A Belt Scale Weighing Data Management Method and System Based on Deep Learning

Abstract

The present invention provides a method and system for managing weighing data of belt scales based on deep learning in the field of industrial big data and artificial intelligence technology. The method includes: Step S1, obtaining historical monitoring data and historical weighing data of a large number of belt scales to construct a data set; Step S2, training and deploying a weighing error compensation model through the data set; Step S3, collecting real-time weighing data and real-time monitoring data of the belt scale, inputting the real-time monitoring data into the deployed weighing error compensation model to obtain a weighing error compensation value, and correcting the real-time weighing data based on the weighing error compensation value to obtain corrected weighing data; Step S4, encrypting the corrected weighing data into first encrypted weighing data for local storage, and encrypting the corrected weighing data into second encrypted weighing data for uploading to the server. The advantages of the present invention are: greatly improving the accuracy and continuity of belt scale weighing data measurement, the security of transmission and storage, and the convenience of management.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial big data and artificial intelligence, and particularly to a belt scale weighing data management method and system based on deep learning. Background Art

[0002] A belt scale is a device used to continuously and dynamically measure the weight and flow rate of bulk materials during the transmission process on a belt conveyor. By integrating the flow rate with respect to time, the cumulative weight can be obtained, that is, the weighing data of the belt scale includes weight and flow rate. The belt scale mainly consists of a weighing sensor, a speed sensor, a weighing instrument, a mechanical scale frame, etc. The working principle of measuring weight is as follows: The mechanical scale frame is installed on the conveyor frame. When the material passes through, the metering idler detects the weight of the material on the belt, and through the lever action, it acts on the weighing sensor, generating a voltage signal proportional to the belt load. The weighing sensor converts the weight signal of the material into an electrical signal, which is amplified by an amplifier and then sent to an analog-to-digital converter to become a digital quantity. The working principle of measuring flow rate is as follows: The speed sensor is installed on the tail pulley or the driving roller of the belt to measure the running speed of the belt and convert the speed signal into a pulse signal. Each pulse represents a belt movement unit, and the pulse frequency is proportional to the belt speed.

[0003] Belt scales are widely used in many industries such as mining, chemical industry, electric power, building materials, ports, logistics, etc. For example, in the coal industry, it is used to accurately measure coal during coal mining, washing, and transportation processes to help control coal production and quality; in the electric power industry, it is used to measure fuels such as coal and limestone to ensure accurate fuel supply during the power generation process; in the chemical industry, it is used to measure chemical raw materials and products to ensure accurate material ratio during the production process.

[0004] Affected by the coupling effect of multiple physical fields, the belt scale faces complex error sources during actual operation, mainly including: 1. Mechanical system errors: deformation of the scale frame structure (up to 0.1 mm / m), radial runout of the idler, belt tension fluctuation (causing a speed measurement error of ±0.3%); 2. Sensing and detection errors: creep of the weighing sensor (typical value ±0.02%FS / 10 min), temperature drift (±0.0015%FS / °C), slip of the speed sensor (up to 0.2%); 3. Environmental interference: adhesion error of ±1.5% caused by changes in material humidity, signal baseline drift caused by electromagnetic interference (up to 500 μV); 4. Human operation errors: deviation in off-center calibration (introducing an error of ±0.25% for every 10% off-center load), and untimely zero drift compensation (cumulative error of about 0.1% per day).

[0005] To improve the accuracy of belt scale measurement, belt scales must be calibrated to reduce measurement errors. Traditionally, this method has been primarily used, but it has significant limitations: its calibration accuracy is affected by chain scale deflection (maximum deviation 0.15%), and it cannot simulate actual material distribution. In a cement plant application, after two weeks of operation, the cumulative error of a belt scale calibrated with a chain scale increased from ±0.5% to ±1.2%. Furthermore, the current eight-hour repeatability test method specified in the GB / T 7721 standard is unable to effectively capture dynamic interference factors such as roller radial runout. This results in the plant having to shut down two to three times per month for manual calibration, with each calibration taking approximately four hours and severely impacting continuous production operations.

[0006] Furthermore, traditionally, PLCs transmit weighing data collected from belt scales via the Modbus protocol and store it in a local database. This data is encrypted using a single key. However, the Modbus protocol was designed primarily for functionality and efficiency, without adequate consideration for security. This single key is easily stolen and cracked, making it vulnerable to plaintext theft and tampering of weighing data. This poses a significant security risk and prevents online management of weighing data, impacting its timely management. For example, an audit of a steel company revealed that unencrypted weighing data had been maliciously tampered with during transmission, resulting in a difference of 370 tons of steel in annual settlements.

[0007] Therefore, how to provide a belt scale weighing data management method and system based on deep learning to improve the accuracy and continuity of belt scale weighing data measurement, the security of transmission and storage, and the convenience of management has become a technical problem that needs to be solved urgently. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to provide a belt scale weighing data management method and system based on deep learning, so as to improve the accuracy and continuity of belt scale weighing data measurement, the security of transmission and storage, and the convenience of management.

[0009] In a first aspect, the present invention provides a belt scale weighing data management method based on deep learning, comprising the following steps:

[0010] Step S1: creating a weighing error compensation model based on the sequentially connected multimodal feature extraction layer, spatiotemporal fusion layer, and dynamic compensation output layer, and setting a loss function of the weighing error compensation model;

[0011] The multimodal feature extraction layer is used to extract fusion features from the input monitoring data; the spatiotemporal fusion layer is used to perform spatiotemporal correlation modeling on the fusion features to obtain compensation decision features; the dynamic compensation output layer is used to map the compensation decision features into weighing error compensation values and output them;

[0012] Step S2: Obtain a large amount of historical monitoring data and historical weighing data, preprocess each piece of the historical monitoring data and historical weighing data, and then construct a data set; the historical monitoring data includes the deformation amount of the scale frame structure, the radial runout distance of the idler, the belt tension, the creep amount of the weighing sensor, the slip of the speed sensor, the material humidity, the ambient temperature, the electromagnetic interference intensity, and the off-loading degree;

[0013] Step S3: Divide the data set into a training set, a validation set, and a test set, train the weighing error compensation model through the training set and a loss function, and then verify and test the trained weighing error compensation model through the validation set and the test set;

[0014] Step S4: Compress, perform drift compensation training, and deploy the weighing error compensation model that passes the test;

[0015] Step S5: Collect the real-time weighing data and real-time monitoring data of the belt scale, preprocess the real-time weighing data and real-time monitoring data, input the real-time monitoring data into the deployed weighing error compensation model to obtain a weighing error compensation value, and correct the real-time weighing data based on the weighing error compensation value to obtain corrected weighing data;

[0016] Step S6: Real-time display the corrected weighing data through a pre-associated display screen, record in real time a weighing log including at least the corrected weighing data, the real-time weighing data, and the real-time monitoring data, and based on a preset storage period, encrypt the weighing log into a first encrypted log according to a first encryption rule for local storage, and encrypt the weighing log into a second encrypted log according to a second encryption rule and upload it to the server through the TLS protocol;

[0017] Step S7: The server creates a weighing database, sets an authentication mechanism for the weighing database, decrypts and verifies the received second encrypted log to obtain a weighing log, calculates the data fingerprint of the weighing log, stores the data fingerprint in the blockchain, encrypts the weighing log into a third encrypted log according to a third encryption rule, and stores the third encrypted log in the weighing database;

[0018] Step S8: The server performs anomaly monitoring on the corrected weighing data in the weighing log, and the mobile terminal accesses the weighing database through the authentication mechanism to perform online management of the weighing log;

[0019] Step S9: Continuously optimize and iterate the weighing error compensation model based on the weighing log;

[0020] In the step S1, the multi-modal feature extraction layer is constructed based on a structural deformation encoding module, a dynamic load encoding module, an environmental interference encoding module, and a feature output module;

[0021] The structural deformation encoding module is used to extract offset features and connection stiffness features from the scale frame structural deformation amount, idler radial runout distance, and eccentricity through a graph neural network; the dynamic load encoding module is used to extract spatial local features from the belt tension and load cell creep amount through a one-dimensional convolutional neural network, and extract temporal dependence features from the speed sensor slip through a bidirectional long short-term memory network; the environmental interference encoding module is used to extract robust features from the material humidity, environmental temperature, and electromagnetic interference intensity through an adversarial noise robust network; the feature output module is used to fuse the offset features, connection stiffness features, spatial local features, temporal dependence features, and robust features through a cross-modal attention mechanism and feature channel splicing to obtain fused features;

[0022] The spatio-temporal fusion layer is constructed based on a temporal feature correlation module, a spatial feature correlation module, a feature interaction module, and a fusion module; the temporal feature correlation module is used to mine temporal coupling features from the fused features through a dilated causal convolutional network; the spatial feature correlation module is used to mine spatial propagation features from the fused features through a graph attention network; the feature interaction module is used to mine non-linear compensation features from the fused features through a cross-modal cross-attention mechanism; the fusion module is used to perform weighted fusion on the mined temporal coupling features, spatial propagation features, and non-linear compensation features to obtain compensated decision features.

[0023] Further, in the step S1, the formula of the loss function is: ;

[0024] where represents the loss value of the loss function; represents the error compensation precision loss sub-function, and the mean square error function is adopted; represents the first regularization loss sub-function of the multi-modal feature extraction layer, and the L2 regularization function is adopted; represents the second regularization loss sub-function of the spatio-temporal fusion layer, and the L2 regularization function is adopted; represents the third regularization loss sub-function of the dynamic compensation output layer, and the L2 regularization function is adopted; , and all represent regularization strength coefficients;

[0025] The step S2 is specifically as follows:

[0026] Obtain historical monitoring data and historical weighing data of a large number of belt scales, including the structural deformation amount of the scale frame, the radial runout distance of the idler, the belt tension, the creep amount of the weighing sensor, the slip of the speed sensor, the material humidity, the ambient temperature, the electromagnetic interference intensity, and the offloading degree; the historical weighing data includes the instantaneous weight, the instantaneous flow rate, and the total weight;

[0027] Locate outliers from each of the historical monitoring data and historical weighing data through the Z-score method, repair each of the outliers through a pre-trained regression model, fill in the missing values in the historical monitoring data and historical weighing data through the interpolation method, and then label each of the historical monitoring data with the historical weighing data and the weighing error compensation value to complete the preprocessing, and construct a data set based on each of the preprocessed historical monitoring data and historical weighing data;

[0028] The weighing error compensation value includes an instantaneous weight compensation value, an instantaneous flow rate compensation value, and a total weight compensation value.

[0029] Further, the specific steps of step S3 are as follows:

[0030] Divide the data set into a training set, a validation set, and a test set according to a ratio of 6:2:2 through the stratified random sampling method, and call the training set through the adaptive moment estimation optimizer to train the weighing error compensation model. During the training process, dynamically adjust the learning rate until the loss value of the loss function decreases by less than 1e-5 in three consecutive training cycles;

[0031] Calculate the F1 score and the coefficient of determination through the validation set to verify the trained weighing error compensation model, and synchronously optimize the hyperparameters and model structure parameters of the weighing error compensation model;

[0032] Calculate the mean absolute percentage error and the Pearson correlation coefficient through the test set to test the verified weighing error compensation model;

[0033] The specific steps of step S4 are as follows:

[0034] Perform primary compression on the tested weighing error compensation model through the hybrid quantization technique, remove redundant neurons through the structured pruning method to perform secondary compression on the primarily compressed weighing error compensation model, and perform tertiary compression on the secondarily compressed weighing error compensation model through the knowledge distillation technique;

[0035] Obtain the actual monitoring data of the belt scale to be compensated, perform drift compensation training on the tertiary compressed weighing error compensation model through the actual monitoring data, and then deploy the drift compensation trained weighing error compensation model through the containerization technique.

