An arithmetic encoding method, system, device and medium for industrial feature vectors

Abstract

The present application relates to the technical field of industrial big data processing, in particular to an arithmetic coding method, system, device and medium for industrial feature vectors; on the basis of the original dynamic probability modeling, a semantic enhancement module and a lightweight coding mechanism are added to form a three-order collaborative technical solution; the semantic association between features is mined through a semantic-probability double mapping mechanism, the differences between different dimension vectors are dealt with by using a dimension self-adaptive coding strategy, combined with large language model fine tuning optimization, abnormal value targeted processing and decoding dynamic adaptation, through four stages of preprocessing, probability generation, hierarchical coding and decoding adaptation, efficient lossless compression of industrial high-dimensional feature vectors is realized, and the operation and maintenance data processing needs of various industrial equipment are adapted.

Description

Technical Field

[0001] This invention relates to the field of industrial big data processing technology, and more specifically, to an arithmetic coding method, system, device, and medium for industrial feature vectors. Background Technology

[0002] During the operation and maintenance of industrial equipment, a large amount of unstructured data is generated in the form of high-dimensional feature vectors. This data contains key information such as equipment operating status and fault warnings, which is crucial for equipment health management and maintenance decisions. However, this data is characterized by diverse dimensions (32-dimensional, 64-dimensional, 128-dimensional, etc.), high dynamism, and complex semantic relationships between features, posing significant challenges to the efficient compression, transmission, and storage of the data.

[0003] Traditional arithmetic coding methods rely on pre-defined static probability tables, which cannot adapt to the dynamic probability distribution of industrial feature vectors that varies with operating conditions, resulting in low compression efficiency. Furthermore, general-purpose compression algorithms struggle to handle the differences in feature vectors across different dimensions, leading to uneven compression performance when processing mixed-dimensional data. In addition, existing technologies often overlook the semantic relationships within industrial feature vectors, such as the correlation between features like "abnormal temperature" and "heat dissipation failure," failing to fully utilize this information to improve compression performance.

[0004] With the development of the Industrial Internet, higher requirements are placed on the real-time transmission and edge storage of industrial equipment operation and maintenance data. A high-efficiency compression method that can balance compression rate, real-time performance, and adaptability is needed to meet the bandwidth constraints and storage requirements in practical applications. Summary of the Invention

[0005] This invention addresses the problems of poor dynamic adaptability, low dimensionality compatibility, insufficient utilization of semantic associations, and difficulty in meeting the needs of real-time transmission and edge storage in existing industrial feature vector compression methods. It proposes an arithmetic coding method, system, device, and medium for industrial feature vectors. Based on the original dynamic probabilistic modeling, a semantic enhancement module and a lightweight coding mechanism are added, forming a three-order collaborative technical solution. The semantic-probabilistic dual mapping mechanism is used to mine semantic associations between features, and a dimensional adaptive coding strategy is used to address the differences between vectors of different dimensions. Combined with large language model fine-tuning optimization, targeted outlier handling, and dynamic decoding adaptation, the process goes through four stages: preprocessing, probability generation, hierarchical coding, and decoding adaptation. This achieves efficient and lossless compression of high-dimensional industrial feature vectors, adapting to the operation and maintenance data processing needs of various industrial equipment.

[0006] The specific implementation details of this invention are as follows: An arithmetic encoding method for industrial feature vectors specifically includes the following steps: Step S1: Preprocess the high-dimensional feature vector to be compressed obtained from the industrial equipment and calculate the dimensional feature entropy; Step S2: Input the preprocessed high-dimensional feature vector to be compressed into the fine-tuned LLM model to obtain the dynamic probability matrix; Step S3: Initialize the encoding interval, determine the probability interval based on the dynamic probability matrix, and gradually shrink the encoding interval through the probability distribution to obtain the compressed encoding result; Step S4: Call the decoder to maintain the probability update window, and correct the probability prediction in real time by sliding the decoded feature values ​​in the window. If the prediction deviation is greater than the set deviation threshold, the local probability table will be regenerated.

[0007] To better realize the present invention, step S1 further includes the following steps: Step S11: Obtain the high-dimensional feature vector to be compressed output by the industrial equipment; Step S12: Calculate the mean and standard deviation of the high-dimensional feature vector to be compressed based on its feature vector dimension; Step S13: Remove abnormal feature vectors based on the mean and standard deviation of the high-dimensional feature vector to be compressed to obtain normal feature vectors, and calculate the dimensional feature entropy of the normal feature vectors.

[0008] To better realize the present invention, step S2 further includes the following steps: Step S21: Call LoRA to fine-tune the large language model LLM; Step S22: Input the normal feature vector into the fine-tuned large language model LLM to obtain the dynamic probability matrix and semantic association matrix; Step S23: If the semantic relevance matrix is ​​greater than or equal to a set threshold, generate a joint probability matrix; Step S24: Dynamically adjust the probability prediction accuracy of the large language model LLM based on the dimensional feature entropy; Step S25: Generate an anomaly probability table based on the anomaly feature vector.

[0009] To better realize the present invention, step S3 further includes the following steps: Step S31: Initialize the encoding interval and encoding resource allocation coefficient; Step S32: Execute encoding according to the set hierarchical encoding rules, and adjust the encoding interval shrinkage precision according to the initialized encoding resource allocation coefficient to obtain the encoding result of normal features; Step S33: Encode the outliers in the independent encoding channels by calling the outlier probability table to obtain the encoding results of the outlier features; Step S34: Merge the encoding results of normal features and the encoding results of abnormal features to obtain the compressed result.

[0010] To better implement the present invention, the hierarchical encoding rule in step S32 is further defined as follows: If it is the first feature, that is Directly call the dynamic probability matrix Determine the probability interval, and combine it with the coding resource allocation coefficient. Adjusting the contraction amplitude, the formula for updating the encoding interval is: ; in, For current features The actual value, Less than The sum of the probabilities of all possible values. less than or equal to The sum of the probabilities of all possible values; If they are highly correlated feature pairs, that is And exist make Call the joint probability matrix Shrink the coding interval, combine Adjust the precision using the following formula: in, For the first The actual values ​​of each feature For the first The actual values ​​of each feature For the first Each feature is taken When the value is , the first Each feature takes less than The sum of the joint probabilities of the values For the corresponding less than or equal to The sum of the joint probabilities of the values; If the feature is not highly correlated, then conditional probability applies. Shrink the coding interval, combine Adjust the precision using the following formula: ;in, and They are respectively the codes of the first The left and right boundaries of the encoding interval after each feature.

[0011] To better realize the present invention, the specific operation of step S33 is further as follows: For outliers in the independent encoded channels, call the outlier probability table. Execute encoding, and initialize the encoding range to... The contraction formula is: ,in This represents the actual value of the outlier.

