difference between the liquid tendency score and the liquid tendency score of the sample to be judged is within a preset floating range, and form a candidate sample case set; The similar sample screening unit is used to calculate the weighted Euclidean distance between the sample to be judged and each candidate sample case in the candidate sample case set, using the separation degree of each CV feature index as a weighting coefficient, combined with the multidimensional CV feature index and the liquid tendency score. The candidate sample cases are sorted from smallest to largest according to the weighted Euclidean distance, and the top K similar sample cases that simultaneously contain positive and negative examples are screened.

[0046] Based on the above embodiments, the result output module includes: The sample segmentation unit is used to inject the multidimensional CV feature index into the prompt words and divide the similar sample cases into two categories, positive examples and negative examples, based on the historical judgment information of similar sample cases. The positive example splicing unit is used to splice the case features and correct historical judgment information corresponding to the positive example into the prompt word when a similar sample case is a positive example. The counterexample splicing unit is used to mark the counterexample in the prompt words when the similar sample case is a counterexample. The counterexample has similar visual features to the sample to be judged and has a history of misjudgment. It also lists the multidimensional CV feature indicators, historical erroneous conclusions and true standard results of the counterexample.

[0047] The memory-enhanced sludge dewatering state cascade identification device provided in this invention can execute the memory-enhanced sludge dewatering state cascade identification method provided in any embodiment of this invention, and has the corresponding functional modules and beneficial effects of the method.

[0048] Example 4 Figure 4 This is a schematic diagram of the structure of a terminal provided in Embodiment 4 of the present invention. Figure 4 A block diagram is shown that is suitable for implementing an exemplary terminal 12 of the present invention. Figure 4 The terminal 12 shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the present invention.

[0049] like Figure 4 As shown, terminal 12 is presented in the form of a general-purpose computing terminal. The components of terminal 12 may include, but are not limited to: one or more processors or processing units 16, system memory 28, and bus 18 connecting different system components (including system memory 28 and processing unit 16).

[0050] Bus 18 represents one or more of several bus architectures, including a memory bus or memory controller, a peripheral bus, a graphics acceleration port, a processor, or a local bus using any of the various bus architectures. For example, these architectures include, but are not limited to, the Industry Standard Architecture (ISA) bus, the Micro Channel Architecture (MAC) bus, the Enhanced ISA bus, the Video Electronics Standards Association (VESA) local bus, and the Peripheral Component Interconnect (PCI) bus.

[0051] Terminal 12 typically includes a variety of computer system readable media. These media can be any available media that can be accessed by terminal 12, including volatile and non-volatile media, removable and non-removable media.

[0052] System memory 28 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) 30 and / or cache 32. Terminal 12 may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, storage system 34 may be used to read and write non-removable, non-volatile magnetic media (… Figure 4 Not shown; usually referred to as a "hard drive"). Although Figure 4 Not shown, a disk drive for reading and writing to a removable non-volatile disk (e.g., a "floppy disk") and an optical disk drive for reading and writing to a removable non-volatile optical disk (e.g., a CD-ROM, DVD-ROM, or other optical media) may be provided. In these cases, each drive may be connected to bus 18 via one or more data media interfaces. System memory 28 may include at least one program product having a set (e.g., at least one) of program modules configured to perform the functions of the embodiments of the present invention.

[0053] A program / utility 40 having a set (at least one) of program modules 42 may be stored, for example, in system memory 28. Such program modules 42 include, but are not limited to, an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment. Program modules 42 typically perform the functions and / or methods described in the embodiments of the present invention.

[0054] Terminal 12 can also communicate with one or more external devices 14 (e.g., keyboard, pointing terminal, display 24, etc.), and with one or more terminals that enable a user to interact with terminal 12, and / or with any terminal (e.g., network card, modem, etc.) that enables terminal 12 to communicate with one or more other computing terminals. This communication can be performed via input / output (I / O) interface 22. Furthermore, terminal 12 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 20. As shown, network adapter 20 communicates with other modules of terminal 12 via bus 18. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with terminal 12, including but not limited to: microcode, terminal drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0055] The processing unit 16 executes various functional applications and data processing by running programs stored in the system memory 28, such as implementing the memory-enhanced sludge dewatering state cascade identification method provided in the embodiments of the present invention.