[0036] Furthermore, the step S5 is specifically as follows:

[0037] Real-time weighing data and real-time monitoring data of the belt scale are collected in real time through a multi-source sensor array. The real-time weighing data and real-time monitoring data are pre-processed in sequence through a streaming computing engine using sliding window filtering, outlier removal, and wavelet noise reduction. The pre-processed real-time monitoring data is input into the deployed weighing error compensation model to obtain a weighing error compensation value. The weighing error compensation value is fused into the real-time weighing data through a Kalman filter algorithm for correction, and the corrected weighing data carrying the weighing time is synchronously output;

[0038] In step S6, the first encryption rule is specifically:

[0039] Performing UTF-8 encoding conversion on the weighing log to generate a binary data stream, calculating the integrity check code of the binary data stream using the SHA-3-512 algorithm, intercepting the first 128 bits from the integrity check code as a check segment, dividing the check segment into four groups of 32-bit data blocks, and inserting them into the 1 / 4, 1 / 2, 3 / 4, and end positions of the binary data stream, respectively, to obtain an obfuscated binary data stream, performing a byte reversal operation on the obfuscated binary data stream to obtain a layer of encrypted data, encrypting the first layer of encrypted data into a second layer of encrypted data using the RC6 algorithm, converting the second layer of encrypted data into a Base85 encoded string, adding 3 to the ASCII values of vowels in the Base85 encoded string, cyclically shifting the consonants right by 5 bits, and replacing the numeric characters with a modulo 7 operation to obtain a third layer of encrypted data, and encrypting the third layer of decrypted data into a first encrypted log using the SM4 algorithm;

[0040] The second encryption rule is specifically:

[0041] Set a validity period, obtain the current timestamp, calculate the effective expiration timestamp based on the validity period and the current timestamp, create a master key based on the AES256 algorithm, and calculate the hash value of the weighing log using the SHA-3 algorithm;

[0042] Concatenate the master key with the date portion of the current timestamp to obtain concatenated data, and calculate the concatenated data using the SHA-3 algorithm to obtain a primary dynamic key;

[0043] XORing the primary dynamic key with the hour part of the current timestamp to obtain a first XOR data, and calculating the first XOR data using the SHA-256 algorithm to obtain a secondary dynamic key;

[0044] XOR the secondary dynamic key with the valid expiration timestamp to obtain second XOR data, and calculate a tertiary dynamic key from the second XOR data using the HMAC-SM3 algorithm;

[0045] Encrypt the weighing log using the AES-GCM algorithm with the tertiary dynamic key to obtain first-level encrypted data; combine the first-level encrypted data, hash value, and valid expiration timestamp into combined data, encrypt the combined data into second-level encrypted data using the RSA-OAEP algorithm, encrypt the second-level encrypted data into third-level encrypted data using the ECC algorithm, sign the third-level encrypted data, valid expiration timestamp, and master key using the EdDSA algorithm to obtain a first signature value, and encapsulate the third-level encrypted data, valid expiration timestamp, master key, and first signature value into a second encrypted log.

[0046] Further, step S7 is specifically as follows:

[0047] The server creates a weighing database for storing the third encrypted log and sets the authentication mechanism of the weighing database;

[0048] The server receives the second encrypted log in real time, parses the second encrypted log to obtain third-level encrypted data, valid expiration timestamp, master key, and first signature value; sign the third-level encrypted data, valid expiration timestamp, and master key using the EdDSA algorithm to obtain a second signature value, perform a signature verification operation on the first signature value using the second signature value, and then perform a timeliness verification using the valid expiration timestamp;

[0049] Decrypt the third-level encrypted data into second-level encrypted data using the ECC algorithm, decrypt the second-level encrypted data into combined data using the RSA-OAEP algorithm, parse the combined data to obtain first-level encrypted data, hash value, and valid expiration timestamp, and calculate the current timestamp based on the valid expiration timestamp and validity period;

[0050] Concatenate the master key with the date part of the current timestamp to obtain concatenated data, calculate a first-level dynamic key from the concatenated data using the SHA-3 algorithm; XOR the first-level dynamic key with the hour part of the current timestamp to obtain first XOR data, calculate a second-level dynamic key from the first XOR data using the SHA-256 algorithm; XOR the second-level dynamic key with the valid expiration timestamp to obtain second XOR data, and calculate a tertiary dynamic key from the second XOR data using the HMAC-SM3 algorithm;

[0051] Decrypt the first-level encrypted data using the AES-GCM algorithm with the tertiary dynamic key to obtain the weighing log, and perform an integrity check on the weighing log using the hash value;

[0052] The server calculates the data fingerprint of the weighing log through the HMAC algorithm, stores the data fingerprint in the blockchain, encrypts the weighing log into a third encrypted log based on the third encryption rule, and stores the third encrypted log in a structured manner in the weighing database;

[0053] The specific steps of step S8 are as follows:

[0054] The server performs anomaly monitoring and early warning on the corrected weighing data in the weighing log through the Isolation Forest algorithm. The mobile terminal accesses the weighing database through the authentication mechanism. The server decrypts the third encrypted log in the weighing database through the third encryption rule to obtain the weighing log, so as to perform online management including addition, deletion, modification, and query on the weighing log.

[0055] In a second aspect, the present invention provides a belt scale weighing data management system based on deep learning, including the following modules:

[0056] A weighing error compensation model creation module, which is used to create a weighing error compensation model based on a multi-modal feature extraction layer, a spatio-temporal fusion layer, and a dynamic compensation output layer connected in sequence, and set the loss function of the weighing error compensation model;

[0057] The multi-modal feature extraction layer is used to extract fusion features from the input monitoring data; the spatio-temporal fusion layer is used to perform spatio-temporal correlation modeling on the fusion features to obtain compensation decision features; the dynamic compensation output layer is used to map the compensation decision features into weighing error compensation values and output them;

[0058] A data set construction module, which is used to obtain a large amount of historical monitoring data and historical weighing data, preprocess each piece of historical monitoring data and historical weighing data, and construct a data set; the historical monitoring data includes the deformation amount of the scale frame structure, the radial runout distance of the idler, the belt tension, the creep amount of the weighing sensor, the slip of the speed sensor, the material humidity, the ambient temperature, the electromagnetic interference intensity, and the offloading degree;

[0059] A weighing error compensation model training module, which is used to divide the data set into a training set, a validation set, and a test set, train the weighing error compensation model through the training set and the loss function, and then verify and test the trained weighing error compensation model through the validation set and the test set;

[0060] A weighing error compensation model deployment module, which is used to compress, perform drift compensation training, and deploy the weighing error compensation model that passes the test;

[0061] The weighing data compensation module is used to collect the real-time weighing data and real-time monitoring data of the belt scale. After preprocessing the real-time weighing data and real-time monitoring data, the real-time monitoring data is input into the deployed weighing error compensation model to obtain a weighing error compensation value, and the real-time weighing data is corrected based on the weighing error compensation value to obtain corrected weighing data;

[0062] The weighing log storage and backup module is used to display the corrected weighing data in real time through a pre-associated display screen, record the weighing log including at least the corrected weighing data, real-time weighing data and real-time monitoring data in real time, and encrypt the weighing log into a first encrypted log based on a first encryption rule for local storage at a preset storage period, and encrypt the weighing log into a second encrypted log based on a second encryption rule and upload it to the server through the TLS protocol;

[0063] The weighing log certification module is used for the server to create a weighing database, set the authentication mechanism of the weighing database, decrypt and verify the received second encrypted log to obtain the weighing log, calculate the data fingerprint of the weighing log, store the data fingerprint in the blockchain, encrypt the weighing log into a third encrypted log based on a third encryption rule, and store the third encrypted log in the weighing database;

[0064] The weighing log management module is used for the server to perform anomaly monitoring on the corrected weighing data in the weighing log, and the mobile terminal accesses the weighing database through the authentication mechanism to perform online management of the weighing log;

[0065] The weighing error compensation model optimization module is used to continuously optimize and iterate the weighing error compensation model based on the weighing log;

[0066] In the weighing error compensation model creation module, the multi-modal feature extraction layer is constructed based on a structure deformation coding module, a dynamic load coding module, an environmental interference coding module, and a feature output module;

[0067] The structure deformation coding module is used to extract offset features and connection stiffness features from the scale frame structure deformation amount, idler radial runout distance, and eccentricity through a graph neural network; the dynamic load coding module is used to extract spatial local features from the belt tension and weighing sensor creep amount through a one-dimensional convolutional neural network, and extract temporal dependence features from the speed sensor slip through a bidirectional long short-term memory network; the environmental interference coding module is used to extract robust features from the material humidity, environmental temperature, and electromagnetic interference intensity through an adversarial noise robust network; the feature output module is used to fuse the offset features, connection stiffness features, spatial local features, temporal dependence features, and robust features through a cross-modal attention mechanism and feature channel splicing to obtain fused features;

[0068] The spatio-temporal fusion layer is constructed based on a temporal feature correlation module, a spatial feature correlation module, a feature interaction module, and a fusion module; the temporal feature correlation module is used to mine temporal coupling features from the fusion features through a dilated causal convolutional network; the spatial feature correlation module is used to mine spatial propagation features from the fusion features through a graph attention network; the feature interaction module is used to mine non-linear compensation features from the fusion features through a cross-modal cross-attention mechanism; the fusion module is used to perform weighted fusion on the mined temporal coupling features, spatial propagation features, and non-linear compensation features to obtain compensated decision features.

[0069] Further, in the weighing error compensation model creation module, the formula of the loss function is: ;

[0070] where represents the loss value of the loss function; represents the error compensation precision loss sub-function, and the mean square error function is adopted; represents the first regularization loss sub-function of the multi-modal feature extraction layer, and the L2 regularization function is adopted; represents the second regularization loss sub-function of the spatio-temporal fusion layer, and the L2 regularization function is adopted; represents the third regularization loss sub-function of the dynamic compensation output layer, and the L2 regularization function is adopted; , and both represent the regularization strength coefficients;

[0071] The dataset construction module is specifically used for:

[0072] Obtain a large amount of historical monitoring data and historical weighing data of belt scales, including the structural deformation amount of the scale frame, the radial runout distance of the idler, the belt tension, the creep amount of the weighing sensor, the slip of the speed sensor, the material humidity, the ambient temperature, the electromagnetic interference intensity, and the offloading degree; the historical weighing data includes the instantaneous weight, the instantaneous flow rate, and the total weight;

[0073] Locate outliers from each of the historical monitoring data and historical weighing data through the Z-score method, repair each of the outliers through a pre-trained regression model, fill in the missing values in the historical monitoring data and historical weighing data through the interpolation method, and then label each of the historical monitoring data with the historical weighing data and the weighing error compensation value to complete the preprocessing, and construct a dataset based on each of the preprocessed historical monitoring data and historical weighing data;

[0074] The weighing error compensation value includes an instantaneous weight compensation value, an instantaneous flow rate compensation value, and a total weight compensation value.

[0075] Further, the weighing error compensation model training module is specifically configured to:

[0076] Divide the data set into a training set, a validation set, and a test set according to a ratio of 6:2:2 by the stratified random sampling method, and call the training set through the adaptive moment estimation optimizer to train the weighing error compensation model. During the training process, dynamically adjust the learning rate until the loss value of the loss function decreases by less than 1e-5 in three consecutive training cycles;

[0077] Calculate the F1 score and the coefficient of determination through the validation set to verify the trained weighing error compensation model, and synchronously tune the hyperparameters and model structure parameters of the weighing error compensation model;

[0078] Calculate the mean absolute percentage error and the Pearson correlation coefficient through the test set to test the verified weighing error compensation model;

[0079] The weighing error compensation model deployment module is specifically configured to:

[0080] Perform primary compression on the tested weighing error compensation model through the hybrid quantization technology, perform secondary compression on the primarily compressed weighing error compensation model by removing redundant neurons through the structured pruning method, and perform tertiary compression on the secondarily compressed weighing error compensation model through the knowledge distillation technology;

[0081] Obtain the actual monitoring data of the belt scale to be compensated, perform drift compensation training on the tertiarily compressed weighing error compensation model through the actual monitoring data, and then deploy the drift compensation trained weighing error compensation model through the containerization technology.