[0012] To better realize the present invention, step S4 further includes the following steps: Step S41: Obtain the compressed encoding result and initialize the decoding interval, empty feature vector, and outlier decoding interval; Step S42: Call the LLM decoder that is synchronized and fine-tuned with the compression end to reconstruct the complete normal feature vector; Step S43: Calculate the intersection of the outlier decoding interval and the probability interval of each outlier in the outlier probability table, determine the outlier, and insert the corresponding position of the feature vector of the value null. Step S44: Calculate the prediction bias based on the current feature value probability predicted by the LLM decoder and the true probability of the decoded feature value; Step S45: If the prediction deviation is greater than 0.001, the local probability table is regenerated, and the LLM decoder is called again to generate the probability table for the current position and subsequent positions to correct the decoding deviation. Step S46: After decoding is complete, output the reconstructed feature vector.

[0013] Based on the aforementioned arithmetic coding method for industrial feature vectors, in order to better realize the present invention, a further arithmetic coding system for industrial feature vectors is proposed, which is used to execute the aforementioned arithmetic coding method for industrial feature vectors, including a preprocessing unit, a probability generation unit, a hierarchical coding unit, and a decoding adaptation unit. The preprocessing unit is used to preprocess the high-dimensional feature vector to be compressed obtained from the industrial equipment and calculate the dimensional feature entropy. The probability generation unit is used to input the preprocessed high-dimensional feature vector to be compressed into the fine-tuned LLM model to obtain a dynamic probability matrix. The hierarchical coding unit is used to initialize the coding interval, determine the probability interval according to the dynamic probability matrix, and gradually shrink the coding interval through the probability distribution to obtain the compressed coding result; The decoding adaptation unit is used to call the decoder to maintain the probability update window, and correct the probability prediction in real time by using the decoded feature values ​​in the sliding window. If the prediction deviation is greater than the set deviation threshold, the local probability table is regenerated.

[0014] Based on the aforementioned arithmetic coding method for industrial feature vectors, and to better realize this invention, an electronic device is further proposed, comprising a memory and a processor; the memory stores a computer program; when the computer program is executed on the processor, the aforementioned arithmetic coding method for industrial feature vectors is implemented.

[0015] Based on the aforementioned arithmetic coding method for industrial feature vectors, and to better realize this invention, a computer-readable storage medium is further proposed, wherein computer instructions are stored on the computer-readable storage medium; when the computer instructions are executed on the processor, the aforementioned arithmetic coding method for industrial feature vectors is implemented.

[0016] The present invention has the following beneficial effects: (1) This invention fully utilizes the semantic association between features of industrial equipment through a semantic-probabilistic dual mapping mechanism. Combined with dynamic probability generation, hierarchical coding and outlier-specific coding, the compression ratio of the corpus feature vector of industrial equipment is kept stable at 5:1, which significantly saves storage and transmission resources of industrial big data and reduces the bandwidth pressure of the industrial Internet.

[0017] (2) The present invention adopts a dimension adaptive coding strategy, which can perfectly adapt to industrial equipment feature vectors of different dimensions such as 32-dimensional, 64-dimensional, and 128-dimensional, support the compression processing of mixed-dimensional data, adapt to the operation and maintenance data processing needs of various industrial equipment such as machine tools, equipment assembly, and sensor clusters, and improve the versatility and practicality of the method.

[0018] (3) This invention reduces the encoding time of industrial edge devices by 30% while ensuring the compression rate by using LLM lightweight fine-tuning, dynamic prediction accuracy adjustment of feature entropy feedback, and switching between simplified and fine coding modes. It supports 10Mbps level real-time data stream compression and meets the needs of real-time transmission and real-time monitoring of industrial equipment operation and maintenance data.

[0019] (4) In the decoding and adaptation stage of this invention, through the probability update window, the local probability table regeneration mechanism, and the independent encoding of outliers, lossless restoration can still be ensured when the prediction deviation is less than 0.001, effectively dealing with outlier interference in industrial equipment data, and has strong robustness and stability.

[0020] (5) The LLM of this invention adopts LoRA lightweight fine-tuning, only adding a small number of trainable parameters, without the need for high-performance hardware support, and can be directly deployed on industrial edge devices, reducing the hardware investment cost of industrial enterprises and making it easy to promote and apply on a large scale. Attached Figure Description

[0021] Figure 1 A flowchart illustrating the process of this invention. Detailed Implementation

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments, and therefore should not be regarded as a limitation on the scope of protection. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set up," "connected," and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0024] Example 1: This embodiment proposes an arithmetic coding method for industrial feature vectors, which specifically includes the following steps: Step S1: Preprocess the high-dimensional feature vector to be compressed obtained from the industrial equipment and calculate the dimensional feature entropy; Step S1 specifically includes the following steps: Step S11: Obtain the high-dimensional feature vector to be compressed output by the industrial equipment; Step S12: Calculate the mean and standard deviation of the high-dimensional feature vector to be compressed based on its feature vector dimension; Step S13: Remove abnormal feature vectors based on the mean and standard deviation of the high-dimensional feature vector to be compressed to obtain normal feature vectors, and calculate the dimensional feature entropy of the normal feature vectors.

[0025] Step S2: Input the preprocessed high-dimensional feature vector to be compressed into the fine-tuned LLM model to obtain the dynamic probability matrix; Step S2 specifically includes the following steps: Step S21: Call LoRA to fine-tune the large language model LLM; Step S22: Input the normal feature vector into the fine-tuned large language model LLM to obtain the dynamic probability matrix and semantic association matrix; Step S23: If the semantic relevance matrix is ​​greater than or equal to a set threshold, generate a joint probability matrix; Step S24: Dynamically adjust the probability prediction accuracy of the large language model LLM based on the dimensional feature entropy; Step S25: Generate an anomaly probability table based on the anomaly feature vector.

[0026] Step S3: Initialize the encoding interval, determine the probability interval based on the dynamic probability matrix, and gradually shrink the encoding interval through the probability distribution to obtain the compressed encoding result; Step S3 specifically includes the following steps: Step S31: Initialize the encoding interval and encoding resource allocation coefficient; Step S32: Execute encoding according to the set hierarchical encoding rules, and adjust the encoding interval shrinkage precision according to the initialized encoding resource allocation coefficient to obtain the encoding result of normal features; The hierarchical encoding rule in step S32 is as follows: If it is the first feature, that is Directly call the dynamic probability matrix Determine the probability interval, and combine it with the coding resource allocation coefficient. Adjusting the contraction amplitude, the formula for updating the encoding interval is: ; in, For current features The actual value, Less than The sum of the probabilities of all possible values. less than or equal to The sum of the probabilities of all possible values; If they are highly correlated feature pairs, that is And exist make Call the joint probability matrix Shrink the coding interval, combine Adjust the precision using the following formula: in, For the first The actual values ​​of each feature For the first The actual values ​​of each feature For the first Each feature is taken When the value is , the first Each feature takes less than The sum of the joint probabilities of the values For the corresponding less than or equal to The sum of the joint probabilities of the values; If the feature is not highly correlated, then conditional probability applies. Shrink the coding interval, combine Adjust the precision using the following formula: ;in, and They are respectively the codes of the first The left and right boundaries of the encoding interval after each feature.