[0056] Note that the above description is merely a preferred embodiment of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, the scope of which is determined by the scope of the appended claims.

Claims

1. A cascaded identification method for sludge dewatering state based on memory enhancement, characterized in that, include: Collect images of the sludge liquid surface and crop the ROI region. Extract multidimensional CV feature indicators from the ROI region. Calculate the separation degree of each CV feature indicator. Determine the weight based on the separation degree of each CV feature indicator. Use the weight to weight and fuse the multidimensional CV feature indicators to obtain a liquid tendency score. The liquid tendency score is partitioned and routed according to a preset threshold. When the liquid tendency score meets the high confidence interval, the sludge dewatering status identification result is directly output. When the liquid tendency score meets the fuzzy interval, the sludge liquid surface image is sent as a sample to be judged into the visual large model. Based on the multidimensional CV feature index and liquid tendency score of the sample to be judged, a two-layer screening mechanism is used to retrieve similar sample cases from the memory enhancement library. The case features and historical judgment information of the similar sample cases are extracted. The similar sample cases include two categories: positive examples and negative examples. The multidimensional CV feature indicators are injected into the prompt words, and the case features and historical judgment information of the similar sample cases are concatenated into the prompt words. Combined with few-shot learning to drive the large visual model, the state recognition of the sample to be judged is performed, and the sludge dewatering state recognition result is output.

2. The method according to claim 1, characterized in that, The step of extracting multidimensional CV feature indicators from the ROI region and calculating the separation degree of each CV feature indicator includes: Multidimensional CV feature indicators are extracted from the ROI region. The multidimensional CV feature indicators include Canny edge density, GLCM contrast, GLCM entropy, HSV brightness uniformity, highlight area proportion, highlight concentration, and 2D-FFT high frequency energy proportion. Separation of each CV feature index The calculation is as follows: in, This represents the mean of the corresponding CV feature index on the normal sample. This represents the mean of the corresponding CV feature index on the outlier samples. This represents the standard deviation of the corresponding CV feature index on normal samples. This represents the standard deviation of the corresponding CV feature index on outlier samples.

3. The method according to claim 1, characterized in that, The step of partitioning and routing the liquid tendency score according to a preset threshold, directly outputting the sludge dewatering state identification result when the liquid tendency score meets the high confidence interval, and sending the sludge surface image as a sample to be judged into the large visual model when the liquid tendency score meets the fuzzy interval, includes: Set a low threshold and a high threshold. When the liquid tendency score is less than the low threshold or greater than the high threshold, the liquid tendency score meets the high confidence interval, and the normal sludge dewatering state identification result or the abnormal sludge dewatering state identification result is directly output. When the liquid tendency score is greater than or equal to the low threshold and less than or equal to the high threshold, the liquid tendency score satisfies the fuzzy interval, and the sludge liquid surface image is sent as a sample to be judged into the large visual model.

4. The method according to claim 1, characterized in that, The multidimensional CV feature index and liquid tendency score based on the sample to be judged employ a two-layer screening mechanism to retrieve similar sample cases from the memory enhancement library, including: Sample cases whose liquid tendency scores differ from those of the sample to be judged within a preset fluctuation range are selected from the memory enhancement library to form a candidate sample case set; Using the separation degree of each CV feature index as a weighting coefficient, and combining the multidimensional CV feature index with the liquid tendency score, the weighted Euclidean distance between the sample to be judged and each candidate sample case in the candidate sample case set is calculated. The candidate sample cases are sorted from smallest to largest according to the weighted Euclidean distance, and the top K similar sample cases that simultaneously contain positive and negative examples are selected.

5. The method according to claim 1, characterized in that, The step of injecting the multidimensional CV feature indicators into the prompt words and concatenating the case features and historical judgment information of the similar sample cases into the prompt words includes: The multidimensional CV feature index is injected into the prompt words, and the similar sample cases are divided into positive examples and negative examples based on the historical judgment information of similar sample cases. When a similar sample case is a positive example, the case features corresponding to that positive example and the correct historical judgment information are concatenated into the prompt words; When a similar sample case is a counterexample, the prompt text indicates that the counterexample has similar visual features to the sample to be judged and has a history of misjudgment, and lists the multidimensional CV feature indicators, historical erroneous conclusions and true standard results of the counterexample.