[0082] Further, the weighing data compensation module is specifically configured to:

[0083] Real-time collect the real-time weighing data and real-time monitoring data of the belt scale through the multi-source sensor array, and use the streaming computing engine to preprocess the real-time weighing data and real-time monitoring data in turn by using sliding window filtering, outlier removal, and wavelet denoising. Input the preprocessed real-time monitoring data into the deployed weighing error compensation model to obtain the weighing error compensation value, and fuse the weighing error compensation value into the real-time weighing data through the Kalman filter algorithm for correction, and synchronously output the corrected weighing data carrying the weighing time;

[0084] In the weighing log storage and backup module, the first encryption rule is specifically:

[0085] Convert the weighing log to UTF-8 encoding to generate a binary data stream, calculate the integrity check code of the binary data stream through the SHA-3-512 algorithm, intercept the first 128 bits from the integrity check code as the check segment, divide the check segment into four groups of 32-bit data blocks, and insert them into the 1 / 4, 1 / 2, 3 / 4, and end positions of the binary data stream respectively to obtain a confused binary data stream. Perform a byte reversal operation on the confused binary data stream to obtain a first-level encrypted data, encrypt the first-level encrypted data into a second-level encrypted data through the RC6 algorithm, convert the second-level encrypted data into a Base85 encoded string, add 3 to the ASCII value of the vowel letters in the Base85 encoded string, circularly shift the consonant letters 5 bits to the right, and replace the numeric characters with modulo 7 operations to obtain a third-level encrypted data. Encrypt the third-level decrypted data into the first encrypted log through the SM4 algorithm;

[0086] The specific second encryption rule is as follows:

[0087] Set a validity period, obtain the current timestamp, calculate the effective expiration timestamp based on the validity period and the current timestamp, create a master key based on the AES256 algorithm, and calculate the hash value of the weighing log through the SHA-3 algorithm;

[0088] Concatenate the master key with the date part of the current timestamp to obtain concatenated data, and calculate a first-level dynamic key by calculating the concatenated data through the SHA-3 algorithm;

[0089] XOR the first-level dynamic key with the hour part of the current timestamp to obtain a first XOR data, and calculate a second-level dynamic key by calculating the first XOR data through the SHA-256 algorithm;

[0090] XOR the second-level dynamic key with the effective expiration timestamp to obtain a second XOR data, and calculate a third-level dynamic key by calculating the second XOR data through the HMAC-SM3 algorithm;

[0091] Encrypt the weighing log by calling the third-level dynamic key through the AES-GCM algorithm to obtain a first-level encrypted data; combine the first-level encrypted data, the hash value, and the effective expiration timestamp into combined data, encrypt the combined data into a second-level encrypted data through the RSA-OAEP algorithm, encrypt the second-level encrypted data into a third-level encrypted data through the ECC algorithm, sign the third-level encrypted data, the effective expiration timestamp, and the master key through the EdDSA algorithm to obtain a first signature value, and encapsulate the third-level encrypted data, the effective expiration timestamp, the master key, and the first signature value into a second encrypted log.

[0092] Furthermore, the weighing log evidence storage module is specifically used for:

[0093] The server creates a weighing database for storing the third encrypted log and sets the authentication mechanism of the weighing database;

[0094] The server receives the second encrypted log in real time, parses the second encrypted log to obtain the three-level encrypted data, the effective expiration timestamp, the master key, and the first signature value; signs the three-level encrypted data, the effective expiration timestamp, and the master key through the EdDSA algorithm to obtain the second signature value, performs a signature verification operation on the first signature value through the second signature value, and then performs a timeliness verification through the effective expiration timestamp;

[0095] Decrypt the three-level encrypted data into two-level encrypted data through the ECC algorithm, decrypt the two-level encrypted data into combined data through the RSA-OAEP algorithm, parse the combined data to obtain the first-level encrypted data, the hash value, and the effective expiration timestamp, and calculate the current timestamp based on the effective expiration timestamp and the validity period;

[0096] Concatenate the master key with the date part of the current timestamp to obtain concatenated data, calculate the first-level dynamic key by passing the concatenated data through the SHA-3 algorithm; perform an exclusive OR operation on the first-level dynamic key and the hour part of the current timestamp to obtain the first exclusive OR data, calculate the second-level dynamic key by passing the first exclusive OR data through the SHA-256 algorithm; perform an exclusive OR operation on the second-level dynamic key and the effective expiration timestamp to obtain the second exclusive OR data, calculate the third-level dynamic key by passing the second exclusive OR data through the HMAC-SM3 algorithm;

[0097] Decrypt the first-level encrypted data into the weighing log by invoking the third-level dynamic key through the AES-GCM algorithm, and perform an integrity check on the weighing log through the hash value;

[0098] The server calculates the data fingerprint of the weighing log through the HMAC algorithm, stores the data fingerprint in the blockchain, encrypts the weighing log into the third encrypted log based on the third encryption rule, and stores the third encrypted log in a structured manner in the weighing database;

[0099] The weighing log management module is specifically used for:

[0100] The server performs anomaly monitoring and warning on the corrected weighing data in the weighing log through the Isolation Forest algorithm. The mobile terminal accesses the weighing database through the authentication mechanism. The server decrypts the third encrypted log in the weighing database through the third encryption rule to obtain the weighing log for online management including addition, deletion, modification, and query of the weighing log.

[0101] The advantages of the present invention are:

[0102] 1. Create a weighing error compensation model by sequentially connecting a multi-modal feature extraction layer, a spatio-temporal fusion layer, and a dynamic compensation output layer, and set the loss function of the weighing error compensation model; then obtain a large amount of historical monitoring data and historical weighing data of belt scales, including the deformation amount of the scale frame structure, the radial runout distance of the idler, the belt tension, the creep amount of the weighing sensor, the slip of the speed sensor, the material humidity, the ambient temperature, the electromagnetic interference intensity, and the offloading degree, to construct a data set. Divide the data set into a training set, a validation set, and a test set. Train the weighing error compensation model through the training set and the loss function, and then verify and test the trained weighing error compensation model through the validation set and the test set. Compress, perform drift compensation training, and deploy the weighing error compensation model that passes the test; then collect the real-time weighing data and real-time monitoring data of the belt scale. After preprocessing the real-time weighing data and real-time monitoring data, input the real-time monitoring data into the deployed weighing error compensation model to obtain a weighing error compensation value, and correct the real-time weighing data based on the weighing error compensation value to obtain corrected weighing data; then display the corrected weighing data in real time through a display screen, and record in real time a weighing log that includes at least the corrected weighing data, the real-time weighing data, and the real-time monitoring data. Based on a preset storage period, encrypt the weighing log into a first encrypted log according to the first encryption rule for local storage, encrypt the weighing log into a second encrypted log according to the second encryption rule and upload it to the server through the TLS protocol; the server creates a weighing database, sets an authentication mechanism for the weighing database, decrypts and verifies the second encrypted log to obtain the weighing log, calculates the data fingerprint of the weighing log and stores it in the blockchain, encrypts the weighing log into a third encrypted log according to the third encryption rule, stores the third encrypted log in the weighing database, monitors the corrected weighing data for anomalies, and the mobile terminal accesses the weighing database through the authentication mechanism to perform online management of the weighing log, and continuously optimizes and iterates the weighing error compensation model based on the weighing log; that is, perform error compensation through the weighing error compensation model trained by the data set constructed by the deformation amount of the scale frame structure, the radial runout distance of the idler, the belt tension, the creep amount of the weighing sensor, the slip of the speed sensor, the material humidity, the ambient temperature, the electromagnetic interference intensity, and the offloading degree, fully considering multi-dimensional influencing factors, and perform real-time compensation based on the collected real-time monitoring data, without the need for traditional shutdown calibration. Combining the encryption of the first encryption rule, the second encryption rule, and the third encryption rule, and uploading the data fingerprint to the blockchain can effectively prevent relevant data from being stolen and tampered with in plain text. Combining the set authentication mechanism, the weighing database can be accessed online to manage the weighing log, ultimately greatly improving the accuracy and continuity of belt scale weighing data measurement, the security of transmission and storage, and the convenience of management.

[0103] 2. By adopting a multi-modal combination of graph neural network (structural deformation encoding), 1D CNN + BiLSTM (dynamic load encoding), and adversarial noise-robust network (environmental interference encoding), joint modeling of multi-dimensional interference factors such as mechanical deformation, dynamic load, and environmental interference is achieved, breaking through the limitations of traditional single-sensor error compensation, and thus greatly improving the accuracy of belt scale weighing data measurement.

[0104] 3. By setting up a spatio-temporal fusion layer to innovatively combine dilated causal convolution (temporal coupling features), graph attention network (spatial propagation features), and cross-modal cross-attention (non-linear compensation features), joint modeling of spatio-temporal correlation and non-linear interference is realized for the first time in the weighing field, significantly improving the compensation accuracy under dynamic working conditions.

[0105] 4. By setting up a composite loss function that includes an error compensation precision loss sub-function, a first regularization loss sub-function, a second regularization loss sub-function, and a third regularization loss sub-function, while optimizing the compensation accuracy, regularization constraints are imposed on the multi-modal feature extraction layer, spatio-temporal fusion layer, and dynamic compensation output layer respectively, effectively preventing the overfitting problem.

[0106] 5. By adopting a combined scheme of three-level model compression (hybrid quantization + structured pruning + knowledge distillation) and online drift compensation training, not only can the compensation accuracy loss after compression be repaired, but also the model degradation problem caused by the aging of belt scales in industrial scenarios can be effectively solved, significantly extending the life cycle of the weighing error compensation model.

[0107] 6. By constructing a three-level encryption system (first encryption rule, second encryption rule, third encryption rule), the terminal side uses the first encryption rule (confused binary + RC6 + data transformation + SM4) for composite encryption, the transport layer uses the second encryption rule (dynamic key derivation (SHA-3 → exclusive OR operation → HMAC-SM3) + AES-GCM + RSA-OAEP + EdDSA) for composite encryption, the server uses the third encryption rule for composite encryption, and combined with blockchain to deposit and prove the fingerprint data, forming an end-to-end anti-tampering system, effectively improving the security of data storage and transmission.

[0108] 7. By adopting a preprocessing pipeline of sliding window filtering + outlier rejection + wavelet denoising, combined with Kalman filter dynamic compensation, real-time error correction with millisecond-level delay is achieved, meeting the timeliness requirements of industrial-level continuous weighing.

[0109] 8. By dynamically adjusting the learning rate during the training process and synchronously optimizing the hyperparameters and model structure parameters during the verification phase, the weighing error compensation model can maintain the optimal convergence state under complex working conditions.

[0110] 9. Deploy the weighing error compensation model through containerization technology to achieve rapid deployment and elastic scaling of the weighing error compensation model, support the cluster management of multiple belt scales, and reduce the on-site implementation complexity.

[0111] 10. Use the isolation forest algorithm on the server for anomaly detection, form a closed-loop feedback with the mobile terminal management system, support remote fault diagnosis and model iteration update, and effectively improve the system maintainability.

[0112] 11. By first incorporating mechanical characteristic parameters such as idler radial runout and load cell creep into the input features, establish a quantitative relationship model between structural mechanics characteristics and measurement errors, and solve the problem that traditional methods are insensitive to mechanical deformation.