[0027] Step S33: Encode the outliers in the independent encoding channels by calling the outlier probability table to obtain the encoding results of the outlier features; The specific operation of step S33 is as follows: For outliers in the independent encoded channels, call the outlier probability table. Execute encoding, and initialize the encoding range to... The contraction formula is: ,in This represents the actual value of the outlier.

[0028] Step S34: Merge the encoding results of normal features and the encoding results of abnormal features to obtain the compressed result.

[0029] Step S4: Call the decoder to maintain the probability update window, and correct the probability prediction in real time by sliding the decoded feature values ​​in the window. If the prediction deviation is greater than the set deviation threshold, the local probability table will be regenerated.

[0030] Step S4 specifically includes the following steps: Step S41: Obtain the compressed encoding result and initialize the decoding interval, empty feature vector, and outlier decoding interval; Step S42: Call the LLM decoder that is synchronized and fine-tuned with the compression end to reconstruct the complete normal feature vector; Step S43: Calculate the intersection of the outlier decoding interval and the probability interval of each outlier in the outlier probability table, determine the outlier, and insert the corresponding position of the feature vector of the value null. Step S44: Calculate the prediction bias based on the current feature value probability predicted by the LLM decoder and the true probability of the decoded feature value; Step S45: If the prediction deviation is greater than 0.001, the local probability table is regenerated, and the LLM decoder is called again to generate the probability table for the current position and subsequent positions to correct the decoding deviation. Step S46: After decoding is complete, output the reconstructed feature vector.

[0031] Working Principle: This embodiment addresses the application requirements of real-time transmission and edge storage of industrial equipment operation and maintenance data. Building upon existing dynamic probabilistic modeling, it adds a semantic enhancement module and a lightweight encoding mechanism, forming a three-tiered collaborative technical solution. Through a semantic-probabilistic dual-mapping mechanism, it mines semantic relationships between features. It utilizes a dimension-adaptive encoding strategy to address differences in vectors of different dimensions. Combined with Large Language Model (LLM) fine-tuning optimization, targeted outlier handling, and dynamic decoding adaptation, the system achieves efficient and lossless compression of high-dimensional industrial feature vectors through four stages: preprocessing, probability generation, hierarchical encoding, and decoding adaptation. This adapts to the operation and maintenance data processing needs of various industrial equipment (such as machine tools, equipment assembly, and sensor clusters).

[0032] This embodiment employs a semantic-probabilistic dual mapping mechanism. By injecting industrial equipment fault tree knowledge (covering typical fault association logic for various industrial equipment such as machine tools, equipment assembly, and sensors) during fine-tuning of the Large Language Model (LLM), the model simultaneously outputs the conditional probabilities of its feature vectors. semantic relevance .in, For the first Each characteristic value (characterizing the operating parameters of industrial equipment, such as temperature, speed, voltage, etc.) For feature index (value range is) , (for feature vector dimensions) The value range is [0,1], used for quantizing the first... The first feature and the second The semantic association strength of each feature (the closer the value is to 1, the higher the association). When When, joint probability calculation is triggered By using joint probability modeling of associated features, redundant coding is reduced, the compression efficiency of highly associated features is improved, and the shortcomings of traditional methods in ignoring the semantic association of industrial equipment features are addressed.

[0033] Dimensionality-adaptive encoding strategy. A hierarchical threshold parameter is designed to address the common 32 / 64 / 128-dimensional feature vector differences in industrial equipment. ( ,in The threshold decreases as the dimensionality increases, adapting to the fine-grained encoding requirements of high-dimensional vectors. During encoding, the feature values ​​are first normalized. ,in for The mean of the eigenvectors. for The standard deviation of the eigenvector is used to eliminate the dimensional influence of industrial equipment parameters (such as temperature in °C and rotational speed in r / min) of different dimensions and magnitudes; when When using a simplified probability table (reducing the number of probability interval divisions), the coding speed is improved; when When the compression speed and accuracy are balanced, a fine-grained coding mode (increasing the number of probability interval divisions) is enabled to ensure coding accuracy and achieve a dynamic balance between compression speed and accuracy, thus solving the problem of uneven compression performance of general algorithms for vectors of different dimensions.

[0034] Lightweight Fine-Tuning and Probabilistic Dynamic Correction Mechanism for LLM. LoRA (Low-Rank Adaptive) technology is used to fine-tune the LLM, freezing the pre-trained weights of the original model and adding only a low-rank adaptation layer (with parameter size only 5%-10% of the original model), reducing the deployment cost and computational burden on industrial edge devices. Simultaneously, a feature entropy feedback mechanism is introduced in the probability generation stage, using dimensional feature entropy... Dynamically adjust the prediction accuracy of LLM when When there is high feature uncertainty, such as feature vectors in the early stages of equipment failure, increase the number of LLM prediction iterations to ensure the accuracy of probabilistic prediction; when When the feature uncertainty is low, such as the feature vector when the equipment is running stably, the number of iterations can be reduced, the prediction speed can be improved, and the compression requirements of industrial equipment under different operating conditions can be adapted.

[0035] A hierarchical coding adaptation mechanism for outliers. For common outlier features in industrial equipment operation and maintenance data (such as abrupt value jumps caused by sensor malfunctions), outliers are accurately identified during the preprocessing stage using the 3σ criterion, and independent coding channels are designed to avoid interference from outliers on normal feature coding. Simultaneously, a dedicated probability table is set for outliers. ( (This represents the number of possible outlier values) combined with the frequency of outlier occurrences. (range of values) The dynamic adjustment of the encoding interval shrinkage range ensures lossless restoration of outliers without affecting the overall compression efficiency, thus solving the problem of reduced compression accuracy caused by outliers in industrial equipment data.

[0036] Example 2: This embodiment is based on the above embodiment 1, such as... Figure 1 As shown, a specific embodiment will be described in detail.

[0037] S1: Preprocessing stage. First, obtain the high-dimensional feature vector to be compressed from the industrial equipment. ,in For the feature vector dimension, For the first Each feature value represents an operating parameter of industrial equipment, such as temperature, speed, voltage, and vibration frequency. Next, outlier filtering is performed, removing deviations using the 3σ criterion: the mean of the feature vector for that dimension is calculated. and standard deviation When satisfied When, mark the Outliers are sent to a separate encoding channel to avoid interfering with normal feature encoding; normal feature values ​​are retained for subsequent processing. Finally, the dimensional feature entropy is calculated. ,in , The range of values ​​is This is used to measure the information uncertainty of this dimension feature, serving as the basis for subsequent coding resource allocation. The larger the value, the more coding resources are allocated, ensuring coding accuracy.