6. The method according to claim 1, characterized in that, After the step of combining few-shot learning to drive a large visual model to identify the state of the sample to be judged and outputting the sludge dewatering state identification result, the method further includes: The sludge dewatering status identification results are collected from three feedback signals: time-series consistency feedback, post-process verification feedback, and manual review feedback, and corresponding confidence levels are configured for each of the three feedback signals. The temporal consistency feedback is based on the short-term continuous and abrupt characteristics of sludge state. It verifies whether the current sample is a misjudgment by the recognition results of multiple consecutive frames, corresponding to the first confidence level. The post-process verification feedback is based on the reverse verification of changes in downstream process indicators of sludge to determine the accuracy of the results, corresponding to a second confidence level higher than the first confidence level; The manual review feedback is directly corrected by the maintenance personnel to correct the identification conclusion, corresponding to the highest third confidence level; The three feedback signals are integrated to determine the authenticity of the recognition result. When the three feedback conclusions are consistent, the sample to be judged is marked as a positive example and stored in the memory enhancement library. When the three feedback conclusions conflict, the feedback with the highest confidence value is selected as the judgment criterion, and the sample to be judged is marked and stored in the memory enhancement library; When the three feedback conclusions conflict and two feedbacks have the same confidence level but contradict each other, a double-blind review process is initiated to determine the authenticity of the sample to be judged, and the sample to be judged is marked and stored in the memory enhancement library.

7. The method according to claim 6, characterized in that, When the three feedback conclusions conflict and two feedbacks have the same confidence level but contradictory conclusions, a double-blind review process is initiated to determine the truth or falsehood of the sample to be judged. The sample to be judged is then marked and stored in a memory enhancement library, including: When the three feedback conclusions conflict and the two feedbacks have the same confidence level but contradict each other, a double-blind review is initiated. The sample to be judged is re-entered into the visual big model for judgment without any additional prompts. The thumbnail and multidimensional CV feature indicators of the sample to be judged are manually reviewed. The true or false of the sample to be judged is determined by combining the second judgment result of the big model and the result of the manual review. The sample to be judged is then marked and stored in the memory enhancement library.

8. The method according to claim 1, characterized in that, After the step of combining few-shot learning to drive a large visual model to identify the state of the sample to be judged and outputting the sludge dewatering state identification result, the method further includes: When the number of sample cases in the enhanced memory bank reaches the upper limit threshold, a quality score is calculated for all sample cases in the enhanced memory bank, and the calculation method is as follows: ； in, This represents the hit accuracy rate of the large model after the sample cases are retrieved and cited. The cumulative number of citations for the sample cases. For time decay term, For conflict penalty items, , , and These are the weighting coefficients; Sample cases in the augmented memory are deleted according to their quality scores, from lowest to highest.

9. A cascaded identification device for sludge dewatering status based on memory enhancement, characterized in that, include: The acquisition and fusion module is used to acquire sludge liquid surface images and crop ROI regions, extract multidimensional CV feature indicators from the ROI regions, calculate the separation degree of each CV feature indicator, determine the weight based on the separation degree of each CV feature indicator, and use the weight to weight and fuse the multidimensional CV feature indicators to obtain a liquid tendency score. The partitioned routing module is used to partition and route the liquid tendency score according to a preset threshold. When the liquid tendency score meets the high confidence interval, the sludge dewatering status identification result is directly output. When the liquid tendency score meets the fuzzy interval, the sludge liquid surface image is sent as a sample to be judged into the visual large model. The extraction and filtering module is used to retrieve similar sample cases from the memory enhancement library based on the multidimensional CV feature index and liquid tendency score of the sample to be judged, and to extract the case features and historical judgment information of the similar sample cases. The similar sample cases include two categories: positive examples and negative examples. The result output module is used to inject the multidimensional CV feature index into the prompt words, and to concatenate the case features and historical judgment information of the similar sample cases into the prompt words. It combines few-shot learning to drive the large visual model to perform state recognition on the sample to be judged, and outputs the sludge dewatering state recognition result.

10. A terminal, characterized in that, The terminal includes: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the memory-enhanced sludge dewatering state cascade identification method as described in any one of claims 1-8.

Images