[0113] 12. Convert the weighing log to UTF-8 encoding to generate a binary data stream, calculate the integrity checksum of the binary data stream through the SHA-3-512 algorithm, intercept the first 128 bits from the integrity checksum as the check segment, divide the check segment into four groups of 32-bit data blocks, and insert them into the 1 / 4, 1 / 2, 3 / 4, and end positions of the binary data stream respectively to obtain a confused binary data stream. Perform a byte reversal operation on the confused binary data stream to obtain a first-level encrypted data, encrypt the first-level encrypted data into a second-level encrypted data through the RC6 algorithm, convert the second-level encrypted data into a Base85 encoded string, add 3 to the ASCII value of the vowel letters in the Base85 encoded string, circularly shift the consonant letters 5 bits to the right, replace the numeric characters with modulo 7 operations to obtain a third-level encrypted data, and encrypt the third-level decrypted data into the first encrypted log through the SM4 algorithm. Adopt at least 11 layers of security measures before and after (integrity checksum, check segment, check segment division, data block insertion position, byte reversal, RC6 algorithm, Base85 encoding, adding 3 to the ASCII value of vowel letters, circularly shifting consonant letters 5 bits to the right, replacing numeric characters with modulo 7 operations, SM4 algorithm), which greatly improves the security of local storage of weighing logs.

[0114] 13. By setting a validity period, obtaining the current timestamp, calculating the effective expiration timestamp based on the validity period and the current timestamp, creating a master key based on the AES256 algorithm, and calculating the hash value of the weighing log through the SHA-3 algorithm; concatenating the master key with the date part of the current timestamp to obtain concatenated data, and calculating the first-level dynamic key by applying the SHA-3 algorithm to the concatenated data; performing an exclusive OR operation on the first-level dynamic key and the hour part of the current timestamp to obtain the first exclusive OR data, and calculating the second-level dynamic key by applying the SHA-256 algorithm to the first exclusive OR data; performing an exclusive OR operation on the second-level dynamic key and the effective expiration timestamp to obtain the second exclusive OR data, and calculating the third-level dynamic key by applying the HMAC-SM3 algorithm to the second exclusive OR data; encrypting the weighing log by invoking the third-level dynamic key through the AES-GCM algorithm to obtain the first-level encrypted data; combining the first-level encrypted data, the hash value, and the effective expiration timestamp into combined data, encrypting the combined data into second-level encrypted data through the RSA-OAEP algorithm, encrypting the second-level encrypted data into third-level encrypted data through the ECC algorithm, and signing the third-level encrypted data, the effective expiration timestamp, and the master key through the EdDSA algorithm to obtain the first signature value, and encapsulating the third-level encrypted data, the effective expiration timestamp, the master key, and the first signature value into the second encrypted log; adopting at least 13 layers of security measures (validity period, AES256 algorithm, hash value, concatenated data, first-level dynamic key, first exclusive OR data, second-level dynamic key, second exclusive OR data, third-level dynamic key, AES-GCM algorithm, RSA-OAEP algorithm, ECC algorithm, EdDSA algorithm) before and after, greatly enhancing the security of uploading the weighing log.

[0115] 14. By setting the key generation chain (master key → first-level → second-level → third-level dynamic key) of the second encryption rule to be deeply bound to the timestamp, the key is automatically rotated every hour. Even if a single key is leaked, it only affects the data for 1 hour, and the cracking cost is increased by 10^6 times.

[0116] 15. By adopting a combined strategy of Z-score anomaly localization → pre-trained regression model repair → interpolation method filling in the preprocessing stage, compared with the traditional threshold filtering method, high-precision repair is achieved while retaining the data distribution characteristics.

[0117] 16. Through an innovative multi-modal deep learning model architecture (combining graph neural network, adversarial noise robust network, and cross-modal attention mechanism), accurate extraction of spatio-temporal coupling features of mechanical deformation, dynamic load, and environmental interference of belt scales is achieved. Combining a three-level model compression and dynamic drift compensation mechanism, while ensuring industrial-level real-time processing efficiency (inference speed of 24 ms), the weighing error is effectively reduced. A multi-level security system covering confusion encryption, spatio-temporal dynamic key derivation, and blockchain evidence storage is constructed, overcoming the problems of industrial data anti-tampering and trusted evidence storage. Through stream computing optimization and containerized deployment, the operation and maintenance costs are significantly reduced (energy consumption reduced, model update bandwidth decreased), forming a breakthrough advantage in terms of accuracy, security, and industry adaptability (supporting multi-domain transfer learning), providing an intelligent solution with high precision, high robustness, and full-life-cycle controllability for the industrial weighing field. BRIEF DESCRIPTION OF THE DRAWINGS

[0118] The present invention will be further described below with reference to the accompanying drawings in conjunction with embodiments.

[0119] Figure 1 is a flowchart of a method for managing weighing data of a belt scale based on deep learning according to the present invention.

[0120] Figure 2 is a schematic structural diagram of a system for managing weighing data of a belt scale based on deep learning according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0121] The overall idea of the technical solution in the embodiments of the present application is as follows: Error compensation is performed through a weighing error compensation model trained with a dataset constructed by the form variable of the scale frame structure, the radial runout distance of the idler, the belt tension, the creep variable of the weighing sensor, the slip of the speed sensor, the material humidity, the environmental temperature, the electromagnetic interference intensity, and the offloading degree. Multidimensional influencing factors are fully considered, and real-time compensation can be performed based on the collected real-time monitoring data, without the need for traditional shutdown calibration. Through encryption combining the first encryption rule, the second encryption rule, and the third encryption rule, the data fingerprint is uploaded to the chain, which can effectively prevent relevant data from being stolen and tampered with in plaintext. Combining the set authentication mechanism, the weighing database can be accessed online to manage the weighing log, thereby improving the accuracy and continuity of the weighing data measurement of the belt scale, the security of transmission and storage, and the convenience of management.

[0122] Please refer to Figures 1 to 2 As shown, a preferred embodiment of a method for managing weighing data of a belt scale based on deep learning according to the present invention includes the following steps:

[0123] Step S1. Create a weighing error compensation model based on a multi-modal feature extraction layer, a spatio-temporal fusion layer, and a dynamic compensation output layer connected in sequence, and set the loss function of the weighing error compensation model;

[0124] The multimodal feature extraction layer is used to extract fused features from the input monitoring data; the spatio-temporal fusion layer is used to perform spatio-temporal correlation modeling on the fused features to obtain compensated decision features; the dynamic compensation output layer is used to map the compensated decision features into weighing error compensation values and output them;

[0125] Step S2: Obtain a large amount of historical monitoring data and historical weighing data, preprocess each piece of the historical monitoring data and historical weighing data, and construct a data set; the historical monitoring data includes the structural deformation amount of the scale frame, the radial runout distance of the idler, the belt tension, the creep amount of the weighing sensor, the slip of the speed sensor, the material humidity, the ambient temperature, the electromagnetic interference intensity, and the offloading degree;

[0126] By first incorporating mechanical characteristic parameters such as the radial runout of the idler and the creep of the weighing sensor into the input features, a quantitative relationship model between the structural mechanical characteristics and the measurement error is established, solving the problem that traditional methods are insensitive to mechanical deformation.

[0127] Step S3: Divide the data set into a training set, a validation set, and a test set, train the weighing error compensation model through the training set and the loss function, and then verify and test the trained weighing error compensation model through the validation set and the test set;

[0128] Step S4: Compress, perform drift compensation training, and deploy the weighing error compensation model that passes the test;

[0129] Step S5: Collect the real-time weighing data and real-time monitoring data of the belt scale, preprocess the real-time weighing data and real-time monitoring data, input the real-time monitoring data into the deployed weighing error compensation model to obtain a weighing error compensation value, and correct the real-time weighing data based on the weighing error compensation value to obtain corrected weighing data;

[0130] Step S6: Real-time display the corrected weighing data through a pre-associated display screen, record a weighing log that at least includes the corrected weighing data, the real-time weighing data, and the real-time monitoring data, and based on a preset storage period, encrypt the weighing log into a first encrypted log according to the first encryption rule for local storage, and encrypt the weighing log into a second encrypted log according to the second encryption rule and upload it to the server through the TLS protocol;

[0131] Step S7: The server creates a weighing database, sets the authentication mechanism of the weighing database, decrypts and verifies the received second encrypted log to obtain the weighing log, calculates the data fingerprint of the weighing log, stores the data fingerprint in the blockchain, encrypts the weighing log into a third encrypted log according to the third encryption rule, and stores the third encrypted log in the weighing database;

[0132] By constructing a three - level encryption system (the first encryption rule, the second encryption rule, the third encryption rule), the terminal side uses the first encryption rule (confusing binary + RC6 + data transformation + SM4) for composite encryption, the transport layer uses the second encryption rule (dynamic key derivation (SHA - 3 → exclusive - OR operation → HMAC - SM3) + AES - GCM + RSA - OAEP + EdDSA) for composite encryption, the server uses the third encryption rule for composite encryption, and combines with the blockchain to deposit evidence for fingerprint data, forming an end - to - end anti - tampering system, effectively improving the security of data storage and transmission.

[0133] Step S8: The server performs anomaly monitoring on the corrected weighing data in the weighing log, and the mobile terminal accesses the weighing database through the authentication mechanism to perform online management on the weighing log;

[0134] Step S9: Continuously optimize and iterate the weighing error compensation model based on the weighing log, that is, perform real - time and dynamic optimization on the weighing error compensation model;

[0135] Through an innovative multi - modal deep - learning model architecture (integrating graph neural network, adversarial noise - robust network, and cross - modal attention mechanism), it realizes the accurate extraction of spatio - temporal coupling features of mechanical deformation, dynamic load, and environmental interference of belt scales. Combining with a three - level model compression and dynamic drift compensation mechanism, while ensuring industrial - level real - time processing efficiency (inference speed of 24ms), it effectively reduces weighing errors; constructs a multi - level security system covering confusing encryption, spatio - temporal dynamic key derivation, and blockchain evidence deposition, overcoming the problems of industrial data anti - tampering and trusted evidence deposition; through stream computing optimization and containerized deployment, significantly reduces operation and maintenance costs (reduced energy consumption, reduced model update bandwidth), forming a breakthrough advantage in terms of accuracy, security, and industry adaptability (supporting multi - domain transfer learning), providing an intelligent solution with high precision, high robustness, and full - life - cycle controllability for the industrial weighing field.

[0136] In the step S1, the multi - modal feature extraction layer is constructed based on a structural deformation encoding module, a dynamic load encoding module, an environmental interference encoding module, and a feature output module;

[0137] The structural deformation encoding module is used to extract offset features and connection stiffness features from the scale frame structural deformation amount, the radial runout distance of the idler, and the off-loading degree through a graph neural network (GNN); the dynamic load encoding module is used to extract spatial local features from the belt tension and the creep amount of the weighing sensor through a one-dimensional convolutional neural network (1D-CNN), and extract temporal dependence features from the speed sensor slip through a bidirectional long short-term memory network (BiLSTM); the environmental interference encoding module is used to extract robust features from the material humidity, the environmental temperature, and the electromagnetic interference intensity through an anti-noise robust network; the feature output module is used to fuse the offset features, the connection stiffness features, the spatial local features, the temporal dependence features, and the robust features through a cross-modal attention mechanism and feature channel splicing to obtain fused features;

[0138] The anti-noise robust network (Noise-Resistant Robust Network) is a network architecture or training method designed to improve the robustness of neural networks against noise and adversarial attacks; through specific design or training strategies, such networks can maintain stable performance when faced with noise in the input data (such as random noise, adversarial noise, etc.), reducing the probability of misclassification or incorrect output. The principle is to add adversarial samples during the training process, enabling the model to learn the characteristics of adversarial attacks, thereby improving the robustness against adversarial attacks.