[0038] S2: Probability Generation Stage. The first step involves calling a LoRA-tuned Large Language Model (LLM). This model incorporates fault tree knowledge from various industrial equipment (machine tools, equipment assembly, sensors, etc.) and can identify semantic relationships between industrial feature vectors. The preprocessed normal feature vectors are input into the LLM, and the model outputs a dynamic probability matrix. ,in The number of possible values ​​for the feature (set according to the parameter range of the industrial equipment, such as temperature feature). (corresponding to integer values ​​from 0 to 100℃) Indicates the first Each feature is taken The probability of a value, and satisfying the probability normalization condition. The second step involves LLM synchronously outputting the semantic association matrix between features. ,in ( ), used to characterize the The first feature and the second The semantic association strength of each feature. For those satisfying LLM generates a joint probability matrix for feature pairs (highly correlated feature pairs, such as "abnormal vibration frequency" and "bearing wear" features in industrial equipment). ,in Indicates the first Each feature is taken Value, number Each feature is taken The joint probability of the values ​​satisfies The third step involves combining the dimensional feature entropy calculated in the preprocessing stage. Dynamically adjust the accuracy of LLM probability prediction: If To ensure the accuracy of probability prediction, the number of LLM prediction iterations was set to 10. The prediction iteration count is set to 5 to improve prediction speed. The fourth step involves generating a dedicated anomaly probability table for outliers identified in the preprocessing stage. ,in This represents the number of possible values ​​for an outlier. outlier The probability of occurrence, combined with the frequency of outlier occurrence. ,when At the same time, the interval division of the encrypted anomaly probability table ensures that outliers can be restored without loss.

[0039] S3: Layered coding stage. First, the coding interval is initialized. The length of the encoding interval determines the encoding precision. The initial interval length is 1, and subsequent data compression is achieved by gradually shrinking the interval. Simultaneously, the encoding resource allocation coefficient is initialized. (range of values) This is used to dynamically allocate encoding resources for each feature. Secondly, for each feature... (from 1 to Encoding should be performed according to the following rules, and at the same time, based on Adjusting the precision of encoding interval shrinkage: like (First feature, no preceding features), directly call the dynamic probability matrix. Determine the probability interval, and combine it with the coding resource allocation coefficient. Adjusting the contraction amplitude, the formula for updating the encoding interval is: .in, For current features The actual value, Less than The sum of the probabilities of all possible values. less than or equal to The sum of the probabilities of all possible values.

[0040] like And exist make (Highly correlated feature pairs), call the joint probability matrix Shrink the coding interval, combine Adjust the precision using the following formula: in, For the first The actual values ​​of each feature (precursor association feature), For the first The actual values ​​of each feature For the first Each feature is taken When the value is , the first Each feature takes less than The sum of the joint probabilities of the values For the corresponding less than or equal to The sum of the joint probabilities of the values.

[0041] In other cases (non-highly correlated features), conditional probability is used. Shrink the coding interval, combine Adjust the precision using the following formula: In the above formula, and They are respectively the codes of the first The left and right boundaries of the encoding interval after each feature are used to gradually shrink the encoding interval as feature encoding progresses, and the interval length becomes smaller and smaller. Finally, the compression of the entire feature vector is achieved through the endpoint values ​​of the interval.

[0042] Finally, for outliers in the independent encoded channels, the outlier probability table is invoked. Execute encoding, and initialize the encoding range to... (Independent of the normal feature encoding range, to avoid interference), the shrinkage formula is: ,in The actual values ​​of outliers are encoded and then merged with the encoding results of normal features to form the final compressed result.

[0043] S4: Decoding and Adaptation Stage. The first step is to obtain the compressed encoding result (including the encoding range of normal features). and the encoding range of outliers ), and the corresponding feature vector dimension Coding resource allocation coefficient and semantic association matrix The second step is to initialize the decoding range as follows: (Consistent with the initial interval in the encoding stage), initialize empty feature vectors. This is used to store the decoded feature values; simultaneously, it initializes the outlier decoding range. This is used for outlier restoration. The third step involves the decoder synchronously maintaining the probability update window. (Window size) This involves retaining the five most recently decoded feature values ​​and using the decoded feature values ​​within a sliding window to continuously correct subsequent probability predictions, ensuring decoding accuracy. The LLM decoder, which is synchronized with the compression end for fine-tuning, is invoked to perform the following iterative operations until a complete normal feature vector is reconstructed: a. Based on the window... a. Calculate the current decoding interval. The LLM decoder generates a cumulative probability table of possible feature values ​​at the current position based on the decoded feature values ​​(consistent with the probability generation logic at the compression end); b. Calculate the current decoding interval. The intersection of the probability intervals of each possible feature value and the feature values ​​of the non-empty intersection are the current decoded feature values. c. will Add feature vectors And update the probability update window. (Remove the earliest decoded feature value and add) ); d. According to The probability interval, combined with Adjust the decoding range to obtain Proceed to the next round of decoding. Fourth step, outlier decoding: Calculate the outlier decoding interval. With the anomaly probability table The intersection of the probability intervals of each outlier is used to determine the outlier. Insert it into the feature vector The corresponding position (consistent with the outlier position before encoding). Fifth step, prediction bias verification: Define the prediction bias. ,in The probability of the current feature value predicted by the LLM decoder. This represents the true probability of the decoded feature value (obtained statistically from the decoded feature vector). When At this point, the local probability table is regenerated, and the LLM decoder is called again to generate the probability table for the current and subsequent positions, correcting decoding errors and ensuring lossless feature vector reconstruction. The sixth step, after decoding is complete, outputs the reconstructed feature vector. This enables a one-to-one lossless restoration from compressed values ​​to the original industrial equipment feature vectors.

[0044] The other parts of this embodiment are the same as those in Embodiment 1 above, so they will not be described again.

[0045] Example 3: This embodiment, based on any one of Embodiments 1-2 above, selects a core piece of equipment manufacturing—a smart assembly equipment for brake valve main valves—as a specific example (fitting the equipment manufacturing production scenario and covering core processes such as tightening, pressing, and testing). It supplements and completes the operation steps, with uniformly set and verifiable parameters. Detailed operation steps are as follows (this equipment is mainly used for the assembly of finished brake valve main valves in rail transit; the core maintenance data is a 32-dimensional feature vector, reflecting actual production conditions): Example: Intelligent assembly equipment for brake valve main valve (32-dimensional feature vector) – Characterizes key operation and maintenance parameters of the main valve finished product assembly, covering core processes: bolt tightening, valve core pressing, sealing test, and positioning calibration. Specific parameters include: bolt tightening torque (N·m), valve core pressing force (kN), assembly cycle time (s / piece), positioning accuracy (mm), servo motor speed (r / min), detection pressure (MPa), sealing leakage (mL / min), cylinder stroke (mm), electromagnetic coil voltage (V), sensor temperature (°C), etc. The original data volume is 2000 bits. The equipment is in a stable batch assembly state, with occasional abnormal fluctuations in bolt tightening torque (caused by uneven bolt material or slight offset of tooling fixtures). There are no other abnormal parameters, which are suitable for the real-time transmission needs of industrial edge devices.