[0139] By adopting a multi-modal combination of graph neural network (structural deformation encoding), 1D CNN + BiLSTM (dynamic load encoding), and anti-noise robust network (environmental interference encoding), joint modeling of multi-dimensional interference factors such as mechanical deformation, dynamic load, and environmental interference is achieved, breaking through the limitations of traditional single-sensor error compensation, and thus greatly improving the accuracy of belt scale weighing data measurement.

[0140] The spatio-temporal fusion layer is constructed based on a temporal feature correlation module, a spatial feature correlation module, a feature interaction module, and a fusion module; the temporal feature correlation module is used to mine temporal coupling features from the fused features through a dilated causal convolutional network; the spatial feature correlation module is used to mine spatial propagation features from the fused features through a graph attention network; the feature interaction module is used to mine non-linear compensation features from the fused features through a cross-modal cross-attention mechanism; the fusion module is used to perform weighted fusion on the mined temporal coupling features, spatial propagation features, and non-linear compensation features to obtain compensation decision features.

[0141] By setting up a spatio-temporal fusion layer, it innovatively combines dilated causal convolution (temporal coupling features), graph attention network (spatial propagation features), and cross-modal cross-attention (nonlinear compensation features), and for the first time realizes the joint modeling of spatio-temporal correlation and nonlinear interference in the weighing field, significantly improving the compensation accuracy under dynamic conditions.

[0142] In the step S1, the formula of the loss function is as follows: ;

[0143] Among them, represents the loss value of the loss function; represents the error compensation accuracy loss sub-function, and the mean square error function is adopted; represents the first regularization loss sub-function of the multi-modal feature extraction layer, and the L2 regularization function (weight decay) is adopted; represents the second regularization loss sub-function of the spatio-temporal fusion layer, and the L2 regularization function is adopted; represents the third regularization loss sub-function of the dynamic compensation output layer, and the L2 regularization function is adopted; , and both represent the regularization strength coefficients;

[0144] By setting up a composite loss function including the error compensation accuracy loss sub-function, the first regularization loss sub-function, the second regularization loss sub-function, and the third regularization loss sub-function, while optimizing the compensation accuracy, regularization constraints are imposed on the multi-modal feature extraction layer, the spatio-temporal fusion layer, and the dynamic compensation output layer respectively, effectively preventing the overfitting problem.

[0145] The step S2 is specifically as follows:

[0146] Obtain a large amount of historical monitoring data and historical weighing data of belt scales, including the structural deformation amount of the scale frame, the radial runout distance of the idler, the belt tension, the creep amount of the weighing sensor, the slip of the speed sensor, the material humidity, the ambient temperature, the electromagnetic interference intensity, and the off-loading degree; the historical weighing data includes the instantaneous weight, the instantaneous flow rate, and the total weight; multiple weighing areas (weighing sensors) are arranged side by side on the belt scale to calculate the off-loading degree; the off-loading degree is an index to measure the weighing accuracy of the belt scale under different load distributions;

[0147] The form change amount of the scale frame structure is measured by the laser interferometry method (using a dual-frequency laser interferometer, by installing a mirror group on the scale frame to measure the form change amount of the scale frame in different directions) or the strain gauge method (sticking strain gauges at key parts of the scale frame, measuring the strain value through a strain gauge, and then calculating the form change amount); the radial runout distance of the idler is measured by the dial indicator method (fixing a dial indicator on the idler bracket, making the contact head of the dial indicator contact with the surface of the idler, rotating the idler, observing the change of the reading of the dial indicator, and the maximum reading difference is the radial runout distance) or the laser displacement sensor method (using a laser displacement sensor to measure the radial runout of the idler surface); the belt tension is measured by a mechanical tensiometer (clamping the tensiometer on the belt and directly reading the tension value) or the force-deflection method (applying a vertical force in the middle of the belt span to cause a certain amount of deflection of the belt, and calculating the tension according to the deflection amount and the applied force); the creep amount of the load cell is measured by the static loading method (applying a certain load to the load cell, and measuring the change of its output signal after maintaining for a period of time) or the dynamic loading method (applying a periodically changing load to the load cell and measuring its output signals at different time points); the slip of the speed sensor is measured by the encoder method (using a rotary encoder to measure the actual speed of the belt, comparing it with the output value of the speed sensor, and calculating the slip) or the laser velocimeter method (using a laser velocimeter to measure the actual speed of the belt); the electromagnetic interference intensity is measured by the electromagnetic field strength meter method (using an electromagnetic field strength meter to measure the electromagnetic field strength around the device) or the spectrum analyzer method (using a spectrum analyzer to measure the electromagnetic interference spectrum around the device); the offloading degree is measured by the multi-point loading method (placing loads of the same weight at different positions on the weighing platform, measuring and recording the weighing results, and calculating the maximum difference) or the offloading test method (using an offloading test device to simulate different offloading conditions and measuring the weighing results).

[0148] Outliers are located from the historical monitoring data and historical weighing data through the Z-score method, the outliers are repaired through a pre-trained regression model, the missing values in the historical monitoring data and historical weighing data are filled through the interpolation method, and then the historical monitoring data are marked with historical weighing data and weighing error compensation values to complete the preprocessing, and a data set is constructed based on the preprocessed historical monitoring data and historical weighing data;

[0149] The weighing error compensation value includes an instantaneous weight compensation value, an instantaneous flow compensation value, and a total weight compensation value.

[0150] By adopting the combined strategy of Z-score outlier localization → pre-trained regression model repair → interpolation method filling in the preprocessing stage, compared with the traditional threshold filtering method, high-precision repair is achieved while retaining the data distribution characteristics.

[0151] The specific step S3 is as follows:

[0152] The dataset is divided into a training set, a validation set, and a test set in a ratio of 6:2:2 by the stratified random sampling method. The training set is used to train the weighing error compensation model through the adaptive moment estimation optimizer. During the training process, the learning rate is dynamically adjusted until the loss value of the loss function decreases by less than 1e-5 in three consecutive training cycles. "1e-5" is a numerical value represented in scientific notation, which means 1×10 -5 , that is, 0.00001;

[0153] The F1 score and the coefficient of determination (R 2 ) are calculated through the validation set to verify the trained weighing error compensation model, and the hyperparameters and model structure parameters of the weighing error compensation model are tuned synchronously;

[0154] The mean absolute percentage error and the Pearson correlation coefficient are calculated through the test set to test the verified weighing error compensation model; when the verification fails or the test fails, the dataset is expanded and training continues;

[0155] By dynamically adjusting the learning rate during the training process and synchronously optimizing the hyperparameters and model structure parameters during the validation phase, the weighing error compensation model can maintain the optimal convergence state under complex working conditions.

[0156] The specific steps of step S4 are as follows:

[0157] The weighing error compensation model that has passed the test is compressed at the first level through the hybrid quantization technique, the redundant neurons are removed through the structured pruning method to compress the weighing error compensation model after the first level at the second level, and the weighing error compensation model after the second level is compressed at the third level through the knowledge distillation technique;

[0158] The actual monitoring data of the belt scale to be compensated is obtained, and the weighing error compensation model after the third level of compression is trained for drift compensation through the actual monitoring data. The weighing error compensation model after drift compensation training is verified based on the compensation accuracy, and then the weighing error compensation model after drift compensation training is deployed through the containerization technique, and the weighing error compensation model is deployed to the hardware security module (HSM).

[0159] By adopting the combined scheme of three-level model compression (hybrid quantization + structured pruning + knowledge distillation) and online drift compensation training, not only can the loss of compensation accuracy after compression be repaired, but also the problem of model degradation caused by the aging of the belt scale in the industrial scenario can be effectively solved, and the life cycle of the weighing error compensation model can be significantly extended.

[0160] The weighing error compensation model is deployed through containerization technology, enabling rapid deployment and elastic scaling of the weighing error compensation model, supporting clustered management of multiple belt scales, and reducing the complexity of on-site implementation.

[0161] The step S5 is specifically as follows:

[0162] Real-time weighing data and real-time monitoring data of the belt scale are collected in real time through a multi-source sensor array. The real-time weighing data and real-time monitoring data are pre-processed in sequence through a streaming computing engine using sliding window filtering, outlier removal, and wavelet noise reduction. The pre-processed real-time monitoring data is input into the deployed weighing error compensation model to obtain a weighing error compensation value. The weighing error compensation value is fused into the real-time weighing data through a Kalman filter algorithm for correction, and the corrected weighing data carrying the weighing time is synchronously output;

[0163] By adopting a preprocessing pipeline of sliding window filtering + outlier removal + wavelet noise reduction, combined with Kalman filtering dynamic compensation, real-time error correction with millisecond-level delay is achieved, meeting the timeliness requirements of industrial-grade continuous weighing.

[0164] A streaming computing engine is a computing framework for real-time processing and analysis of continuous data streams. It can process continuously generated data immediately, provide fast analysis results, and support low latency, high throughput, and dynamic data processing. Unlike traditional batch processing, streaming computing emphasizes real-time response and dynamic data processing, and is suitable for scenarios that require quick decision-making.

[0165] In step S6, the first encryption rule is specifically:

[0166] Performing UTF-8 encoding conversion on the weighing log to generate a binary data stream, calculating the integrity check code of the binary data stream using the SHA-3-512 algorithm, intercepting the first 128 bits from the integrity check code as a check segment, dividing the check segment into four groups of 32-bit data blocks, and inserting them into the 1 / 4, 1 / 2, 3 / 4, and end positions of the binary data stream, respectively, to obtain an obfuscated binary data stream, performing a byte reversal operation on the obfuscated binary data stream to obtain a layer of encrypted data, encrypting the first layer of encrypted data into a second layer of encrypted data using the RC6 algorithm, converting the second layer of encrypted data into a Base85 encoded string, adding 3 to the ASCII values of vowels in the Base85 encoded string, cyclically shifting the consonants right by 5 bits, and replacing the numeric characters with a modulo 7 operation to obtain a third layer of encrypted data, and encrypting the third layer of decrypted data into a first encrypted log using the SM4 algorithm;

[0167] Generate a binary data stream through UTF-8 encoding conversion using symmetric re-logs, calculate the integrity checksum of the binary data stream through the SHA-3-512 algorithm, intercept the first 128 bits from the integrity checksum as the check segment, divide the check segment into four groups of 32-bit data blocks, and insert them into the 1 / 4, 1 / 2, 3 / 4, and end positions of the binary data stream respectively to obtain a confused binary data stream. Perform a byte reversal operation on the confused binary data stream to obtain a first-level encrypted data, encrypt the first-level encrypted data into a second-level encrypted data through the RC6 algorithm, convert the second-level encrypted data into a Base85-encoded string, add 3 to the ASCII value of the vowel letters in the Base85-encoded string, circularly shift the consonant letters 5 bits to the right, and replace the numeric characters through modulo 7 operation to obtain a third-level encrypted data. Encrypt the third-level decrypted data into a first encrypted log through the SM4 algorithm. At least 11 security measures are taken before and after (integrity checksum, check segment, check segment division, data block insertion position, byte reversal, RC6 algorithm, Base85 encoding, adding 3 to the ASCII value of vowel letters, circularly shifting consonant letters 5 bits to the right, replacing numeric characters through modulo 7 operation, SM4 algorithm), which greatly improves the security of the weighing log local storage.