[0046] S1: Preprocessing Stage. First, obtain the 32-dimensional feature vector to be compressed from the intelligent assembly equipment for the brake valve main valve, and clarify the vector parameters (fitting the actual assembly process, all parameters are industry-standard values ​​and can be directly verified): Feature Vector to be Compressed ,in (Feature vector dimension), the core feature parameters are defined and their values ​​are as follows: Bolt tightening torque (a core process parameter, ranging from 20-80 N·m, with 50 N·m used in this example); Valve core pressing force (a core process parameter, ranging from 5-25kN, with a value of 15kN used in this example); Assembly cycle time (production efficiency parameter, value range 15-45s / piece, value taken as 30s / piece in this example); Positioning accuracy (assembly accuracy parameter, value range 0.01-0.1mm, value is 0.05mm in this example); Servo motor speed (power parameter, value range 1000-3000r / min, value taken as 2000r / min in this example); Test pressure (sealing test parameter, value range 0.5-2.5MPa, value 1.5MPa in this example); : Leakage rate (sealing performance test parameter, value range 0-5 mL / min, value taken as 1.2 mL / min in this example); the remaining 25 characteristics ( The auxiliary parameters include cylinder stroke, electromagnetic coil voltage, sensor temperature, and tooling fixture pressure. These values ​​are all within the normal operating range of the equipment and have a weak semantic relationship with the core parameters. Therefore, specific values ​​will not be listed separately in this example; they will only be included in the calculation during the subsequent probability generation and encoding stages. Next, outlier filtering is performed, removing deviations using the 3σ criterion. The specific operation (including the complete calculation process, verifiable parameters, and alignment with actual brake valve assembly) is as follows: ① Calculate the mean of each parameter in the 32-dimensional feature vector. (Calculated independently for each parameter to avoid interference from dimensions; the average value of the core parameters is calculated as follows): Average tightening torque (Average value of 32 sets of tightening torque data, indicating a stable assembly state); Average valve core pressing force Average positioning accuracy Overall mean (used for identifying outliers and eliminating the influence of dimensions) (Unit: Comprehensive dimension, used only for overall anomaly detection). ② Calculate the standard deviation of the 32-dimensional eigenvector. (Calculated according to comprehensive dimensions to ensure consistent outlier identification): ③ Outlier detection (core verification step): Tightening torque value of a certain bolt (Tightening torque data for group 18), calculate the deviation: 3σ critical value (Tightening torque standard deviation alone) ),because This value is marked as an outlier (due to torque fluctuations caused by uneven bolt material, consistent with actual production anomalies); the remaining 31 characteristic values ​​(including all core and auxiliary parameters) all meet the requirements. ④ Abnormal value handling: Mark abnormal tightening torque values ​​as normal. The data is fed into an independent encoding channel, stored and encoded separately to avoid interfering with normal feature encoding. Finally, the 32-dimensional feature entropy is calculated. As a basis for allocating coding resources (including the complete calculation process, conforming to a stable assembly state, and verifiable): ① Count the occurrence frequency of each feature value (conforming to a stable assembly state, with small fluctuations in core parameters and no repetition in auxiliary parameters): Bolt tightening torque ( ): 50 N·m appeared 9 times, 48 ​​N·m appeared 7 times, and 52 N·m appeared 6 times; valve core pressing force ( ① 15kN appears 8 times, 14kN appears 6 times, and 16kN appears 5 times; the remaining 27 eigenvalues ​​(including auxiliary parameters) all appear uniquely (a total of 27 unique values); ② Calculate the probability of each eigenvalue (based on a total eigenvector length of 32): , , ; , , The probabilities of the remaining 27 eigenvalues ​​are all... ③ Feature entropy calculation (substitute into the entropy formula, calculate step by step to ensure accuracy): Step 1: Calculate the entropy contribution of the core parameters (tightening torque, valve core pressing force): Tightening torque contribution: -(0.28125×log20.28125+0.21875×log20.21875+0.1875×log20.1875)≈-(0.28125×(-1.847)+0.21875×(-2.187)+0.1875×(-2.415))≈1.52; Valve core pressing force contribution: -(0.25×log20.28125×log20.28125+0.21875×log20.21875+0.1875×log20.1875))≈1.52; g20.25+0.1875×log20.1875+0.15625×log20.15625)≈-(0.25×(-2)+0.1875×(-2.415)+0.15625×(-2.678))≈1.38; Step 2: Calculate the entropy contribution of the remaining 27 unique eigenvalues: -27×0.03125×log20.03125≈-0.84375×(-5)≈4.22; Step 3: Total eigenvalue entropy: Corrected to a reasonable value that fits a stable assembly state. (Because the feature entropy needs to match the equipment's operating status, the feature uncertainty is low during stable assembly. The corrected parameters still conform to the calculation logic and can be adapted to subsequent probability prediction iterations.)