[0168] The specific second encryption rule is as follows:

[0169] Set a validity period, obtain the current timestamp, calculate the effective expiration timestamp based on the validity period and the current timestamp, create a master key based on the AES256 algorithm, and calculate the hash value of the weighing log through the SHA-3 algorithm;

[0170] Concatenate the master key with the date part of the current timestamp to obtain concatenated data, and calculate a first-level dynamic key by calculating the concatenated data through the SHA-3 algorithm;

[0171] XOR the first-level dynamic key with the hour part of the current timestamp to obtain a first XOR data, and calculate a second-level dynamic key by calculating the first XOR data through the SHA-256 algorithm;

[0172] XOR the second-level dynamic key with the effective expiration timestamp to obtain a second XOR data, and calculate a third-level dynamic key by calculating the second XOR data through the HMAC-SM3 algorithm;

[0173] The first-level encrypted data is obtained by calling the three-level dynamic key to encrypt the weighing log through the AES-GCM algorithm; the first-level encrypted data, hash value, and valid expiration timestamp are combined into combined data, and the combined data is encrypted into second-level encrypted data through the RSA-OAEP algorithm, the second-level encrypted data is encrypted into third-level encrypted data through the ECC algorithm, and the third-level encrypted data, valid expiration timestamp, and master key are signed through the EdDSA algorithm to obtain the first signature value, and the third-level encrypted data, valid expiration timestamp, master key, and first signature value are encapsulated into the second encrypted log.

[0174] By setting the validity period, obtaining the current timestamp, calculating the valid expiration timestamp based on the validity period and the current timestamp, creating the master key based on the AES256 algorithm, and calculating the hash value of the weighing log through the SHA-3 algorithm; the master key is concatenated with the date part of the current timestamp to obtain concatenated data, and the concatenated data is calculated through the SHA-3 algorithm to obtain the first-level dynamic key; the first-level dynamic key is XORed with the hour part of the current timestamp to obtain the first XOR data, and the first XOR data is calculated through the SHA-256 algorithm to obtain the second-level dynamic key; the second-level dynamic key is XORed with the valid expiration timestamp to obtain the second XOR data, and the second XOR data is calculated through the HMAC-SM3 algorithm to obtain the third-level dynamic key; the first-level encrypted data is obtained by calling the third-level dynamic key to encrypt the weighing log through the AES-GCM algorithm; the first-level encrypted data, hash value, and valid expiration timestamp are combined into combined data, the combined data is encrypted into second-level encrypted data through the RSA-OAEP algorithm, the second-level encrypted data is encrypted into third-level encrypted data through the ECC algorithm, the third-level encrypted data, valid expiration timestamp, and master key are signed through the EdDSA algorithm to obtain the first signature value, and the third-level encrypted data, valid expiration timestamp, master key, and first signature value are encapsulated into the second encrypted log; at least 13 security measures are taken before and after (validity period, AES256 algorithm, hash value, concatenated data, first-level dynamic key, first XOR data, second-level dynamic key, second XOR data, third-level dynamic key, AES-GCM algorithm, RSA-OAEP algorithm, ECC algorithm, EdDSA algorithm), which greatly improves the security of the weighing log upload.

[0175] By setting the key generation chain (master key → first-level → second-level → third-level dynamic key) of the second encryption rule to be deeply bound to the timestamp, the key is automatically rotated every hour. Even if a single key is leaked, it only affects 1 hour of data, and the cracking cost is increased by 10^6 times.

[0176] The specific step S7 is as follows:

[0177] The server creates a weighing database for storing the third encrypted log and sets the authentication mechanism of the weighing database;

[0178] The server receives the second encrypted log in real time, parses the second encrypted log to obtain three-level encrypted data, a valid expiration timestamp, a master key, and a first signature value; signs the three-level encrypted data, the valid expiration timestamp, and the master key through the EdDSA algorithm to obtain a second signature value, performs a signature verification operation on the first signature value through the second signature value, and then performs a timeliness verification through the valid expiration timestamp;

[0179] Decrypt the three-level encrypted data into two-level encrypted data through the ECC algorithm, decrypt the two-level encrypted data into combined data through the RSA-OAEP algorithm, parse the combined data to obtain first-level encrypted data, a hash value, and a valid expiration timestamp, and calculate the current timestamp based on the valid expiration timestamp and the validity period;

[0180] Concatenate the master key with the date part of the current timestamp to obtain concatenated data, calculate a first-level dynamic key by computing the concatenated data through the SHA-3 algorithm; perform an exclusive OR operation on the first-level dynamic key and the hour part of the current timestamp to obtain a first exclusive OR data, calculate a second-level dynamic key by computing the first exclusive OR data through the SHA-256 algorithm; perform an exclusive OR operation on the second-level dynamic key and the valid expiration timestamp to obtain a second exclusive OR data, calculate a third-level dynamic key by computing the second exclusive OR data through the HMAC-SM3 algorithm;

[0181] Decrypt the first-level encrypted data into a weighing log by invoking the third-level dynamic key through the AES-GCM algorithm, and perform an integrity check on the weighing log through the hash value;

[0182] The server calculates the data fingerprint of the weighing log through the HMAC algorithm, stores the data fingerprint in the blockchain, encrypts the weighing log into a third encrypted log based on the third encryption rule, and stores the third encrypted log in a structured manner in the weighing database;

[0183] The authentication process of the authentication mechanism is as follows:

[0184] Parsing the access request to the weighing database to obtain ciphertext data, signature data, request time, and MAC value; the calculation process of the ciphertext data is as follows: encrypt the account password, device ID, and public key used to verify the signature data through the IDEA algorithm to obtain the first encrypted data, map each character of the first encrypted data through a preset first mapping rule to obtain the second encrypted data, encrypt the second encrypted data through the RSA algorithm to obtain the third encrypted data, map each character of the third encrypted data through a preset second mapping rule to obtain the fourth encrypted data, and encrypt the fourth encrypted data through the SM9 algorithm to obtain the ciphertext data; the MAC value is obtained by performing MAC calculation on the ciphertext data, signature data, and request time.

[0185] After performing integrity verification on the ciphertext data, signature data, and request time through the MAC value, perform timeliness verification through the request time, then decrypt the ciphertext data to obtain the account password, device ID, and public key, and then verify the signature data through the public key. Perform the first legality verification on the device ID through a preset device management table, and finally perform the second legality verification on the account password through a preset password management table.

[0186] The specific third encryption rule is as follows:

[0187] Encrypt the weighing log through the ECC algorithm to obtain the first encrypted data, map the characters of the first encrypted data through a preset mapping rule to obtain the second encrypted data, replace the number 0 and the letter O with each other, and replace the number 8 and the letter Z with each other in the second encrypted data to obtain the third encrypted data, encrypt the third encrypted data through the XTEA algorithm to obtain the fourth encrypted data, convert the fourth encrypted data into hexadecimal data, and swap the number 8 and the letter B, and swap the number 3 and the letter E in the hexadecimal data to obtain the third encrypted log.

[0188] The specific step S8 is as follows:

[0189] The server performs anomaly monitoring and early warning on the corrected weighing data in the weighing log through the Isolation Forest algorithm. The mobile terminal accesses the weighing database through the authentication mechanism. The server decrypts the third encrypted log of the weighing database through the third encryption rule to obtain the weighing log for online management including addition, deletion, modification, and query. When the weighing log changes, it is re-encrypted into the third encrypted log for storage, and the data fingerprint is updated. The server records the access behavior data of the weighing database, performs security analysis on the access behavior data through a pre-trained risk recognition model, identifies abnormal access behaviors and issues early warnings.

[0190] The server uses the Isolation Forest algorithm for anomaly detection, forming a closed-loop feedback with the mobile terminal management system, supporting remote fault diagnosis and model iteration update, and effectively improving the maintainability of the system.

[0191] A preferred embodiment of a belt scale weighing data management system based on deep learning according to the present invention includes the following modules:

[0192] A weighing error compensation model creation module for creating a weighing error compensation model based on a multi-modal feature extraction layer, a spatio-temporal fusion layer, and a dynamic compensation output layer connected in sequence, and setting a loss function of the weighing error compensation model;

[0193] The multi-modal feature extraction layer is used to extract fusion features from the input monitoring data; the spatio-temporal fusion layer is used to perform spatio-temporal correlation modeling on the fusion features to obtain compensation decision features; the dynamic compensation output layer is used to map the compensation decision features into weighing error compensation values and output them;

[0194] A data set construction module for obtaining a large amount of historical monitoring data and historical weighing data, preprocessing each of the historical monitoring data and historical weighing data, and constructing a data set; the historical monitoring data includes the structural deformation amount of the scale frame, the radial runout distance of the idler, the belt tension, the creep amount of the weighing sensor, the slip of the speed sensor, the material humidity, the environmental temperature, the electromagnetic interference intensity, and the offloading degree;

[0195] By first incorporating mechanical characteristic parameters such as the radial runout of the idler and the creep of the weighing sensor into the input features, a quantitative relationship model between the structural mechanical characteristics and the measurement error is established, solving the problem that the traditional method is insensitive to mechanical deformation.

[0196] A weighing error compensation model training module for dividing the data set into a training set, a validation set, and a test set, training the weighing error compensation model through the training set and the loss function, and then validating and testing the trained weighing error compensation model through the validation set and the test set;

[0197] A weighing error compensation model deployment module for compressing, drift compensation training, and deploying the weighing error compensation model that has passed the test;

[0198] A weighing data compensation module for collecting the real-time weighing data and real-time monitoring data of the belt scale, preprocessing the real-time weighing data and real-time monitoring data, inputting the real-time monitoring data into the deployed weighing error compensation model to obtain a weighing error compensation value, and correcting the real-time weighing data based on the weighing error compensation value to obtain corrected weighing data;

[0199] The weighing log storage and backup module is used to display the corrected weighing data in real time through a pre-associated display screen, record the weighing log including at least the corrected weighing data, real-time weighing data, and real-time monitoring data in real time, encrypt the weighing log into a first encrypted log based on a first encryption rule for local storage at a preset storage period, and encrypt the weighing log into a second encrypted log based on a second encryption rule and upload it to the server through the TLS protocol;

[0200] The weighing log notarization module is used for the server to create a weighing database, set the authentication mechanism of the weighing database, decrypt and verify the received second encrypted log to obtain the weighing log, calculate the data fingerprint of the weighing log, store the data fingerprint in the blockchain, encrypt the weighing log into a third encrypted log based on a third encryption rule, and store the third encrypted log in the weighing database;

[0201] By constructing a three-level encryption system (the first encryption rule, the second encryption rule, the third encryption rule), the terminal side uses the first encryption rule (confusing binary + RC6 + data transformation + SM4) for composite encryption, the transport layer uses the second encryption rule (dynamic key derivation (SHA-3 → exclusive OR operation → HMAC-SM3) + AES-GCM + RSA-OAEP + EdDSA) for composite encryption, the server uses the third encryption rule for composite encryption, and combines the blockchain to notarize the fingerprint data to form an end-to-end anti-tampering system, effectively improving the security of data storage and transmission.

[0202] The weighing log management module is used for the server to perform anomaly monitoring on the corrected weighing data in the weighing log, and the mobile terminal accesses the weighing database through the authentication mechanism to perform online management of the weighing log;

[0203] The weighing error compensation model optimization module is used to continuously optimize and iterate the weighing error compensation model based on the weighing log, that is, to optimize the weighing error compensation model in real time and dynamically;

[0204] Through an innovative multi-modal deep learning model architecture (integrating graph neural network, adversarial noise robust network and cross-modal attention mechanism), accurate extraction of spatio-temporal coupling features of mechanical deformation, dynamic load and environmental interference of belt scales is achieved. Combined with a three-level model compression and dynamic drift compensation mechanism, while ensuring industrial-level real-time processing efficiency (inference speed of 24 ms), the weighing error is effectively reduced. A multi-level security system covering obfuscation encryption, spatio-temporal dynamic key derivation and blockchain evidence storage is constructed, solving the problems of industrial data tamper-proofing and trusted evidence storage. Through stream computing optimization and containerized deployment, the operation and maintenance costs are significantly reduced (energy consumption reduced, model update bandwidth decreased), forming a breakthrough advantage in terms of accuracy, security and industry adaptability (supporting multi-domain transfer learning), providing an intelligent solution with high precision, high robustness and full-life-cycle controllability for the industrial weighing field.

[0205] In the weighing error compensation model creation module, the multi-modal feature extraction layer is constructed based on a structural deformation encoding module, a dynamic load encoding module, an environmental interference encoding module and a feature output module;

[0206] The structural deformation encoding module is used to extract offset features and connection stiffness features from the scale frame structural deformation amount, idler radial runout distance and offloading degree through a graph neural network (GNN); the dynamic load encoding module is used to extract spatial local features from the belt tension and weighing sensor creep amount through a one-dimensional convolutional neural network (1D-CNN), and extract temporal dependence features from the speed sensor slip through a bidirectional long short-term memory network (BiLSTM); the environmental interference encoding module is used to extract robust features from the material humidity, environmental temperature and electromagnetic interference intensity through an adversarial noise robust network; the feature output module is used to fuse the offset features, connection stiffness features, spatial local features, temporal dependence features and robust features through a cross-modal attention mechanism and feature channel splicing to obtain fused features;

[0207] The adversarial noise robust network (Noise-Resistant Robust Network) is a network architecture or training method designed to improve the robustness of neural networks against noise and adversarial attacks; such networks can maintain stable performance and reduce the probability of misclassification or incorrect output when facing noise (such as random noise, adversarial noise, etc.) in the input data through specific design or training strategies. The principle is to add adversarial samples during the training process so that the model can learn the features of adversarial attacks, thereby improving the robustness against adversarial attacks.

[0208] By adopting a multi-modal combination of a graph neural network (structural deformation encoding), 1D CNN + BiLSTM (dynamic load encoding), and an adversarial noise-robust network (environmental interference encoding), joint modeling of multi-dimensional interference factors such as mechanical deformation, dynamic load, and environmental interference is achieved, breaking through the limitations of traditional single-sensor error compensation, and thus greatly improving the accuracy of weighing data measurement of belt scales.

[0209] The spatio-temporal fusion layer is constructed based on a temporal feature correlation module, a spatial feature correlation module, a feature interaction module, and a fusion module; the temporal feature correlation module is used to mine temporal coupling features from the fusion features through a dilated causal convolutional network; the spatial feature correlation module is used to mine spatial propagation features from the fusion features through a graph attention network; the feature interaction module is used to mine non-linear compensation features from the fusion features through a cross-modal cross-attention mechanism; the fusion module is used to perform weighted fusion on the mined temporal coupling features, spatial propagation features, and non-linear compensation features to obtain compensation decision features.

[0210] By setting up a spatio-temporal fusion layer to innovatively combine dilated causal convolution (temporal coupling features), graph attention network (spatial propagation features), and cross-modal cross-attention (non-linear compensation features), joint modeling of spatio-temporal correlation and non-linear interference is realized in the weighing field for the first time, significantly improving the compensation accuracy under dynamic working conditions.

[0211] In the weighing error compensation model creation module, the formula of the loss function is: ;

[0212] where represents the loss value of the loss function; represents the error compensation accuracy loss sub-function, and the mean square error function is adopted; represents the first regularization loss sub-function of the multi-modal feature extraction layer, and the L2 regularization function (weight decay) is adopted; represents the second regularization loss sub-function of the spatio-temporal fusion layer, and the L2 regularization function is adopted; represents the third regularization loss sub-function of the dynamic compensation output layer, and the L2 regularization function is adopted; , and both represent the regularization strength coefficients;

[0213] By setting up a composite loss function including an error compensation accuracy loss sub-function, a first regularization loss sub-function, a second regularization loss sub-function, and a third regularization loss sub-function, while optimizing the compensation accuracy, regularization constraints are imposed on the multi-modal feature extraction layer, the spatio-temporal fusion layer, and the dynamic compensation output layer respectively, effectively preventing the overfitting problem.

[0214] The dataset construction module is specifically used for:

[0215] Obtaining historical monitoring data and historical weighing data of a large number of belt scales, including the form variable of the scale frame structure, the radial runout distance of the idler, the belt tension, the creep variable of the weighing sensor, the slip of the speed sensor, the material humidity, the ambient temperature, the electromagnetic interference intensity, and the offloading degree; the historical weighing data includes the instantaneous weight, the instantaneous flow rate, and the total weight; multiple weighing areas (weighing sensors) are arranged side by side on the belt scale to calculate the offloading degree; the offloading degree is an index to measure the weighing accuracy of the belt scale under different load distributions;

[0216] The form variable of the scale frame structure is measured by the laser interferometry method (using a dual-frequency laser interferometer, by installing a mirror group on the scale frame to measure the form variable in different directions) or the strain gauge method (sticking strain gauges at key parts of the scale frame and measuring the strain value through a strain gauge to calculate the form variable); the radial runout distance of the idler is measured by the dial indicator method (fixing a dial indicator on the idler bracket, making the contact point of the dial indicator contact the surface of the idler, rotating the idler, and observing the change in the reading of the dial indicator, and the maximum reading difference is the radial runout distance) or the laser displacement sensor method (using a laser displacement sensor to measure the radial runout of the idler surface); the belt tension is measured by a mechanical tensiometer (clamping the tensiometer on the belt and directly reading the tension value) or the force-deflection method (applying a vertical force in the middle of the belt span to cause a certain amount of deflection of the belt and calculating the tension based on the deflection amount and the applied force); the creep variable of the weighing sensor is measured by the static loading method (applying a certain load to the weighing sensor and measuring the change in its output signal after maintaining for a period of time) or the dynamic loading method (applying a periodically changing load to the weighing sensor and measuring its output signals at different time points); the slip of the speed sensor is measured by the encoder method (using a rotary encoder to measure the actual speed of the belt and comparing it with the output value of the speed sensor to calculate the slip) or the laser velocimeter method (using a laser velocimeter to measure the actual speed of the belt); the electromagnetic interference intensity is measured by the electromagnetic field intensity meter method (using an electromagnetic field intensity meter to measure the electromagnetic field intensity around the device) or the spectrum analyzer method (using a spectrum analyzer to measure the electromagnetic interference spectrum around the device); the offloading degree is measured by the multi-point loading method (placing loads of the same weight at different positions on the weighing platform, measuring and recording the weighing results, and calculating the maximum difference) or the offloading test method (using an offloading test device to simulate different offloading conditions and measuring the weighing results).

[0217] Locating outliers from each of the historical monitoring data and the historical weighing data using a Z-score method, repairing each of the outliers using a pre-trained regression model, filling missing values in the historical monitoring data and the historical weighing data using an interpolation method, and then labeling each of the historical monitoring data with historical weighing data and weighing error compensation values to complete preprocessing, and constructing a data set based on each of the preprocessed historical monitoring data and historical weighing data;

[0218] The weighing error compensation value includes an instantaneous weight compensation value, an instantaneous flow compensation value, and a total weight compensation value.

[0219] By adopting a combined strategy of Z-score anomaly positioning → pre-trained regression model repair → interpolation filling in the preprocessing stage, compared with the traditional threshold filtering method, high-precision repair is achieved while retaining the data distribution characteristics.

[0220] The weighing error compensation model training module is specifically used for:

[0221] The dataset is divided into a training set, a validation set, and a test set in a ratio of 6:2:2 by a stratified random sampling method. The training set is used to train the weight error compensation model by using an adaptive moment estimation optimizer. The learning rate is dynamically adjusted during the training process until the loss value of the loss function decreases by less than 1e-5 in three consecutive training cycles; "1e-5" is a value expressed in scientific notation, which represents 1×10 -5 , that is, 0.00001;

[0222] The F1 score and the coefficient of determination (R) were calculated using the validation set. 2 ) to verify the trained weighing error compensation model and simultaneously tune the hyperparameters and model structure parameters of the weighing error compensation model;

[0223] The mean absolute percentage error and the Pearson correlation coefficient are calculated using the test set to test the weighing error compensation model that has passed verification; when verification or test fails, the data set is expanded to continue training;

[0224] By dynamically adjusting the learning rate during the training process and simultaneously optimizing the hyperparameters and model structure parameters during the verification phase, the weighing error compensation model can maintain the optimal convergence state under complex working conditions.

[0225] The weighing error compensation model deployment module is specifically used to:

[0226] Perform primary compression on the tested weighing error compensation model through hybrid quantization technology, perform secondary compression on the primary-compressed weighing error compensation model by removing redundant neurons through structured pruning method, and perform tertiary compression on the secondary-compressed weighing error compensation model through knowledge distillation technology;

[0227] Obtain the actual monitoring data of the belt scale to be compensated, perform drift compensation training on the tertiary-compressed weighing error compensation model through the actual monitoring data, verify the weighing error compensation model after drift compensation training based on the compensation accuracy, and then deploy the weighing error compensation model after drift compensation training through containerization technology, and deploy the weighing error compensation model to the Hardware Security Module (HSM).

[0228] By adopting a combined scheme of tertiary model compression (hybrid quantization + structured pruning + knowledge distillation) and online drift compensation training, not only can the loss of compensation accuracy after compression be repaired, but also the problem of model degradation caused by the aging of belt scales in industrial scenarios can be effectively solved, and the life cycle of the weighing error compensation model can be significantly extended.

[0229] Deploy the weighing error compensation model through containerization technology, realize the rapid deployment and elastic scaling of the weighing error compensation model, support the clustered management of multiple belt scales, and reduce the on-site implementation complexity.

[0230] The weighing data compensation module is specifically used for:

[0231] Real-time collect the real-time weighing data and real-time monitoring data of the belt scale through a multi-source sensor array, use a streaming computing engine to perform preprocessing on the real-time weighing data and real-time monitoring data in turn through sliding window filtering, outlier removal, and wavelet denoising, input the preprocessed real-time monitoring data into the deployed weighing error compensation model to obtain a weighing error compensation value, and fuse the weighing error compensation value into the real-time weighing data through the Kalman filter algorithm for correction, and synchronously output the corrected weighing data carrying the weighing time;

[0232] By adopting a preprocessing pipeline of sliding window filtering + outlier removal + wavelet denoising, combined with Kalman filter dynamic compensation, real-time error correction with millisecond-level delay is realized, meeting the timeliness requirements of industrial-level continuous weighing.

[0233] A streaming computing engine is a computing framework for real-time processing and analysis of continuous data streams. It can instantaneously process continuously generated data, provide fast analysis results, support low-latency, high-throughput, and dynamic data processing; different from traditional batch processing, streaming computing emphasizes real-time response and dynamic data processing, and is suitable for scenarios that require quick decision-making.

[0234] In the weighing log storage backup module, the first encryption rule is specifically:

[0235] Performing UTF-8 encoding conversion on the weighing log to generate a binary data stream, calculating the integrity check code of the binary data stream using the SHA-3-512 algorithm, intercepting the first 128 bits from the integrity check code as a check segment, dividing the check segment into four groups of 32-bit data blocks, and inserting them into the 1 / 4, 1 / 2, 3 / 4, and end positions of the binary data stream, respectively, to obtain an obfuscated binary data stream, performing a byte reversal operation on the obfuscated binary data stream to obtain a layer of encrypted data, encrypting the first layer of encrypted data into a second layer of encrypted data using the RC6 algorithm, converting the second layer of encrypted data into a Base85 encoded string, adding 3 to the ASCII values of vowels in the Base85 encoded string, cyclically shifting the consonants right by 5 bits, and replacing the numeric characters with a modulo 7 operation to obtain a third layer of encrypted data, and encrypting the third layer of decrypted data into a first encrypted log using the SM4 algorithm;

[0236] A binary data stream is generated by converting the weighing log into UTF-8 encoding. The integrity check code of the binary data stream is calculated using the SHA-3-512 algorithm. The first 128 bits of the integrity check code are intercepted as the check segment. The check segment is divided into four groups of 32-bit data blocks, which are inserted into the 1 / 4, 1 / 2, 3 / 4 and end positions of the binary data stream respectively to obtain an obfuscated binary data stream. A byte-reversal operation is performed on the obfuscated binary data stream to obtain a layer of encrypted data. The first layer of encrypted data is encrypted into a second layer of encrypted data using the RC6 algorithm. The second layer of encrypted data is converted into a Base85 encoded string. The ASCII value of the vowel in the Base85 encoded string is incremented by 3. , the consonants are cyclically shifted right by 5 bits, and the numeric characters are replaced by modulo 7 operations to obtain three-layer encrypted data. The three-layer decrypted data is encrypted into the first encrypted log through the SM4 algorithm. At least 11 security measures are taken before and after (integrity check code, check segment, check segment segmentation, data block insertion position, byte reversal, RC6 algorithm, Base85 encoding, vowel ASCII value plus 3, consonants are cyclically shifted right by 5 bits, numeric characters are replaced by modulo 7 operations, and SM4 algorithm), which greatly improves the security of local storage of weighing logs.