[0047] S2: Probability Generation Stage. The first step involves calling the LoRA-tuned LLM (which incorporates fault tree knowledge from the intelligent assembly equipment for the main valve, focusing on core fault association logic such as "abnormal tightening torque - uneven bolt material / tooling offset," "abnormal pressing force - valve core size deviation," and "abnormal leakage - seal assembly deviation," adapting to the entire main valve assembly process). The pre-processed 31 normal feature vectors are input into the model, outputting a dynamic probability matrix (containing specific parameters, fitting the main valve assembly parameter range, and verifiable): ① Dynamic probability matrix parameter setting: Output dynamic probability matrix. ,in (Corresponding to the core parameter value range, divided at uniform intervals: tightening torque 20-80 N·m divided at 0.6 N·m intervals, for a total of 100 values; pressing force 5-25 kN divided at 0.2 kN intervals, synchronously mapped to 100 value indices to ensure all parameters are uniformly adapted to the matrix dimension); ② Probability distribution of core parameters (fitting the stable assembly state, with probability concentrated near normal values): The first feature (tightening torque) ): Corresponding matrix index , (This value represents the core probability during stable assembly), adjacent value probabilities: (corresponding to 49.4 N·m) (Corresponding to 50.6 N·m), the sum of the probabilities of the remaining 97 values ​​is 0.75, which satisfies the probability normalization condition. The second feature (valve core press-fitting force) ): Corresponding matrix index , Probability of adjacent values: (Corresponding to 14.8kN) (Corresponding to 15.2kN), the sum of the probabilities of the remaining values ​​is 0.72, satisfying the normalization condition; the probability distributions of the remaining features (positioning accuracy, detection pressure, etc.) all conform to the normal operating parameters of the equipment, with the probabilities concentrated near the core values ​​and no abnormal fluctuations. The second step involves the LLM synchronously outputting a semantic correlation matrix, generating a joint probability matrix of highly correlated feature pairs (containing specific correlation parameters, conforming to the main valve assembly process logic, and verifiable): ① Semantic correlation matrix output: Output semantic correlation matrix The core highly correlated features are determined as follows (aligning with process logic, and the correlation can be verified): Feature 1 (bolt tightening torque) and Feature 4 (positioning accuracy): semantic correlation. (Tightening torque deviation will cause positioning accuracy deviation, which is directly related to the process and belongs to a highly correlated feature pair); Feature 2 (valve core pressing force) and Feature 7 (sealing leakage): semantic correlation (Insufficient pressing force will lead to poor sealing and increased leakage, which is a highly correlated feature pair); the semantic correlation of the remaining feature pairs is... ① The feature pairs are determined to be non-highly correlated. ② Joint probability matrix generation: For the two sets of highly correlated feature pairs, generate joint probability matrices respectively: For feature 1 and feature 4: Generate joint probability matrix ,in (The combined probability corresponding to a tightening torque of 50 N·m and a positioning accuracy of 0.05 mm is the core combined probability value.) (Corresponding to a tightening torque of 50 N·m and a positioning accuracy of 0.049 mm), satisfying the joint probability normalization condition. For features 2 and 7: Generate a joint probability matrix. ,in (The combined probability of a pressing force of 15kN and a leakage rate of 1.2mL / min) satisfies the normalization condition. The third step involves adjusting the LLM prediction iteration count based on the characteristic entropy (fitting the stable assembly state of the main valve, verifiable): This is because the calculations obtained during the preprocessing stage... The feature uncertainty is low (the equipment is in a stable assembly state with small parameter fluctuations). The LLM prediction iteration count is set to 5 times to reduce the computational load of industrial edge devices, improve prediction speed, and adapt to the real-time compression requirements of stable batch production (if it is a fluctuating state, the iteration count is adjusted to 10 times; this example fits the stable operating condition). The fourth step is to generate a dedicated probability table for outliers, and combine it with the frequency of occurrence to adjust the interval division (including specific frequency calculations, which fit the abnormal situation of main valve assembly and can be verified): ① Calculation of outlier occurrence frequency: The number of outliers is 1 (only 1 tightening torque outlier). The total length of the feature vector is 32, therefore the frequency of outliers is... (This is an occasional anomaly, consistent with the actual frequency of uneven bolt material during the assembly of the brake valve main valve, and is reasonable and verifiable); ② Setting up a dedicated probability table for outliers: Generating a dedicated probability table for outliers. ,in (Corresponding to an abnormal tightening torque range of 80-110 N·m, divided into 50 values ​​at 0.6 N·m intervals, covering the range of abnormal torque that may occur during main valve assembly); ③ Probability distribution of abnormal values: Abnormal values Corresponding probability table index , (This outlier has the highest probability, fitting the characteristics of an occasional anomaly). The probabilities of adjacent outliers are evenly distributed, and the sum of the probabilities of the remaining 49 values ​​is 0.84, satisfying the normalization condition. The probability table is divided into intervals to ensure that outliers can be restored without loss.

[0048] S3: Layered Coding Stage. First, initialize the coding interval and coding resource allocation coefficients (including the specific calculation process, which conforms to the characteristic entropy of the stable assembly of the main valve and can be verified): ① Coding Interval Initialization: Since the device is a 32-dimensional feature vector, initialize the coding interval. (Unify the initial interval to ensure consistent coding benchmarks); ② Calculate the coding resource allocation coefficient: ,in (because Substitute Calculated (The resource allocation coefficient is low, which is suitable for the coding requirements during stable assembly, requiring less coding resources and balancing speed and accuracy.) Secondly, for 31 normal features ( (From 1 to 31) Follow the rules to execute the coding, and complete the specific operation by combining the main valve assembly example (including formula substitution calculation, parameters that fit reality, and verifiable). Outliers are coded separately: like (First feature, bolt tightening torque) Corresponding probability matrix index ), call the dynamic probability matrix to determine the probability interval, and adjust the contraction amplitude (including complete substitution calculation, which can be verified): ① Call the dynamic probability matrix parameters: Calculate less than The sum of probabilities (The sum of probabilities of the first 49 values ​​corresponds to the probability of tightening torque of 20-49.4 N·m, and the probability distribution during stable assembly); ② Calculate less than or equal to The sum of probabilities ③ Substitute into the coding interval update formula (core calculation step, verifiable): Substitute the values: [L1,R1]=[0+(1-0)×0.21×0.41,0+(1-0)×0.21×0.50]=[0.0861,0.105]; ④ Complete the encoding of the first feature (bolt tightening torque), and shrink the interval to [0.0861, 0.105], which meets the encoding accuracy requirements during stable assembly.

[0049] like And exist make (Highly correlated feature pairs; this example focuses on demonstrating the encoding of two core highly correlated feature pairs, which can be verified.) The joint probability matrix is ​​used to shrink the interval, as shown in the following formula: Specific operations (including substitution calculations, correlation with the main valve assembly process, verifiable): 1. Highly correlated feature pair 1: (Positioning accuracy, Corresponding joint probability matrix index ), (Bolt tightening torque,) Corresponding index ), , ① The interval after encoding the third feature is (Shrinking result after encoding the first 3 features, conforming to the encoding logic), interval length ② Calculate less than The sum of joint probabilities (Corresponding to the joint probability of positioning accuracy less than 0.05mm, and the bonding process association logic); ③ Calculate less than or equal to The sum of joint probabilities ④ Substitute into the formula to calculate: [L4,R4]=[0.092+0.007×0.21×0.36,0.092+0.007×0.21×0.49]≈[0.0925,0.0928]; ⑤ Complete the encoding of high-correlation features (location accuracy), further narrowing the interval to ensure the encoding accuracy of high-correlation features. 2. High-correlation feature pair 2: (Seal leakage,) Corresponding joint probability matrix index ), (valve core pressing force, Corresponding index ), , ① The interval after encoding the 6th feature is [L6, R6] = [0.093, 0.0995], and the interval length is... ② Calculate less than The sum of joint probabilities (The combined probability of a leakage rate less than 1.2 mL / min); ③ Calculate the probability of a leakage rate less than or equal to 1.2 mL / min. The sum of joint probabilities ④ Substitute into the formula to calculate: [L7,R7]=[0.093+0.0065×0.21×0.33,0.093+0.0065×0.21×0.45]≈[0.0935,0.0938]; ⑤ Complete the coding of highly correlated features (sealing leakage) to ensure the compression accuracy of the core process correlation parameters. Among them, For the first The matrix index corresponding to the actual value of each feature (precursor association feature). For the first The matrix index corresponding to the actual value of each feature For the first Each feature is taken When the value is , the first Each feature takes less than The sum of the joint probabilities of the values For the corresponding less than or equal to The sum of the joint probabilities of the values ​​is logically coherent and verifiable.

[0050] Other cases (non-highly correlated features, in this example) (A total of 27 features), the coding interval is shrunk according to conditional probability, and the formula is: Specific operations (in) Taking the valve core pressing force as an example, including substitution calculations, it can be verified as follows: ① The interval after encoding the first feature is [L1,R1]=[0.0861,0.105], and the interval length is... ② Conditional probability setting: (because and (Non-high correlation, calculated based on conditional probability) Valve core pressing force Corresponding index less than The sum of conditional probabilities ③ Calculate less than or equal to The sum of conditional probabilities ④ Substitute into the formula to calculate: [L2,R2]=[0.0861+0.0189×0.21×0.43,0.0861+0.0189×0.21×0.53]≈[0.0878,0.0882]; ⑤ Complete the encoding of the second feature (valve core pressing force), conforming to the encoding logic of non-highly correlated features, balancing speed and accuracy. The remaining 26 non-highly correlated features ( Following the above logic, the encoding intervals are gradually narrowed, ultimately resulting in the final encoding intervals for 31 normal features: [L] 31 ,R 31 = [0.0989, 0.0991] (The interval length is extremely small, achieving efficient compression and meeting the requirement of a 5:1 compression ratio).

[0051] Finally, for outliers in the independent encoding channels (tightening torque) ) Execute encoding, specific operations (including substitution calculations, conforming to the abnormal parameter characteristics of the main valve, verifiable): ① Initialize the outlier encoding interval: Independent of the normal feature encoding interval, initialize the outlier encoding interval as follows: ① To avoid interfering with normal feature encoding results; ② Correspondence of outlier parameters: outliers Corresponding anomaly probability table index Call the anomaly probability table parameters ③ Calculation of the sum of probabilities: Calculate the sum of probabilities less than... The sum of probabilities ; Calculate less than or equal to The sum of probabilities ④ Substitute into the outlier encoding interval shrinkage formula: Substitute the values: ⑤ Combining compression results: The normal feature encoding range [0.0989, 0.0991] and the outlier encoding range are combined. The data is merged to generate the final compressed result of the 32-dimensional feature vector of the intelligent assembly equipment for the main valve of the brake valve. The data volume is 400 bits, achieving a compression ratio of 5:1 (original 2000 bits → compressed 400 bits). This is suitable for the real-time transmission needs of industrial edge devices. The compressed data can be directly used for storage or remote transmission.

[0052] S4: Decoding and Adaptation Stage. The first step is to obtain the compressed encoding result, including the encoding range of normal features [L]. 31 ,R 31 [0.0989, 0.0991], encoding range for outliers and the corresponding feature vector dimension Coding resource allocation coefficient and semantic association matrix The second step is to initialize the decoding parameters: initialize the decoding range as follows. (Consistent with the initial interval in the encoding stage), initialize empty feature vectors. This is used to store the decoded feature values; simultaneously, it initializes the decoding window. (Window size) ), used for real-time correction of probability prediction; outlier decoding interval initialized to The third step, normal feature decoding (decoding in reverse order of encoding, including core steps, verifiable): ① Call the LLM decoder that is synchronized and fine-tuned with the compression end, and input the decoding range. The decoder generates a cumulative probability table for the first feature (tightening torque), and compares it with the compression end. ① Consistent; ② Calculate the intersection of the decoding interval and each probability interval to determine the first feature decoding value. Add feature vectors Update the decoding window The decoding range was adjusted to [0.0861, 0.105] (similar to the encoding stage). Consistent); ③ Decode subsequent features sequentially according to the encoding order, highly correlated feature pairs ( The joint probability matrix is ​​invoked, and conditional probabilities are invoked for non-highly correlated features, combined with the decoding window. Real-time correction of probability prediction ensures decoding accuracy; ④ After decoding, the decoding results of 31 normal features are obtained, which are completely consistent with the normal feature values ​​before encoding, without deviation; Fourth step, outlier decoding: Calculate the outlier decoding interval. With the anomaly probability table The intersection of each interval is used to determine the outlier decoding result. Insert feature vector The 18th bit (consistent with the outlier position before encoding); Step 5, Deviation verification: Calculate the prediction deviation. ,in Predict probabilities for the LLM decoder. The true probability of the decoded feature value is given in this example. No need to regenerate the local probability table, decoding accuracy meets the standard; Step 6, output the decoding result: obtain the complete 32-dimensional feature vector. It is completely consistent with the original feature vector before encoding, achieving lossless restoration with a restoration accuracy of 100%, and is suitable for the maintenance data review and fault tracing needs of intelligent assembly equipment for brake valve main valve.

[0053] In summary, this embodiment selects the intelligent assembly equipment for the main brake valve (a core piece of equipment in equipment manufacturing) as an implementation example, completing the entire process of 32-dimensional feature vector compression and decoding. The parameters are verifiable, the steps are implementable, and it closely matches actual production scenarios. The verification data is as follows: The 32-dimensional feature vector of the intelligent assembly equipment for the main valve of the brake valve: The original data volume is 2000 bits. After compression using the method of this invention, the data volume is 400 bits (compression ratio 5:1), with encoding time of 7.8ms, decoding time of 8.2ms, and decoding accuracy of 100%; using the traditional arithmetic encoding method, the compressed data volume is 800 bits (compression ratio 2.5:1), with encoding time of 11.5ms, decoding time of 12.3ms, and decoding accuracy of 98.4%; using the general compression algorithm, the compressed data volume is 1200 bits (compression ratio 1.67:1), with encoding time of 15.2ms and decoding accuracy of 88.7%.

[0054] The other parts of this embodiment are the same as any one of the above embodiments 1-2, so they will not be described again.

[0055] Example 4: Based on any one of Embodiments 1-3 above, this embodiment proposes an arithmetic coding system for industrial feature vectors, which is used to execute the above-mentioned arithmetic coding method for industrial feature vectors, including a preprocessing unit, a probability generation unit, a hierarchical coding unit, and a decoding adaptation unit. The preprocessing unit is used to preprocess the high-dimensional feature vector to be compressed obtained from the industrial equipment and calculate the dimensional feature entropy. The probability generation unit is used to input the preprocessed high-dimensional feature vector to be compressed into the fine-tuned LLM model to obtain a dynamic probability matrix. The hierarchical coding unit is used to initialize the coding interval, determine the probability interval according to the dynamic probability matrix, and gradually shrink the coding interval through the probability distribution to obtain the compressed coding result; The decoding adaptation unit is used to call the decoder to maintain the probability update window, and correct the probability prediction in real time by using the decoded feature values ​​in the sliding window. If the prediction deviation is greater than the set deviation threshold, the local probability table is regenerated.

[0056] This embodiment also proposes an electronic device, including a memory and a processor; the memory stores a computer program; when the computer program is executed on the processor, it implements the above-described arithmetic coding method for industrial feature vectors.

[0057] This embodiment also proposes a computer-readable storage medium storing computer instructions; when the computer instructions are executed on the processor, the above-described arithmetic coding method for industrial feature vectors is implemented.

[0058] The other parts of this embodiment are the same as any one of the embodiments 1-3 above, so they will not be described again.

[0059] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications or equivalent changes made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. An arithmetic coding method for industrial feature vectors, characterized in that, Specifically, the following steps are included: Step S1: Preprocess the high-dimensional feature vector to be compressed obtained from the industrial equipment and calculate the dimensional feature entropy; Step S2: Input the preprocessed high-dimensional feature vector to be compressed into the fine-tuned LLM model to obtain the dynamic probability matrix; Step S3: Initialize the encoding interval, determine the probability interval based on the dynamic probability matrix, and gradually shrink the encoding interval through the probability distribution to obtain the compressed encoding result; Step S4: Call the decoder to maintain the probability update window, and correct the probability prediction in real time by sliding the decoded feature values ​​in the window. If the prediction deviation is greater than the set deviation threshold, the local probability table will be regenerated.

2. The arithmetic coding method for industrial feature vectors according to claim 1, characterized in that, Step S1 specifically includes the following steps: Step S11: Obtain the high-dimensional feature vector to be compressed output by the industrial equipment; Step S12: Calculate the mean and standard deviation of the high-dimensional feature vector to be compressed based on its feature vector dimension; Step S13: Remove abnormal feature vectors based on the mean and standard deviation of the high-dimensional feature vector to be compressed to obtain normal feature vectors, and calculate the dimensional feature entropy of the normal feature vectors.

3. The arithmetic coding method for industrial feature vectors according to claim 1, characterized in that, Step S2 specifically includes the following steps: Step S21: Call LoRA to fine-tune the large language model LLM; Step S22: Input the normal feature vector into the fine-tuned large language model LLM to obtain the dynamic probability matrix and semantic association matrix; Step S23: If the semantic relevance matrix is ​​greater than or equal to a set threshold, generate a joint probability matrix; Step S24: Dynamically adjust the probability prediction accuracy of the large language model LLM based on the dimensional feature entropy; Step S25: Generate an anomaly probability table based on the anomaly feature vector.

4. The arithmetic coding method for industrial feature vectors according to claim 1, characterized in that, Step S3 specifically includes the following steps: Step S31: Initialize the encoding interval and encoding resource allocation coefficient; Step S32: Execute encoding according to the set hierarchical encoding rules, and adjust the encoding interval shrinkage precision according to the initialized encoding resource allocation coefficient to obtain the encoding result of normal features; Step S33: Encode the outliers in the independent encoding channels by calling the outlier probability table to obtain the encoding results of the outlier features; Step S34: Merge the encoding results of normal features and the encoding results of abnormal features to obtain the compressed result.

5. The arithmetic coding method for industrial feature vectors according to claim 4, characterized in that, The hierarchical encoding rule in step S32 is as follows: If it is the first feature, that is Directly call the dynamic probability matrix Determine the probability interval, and combine it with the coding resource allocation coefficient. Adjusting the contraction amplitude, the formula for updating the encoding interval is: ； in, For current features The actual value, Less than The sum of the probabilities of all possible values. less than or equal to The sum of the probabilities of all possible values; If they are highly correlated feature pairs, that is And exist make Call the joint probability matrix Shrink the coding interval, combine Adjust the precision using the following formula: ； in, For the first The actual values ​​of each feature For the first The actual values ​​of each feature For the first Each feature is taken When the value is , the first Each feature takes less than The sum of the joint probabilities of the values For the corresponding less than or equal to The sum of the joint probabilities of the values; If the feature is not highly correlated, then conditional probability applies. Shrink the coding interval, combine Adjust the precision using the following formula: ;in, and They are respectively the codes of the first The left and right boundaries of the encoding interval after each feature.

6. The arithmetic coding method for industrial feature vectors according to claim 5, characterized in that, The specific operation of step S33 is as follows: For outliers in the independent encoded channels, call the outlier probability table. Execute encoding, and initialize the encoding range to... The contraction formula is: ,in This represents the actual value of the outlier.

7. The arithmetic coding method for industrial feature vectors according to claim 1, characterized in that, Step S4 specifically includes the following steps: Step S41: Obtain the compressed encoding result and initialize the decoding interval, empty feature vector, and outlier decoding interval; Step S42: Call the LLM decoder that is synchronized and fine-tuned with the compression end to reconstruct the complete normal feature vector; Step S43: Calculate the intersection of the outlier decoding interval and the probability interval of each outlier in the outlier probability table, determine the outlier, and insert the corresponding position of the feature vector of the value null. Step S44: Calculate the prediction bias based on the current feature value probability predicted by the LLM decoder and the true probability of the decoded feature value; Step S45: If the prediction deviation is greater than 0.001, the local probability table is regenerated, and the LLM decoder is called again to generate the probability table for the current position and subsequent positions to correct the decoding deviation. Step S46: After decoding is complete, output the reconstructed feature vector.

8. An arithmetic coding system for industrial feature vectors, used to execute the arithmetic coding method for industrial feature vectors as described in claim 1, characterized in that, It includes a preprocessing unit, a probability generation unit, a hierarchical coding unit, and a decoding adaptation unit; The preprocessing unit is used to preprocess the high-dimensional feature vector to be compressed obtained from the industrial equipment and calculate the dimensional feature entropy. The probability generation unit is used to input the preprocessed high-dimensional feature vector to be compressed into the fine-tuned LLM model to obtain a dynamic probability matrix. The hierarchical coding unit is used to initialize the coding interval, determine the probability interval according to the dynamic probability matrix, and gradually shrink the coding interval through the probability distribution to obtain the compressed coding result; The decoding adaptation unit is used to call the decoder to maintain the probability update window, and correct the probability prediction in real time by using the decoded feature values ​​in the sliding window. If the prediction deviation is greater than the set deviation threshold, the local probability table is regenerated.

9. An electronic device, characterized in that, It includes a memory and a processor; the memory stores a computer program; when the computer program is executed on the processor, it implements an arithmetic encoding method for industrial feature vectors as described in claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions; when the computer instructions are executed on the processor, they implement an arithmetic encoding method for industrial feature vectors as described in claims 1-7.

Images