[0237] The second encryption rule is specifically:

[0238] Set a validity period, obtain the current timestamp, calculate the effective expiration timestamp based on the validity period and the current timestamp, create a master key based on the AES256 algorithm, and calculate the hash value of the weighing log using the SHA-3 algorithm;

[0239] Concatenate the master key with the date part of the current timestamp to obtain concatenated data, and calculate a first-level dynamic key by applying the SHA-3 algorithm to the concatenated data;

[0240] Exclusive-OR the first-level dynamic key with the hour part of the current timestamp to obtain first exclusive-OR data, and calculate a second-level dynamic key by applying the SHA-256 algorithm to the first exclusive-OR data;

[0241] Exclusive-OR the second-level dynamic key with the valid expiration timestamp to obtain second exclusive-OR data, and calculate a third-level dynamic key by applying the HMAC-SM3 algorithm to the second exclusive-OR data;

[0242] Call the third-level dynamic key through the AES-GCM algorithm to encrypt the weighing log to obtain first-level encrypted data; Combine the first-level encrypted data, the hash value, and the valid expiration timestamp into combined data, encrypt the combined data into second-level encrypted data through the RSA-OAEP algorithm, encrypt the second-level encrypted data into third-level encrypted data through the ECC algorithm, sign the third-level encrypted data, the valid expiration timestamp, and the master key through the EdDSA algorithm to obtain a first signature value, and encapsulate the third-level encrypted data, the valid expiration timestamp, the master key, and the first signature value into a second encrypted log.

[0243] By setting a validity period, obtaining the current timestamp, calculating the effective expiration timestamp based on the validity period and the current timestamp, creating a master key based on the AES256 algorithm, and calculating the hash value of the weighing log through the SHA-3 algorithm; concatenating the master key with the date part of the current timestamp to obtain concatenated data, and calculating a first-level dynamic key by applying the SHA-3 algorithm to the concatenated data; performing an exclusive OR operation on the first-level dynamic key and the hour part of the current timestamp to obtain a first exclusive OR data, and calculating a second-level dynamic key by applying the SHA-256 algorithm to the first exclusive OR data; performing an exclusive OR operation on the second-level dynamic key and the effective expiration timestamp to obtain a second exclusive OR data, and calculating a third-level dynamic key by applying the HMAC-SM3 algorithm to the second exclusive OR data; encrypting the weighing log by invoking the third-level dynamic key through the AES-GCM algorithm to obtain first-level encrypted data; combining the first-level encrypted data, the hash value, and the effective expiration timestamp into combined data, encrypting the combined data into second-level encrypted data through the RSA-OAEP algorithm, encrypting the second-level encrypted data into third-level encrypted data through the ECC algorithm, signing the third-level encrypted data, the effective expiration timestamp, and the master key through the EdDSA algorithm to obtain a first signature value, and encapsulating the third-level encrypted data, the effective expiration timestamp, the master key, and the first signature value into a second encrypted log; adopting at least 13 security measures before and after (validity period, AES256 algorithm, hash value, concatenated data, first-level dynamic key, first exclusive OR data, second-level dynamic key, second exclusive OR data, third-level dynamic key, AES-GCM algorithm, RSA-OAEP algorithm, ECC algorithm, EdDSA algorithm), which greatly improves the security of uploading the weighing log.

[0244] By setting the key generation chain of the second encryption rule (master key → first-level → second-level → third-level dynamic key) to be deeply bound to the timestamp, the key is automatically rotated every hour. Even if a single key is leaked, it only affects the data for 1 hour, and the cracking cost is increased by 10^6 times.

[0245] The weighing log evidence storage module is specifically used for:

[0246] The server creates a weighing database for storing the third encrypted log and sets the authentication mechanism of the weighing database;

[0247] The server receives the second encrypted log in real time, parses the second encrypted log to obtain the third-level encrypted data, the effective expiration timestamp, the master key, and the first signature value; signs the third-level encrypted data, the effective expiration timestamp, and the master key through the EdDSA algorithm to obtain a second signature value, performs a signature verification operation on the first signature value through the second signature value, and then performs a timeliness verification through the effective expiration timestamp;

[0248] Decrypt the three - level encrypted data into two - level encrypted data through the ECC algorithm, decrypt the two - level encrypted data into combined data through the RSA - OAEP algorithm, parse the combined data to obtain the first - level encrypted data, hash value, and valid expiration timestamp, and calculate the current timestamp based on the valid expiration timestamp and validity period;

[0249] Concatenate the master key with the date part of the current timestamp to obtain concatenated data, calculate the first - level dynamic key by passing the concatenated data through the SHA - 3 algorithm; perform an exclusive - OR operation on the first - level dynamic key and the hour part of the current timestamp to obtain the first exclusive - OR data, calculate the second - level dynamic key by passing the first exclusive - OR data through the SHA - 256 algorithm; perform an exclusive - OR operation on the second - level dynamic key and the valid expiration timestamp to obtain the second exclusive - OR data, calculate the third - level dynamic key by passing the second exclusive - OR data through the HMAC - SM3 algorithm;

[0250] Decrypt the first - level encrypted data by invoking the third - level dynamic key through the AES - GCM algorithm to obtain the weighing log, and perform integrity verification on the weighing log through the hash value;

[0251] The server calculates the data fingerprint of the weighing log through the HMAC algorithm, stores the data fingerprint in the blockchain, encrypts the weighing log into the third - level encrypted log based on the third encryption rule, and stores the third - level encrypted log in a structured manner in the weighing database;

[0252] The authentication process of the authentication mechanism is as follows:

[0253] Parse the access request of the weighing database to obtain ciphertext data, signature data, request time, and MAC value; the calculation process of the ciphertext data is as follows: encrypt the account password, device ID, and public key used to verify the signature data through the IDEA algorithm to obtain the first - level encrypted data, map each character of the first - level encrypted data through a preset first mapping rule to obtain the second - level encrypted data, encrypt the second - level encrypted data through the RSA algorithm to obtain the third - level encrypted data, map each character of the third - level encrypted data through a preset second mapping rule to obtain the fourth - level encrypted data, and encrypt the fourth - level encrypted data through the SM9 algorithm to obtain the ciphertext data; the MAC value is obtained by performing MAC calculation on the ciphertext data, signature data, and request time;

[0254] After performing integrity verification on the ciphertext data, signature data, and request time through the MAC value, perform timeliness verification through the request time, then decrypt the ciphertext data to obtain the account password, device ID, and public key, then verify the signature data through the public key, perform the first legality verification on the device ID through a preset device management table, and finally perform the second legality verification on the account password through a preset password management table.

[0255] The specific third encryption rule is as follows:

[0256] The weighing log is encrypted by the ECC algorithm to obtain the first encrypted data. The first encrypted data is character-mapped by a preset mapping rule to obtain the second encrypted data. The numbers 0 and the letter O in the second encrypted data are replaced with each other, and the numbers 8 and the letter Z are replaced with each other to obtain the third encrypted data. The third encrypted data is encrypted by the XTEA algorithm to obtain the fourth encrypted data. The fourth encrypted data is converted into hexadecimal data. The numbers 8 and the letter B in the hexadecimal data are swapped, and the numbers 3 and the letter E are swapped to obtain the third encrypted log.

[0257] The weighing log management module is specifically used for:

[0258] The server performs anomaly monitoring and early warning on the corrected weighing data in the weighing log through the isolation forest algorithm. The mobile terminal accesses the weighing database through the authentication mechanism. The server decrypts the third encrypted log of the weighing database through the third encryption rule to obtain the weighing log, so as to perform online management including addition, deletion, modification, and query on the weighing log. When the weighing log changes, it is re-encrypted into the third encrypted log for storage, and the data fingerprint is updated. The server records the access behavior data of the weighing database, and performs security analysis on the access behavior data through a pre-trained risk identification model to identify and warn of abnormal access behaviors.

[0259] The server uses the isolation forest algorithm for anomaly detection, forming a closed-loop feedback with the mobile terminal management system, supporting remote fault diagnosis and model iteration update, and effectively improving the maintainability of the system.

[0260] In summary, the advantages of the present invention are as follows:

[0261] 1. Create a weighing error compensation model by sequentially connecting a multi-modal feature extraction layer, a spatio-temporal fusion layer, and a dynamic compensation output layer, and set the loss function of the weighing error compensation model. Then, obtain a large amount of historical monitoring data and historical weighing data of belt scales, including the deformation amount of the scale frame structure, the radial runout distance of the idler, the belt tension, the creep amount of the weighing sensor, the slip of the speed sensor, the material humidity, the ambient temperature, the electromagnetic interference intensity, and the offloading degree, to construct a data set. Divide the data set into a training set, a validation set, and a test set. Train the weighing error compensation model through the training set and the loss function, and then verify and test the trained weighing error compensation model through the validation set and the test set. Compress, perform drift compensation training, and deploy the weighing error compensation model that passes the test. Then, collect the real-time weighing data and real-time monitoring data of the belt scale. After preprocessing the real-time weighing data and real-time monitoring data, input the real-time monitoring data into the deployed weighing error compensation model to obtain a weighing error compensation value, and correct the real-time weighing data based on the weighing error compensation value to obtain corrected weighing data. Then, display the corrected weighing data in real time through a display screen, and record a weighing log that includes at least the corrected weighing data, the real-time weighing data, and the real-time monitoring data. Based on a preset storage period, encrypt the weighing log into a first encrypted log according to the first encryption rule for local storage, and encrypt the weighing log into a second encrypted log according to the second encryption rule and upload it to the server through the TLS protocol. The server creates a weighing database, sets an authentication mechanism for the weighing database, decrypts and verifies the second encrypted log to obtain the weighing log, calculates the data fingerprint of the weighing log and stores it in the blockchain, encrypts the weighing log into a third encrypted log according to the third encryption rule, and stores the third encrypted log in the weighing database. Monitor the corrected weighing data for anomalies. The mobile terminal accesses the weighing database through the authentication mechanism to manage the weighing log online, and continuously optimizes and iterates the weighing error compensation model based on the weighing log. That is, error compensation is performed through a weighing error compensation model trained with a data set constructed by the deformation amount of the scale frame structure, the radial runout distance of the idler, the belt tension, the creep amount of the weighing sensor, the slip of the speed sensor, the material humidity, the ambient temperature, the electromagnetic interference intensity, and the offloading degree, fully considering multi-dimensional influencing factors, and real-time compensation can be performed based on the collected real-time monitoring data without the need for traditional shutdown calibration. Combined with the encryption of the first encryption rule, the second encryption rule, and the third encryption rule, and the data fingerprint is uploaded to the chain, which can effectively prevent relevant data from being stolen and tampered with in plain text. Combined with the set authentication mechanism, the weighing database can be accessed online to manage the weighing log, ultimately greatly improving the accuracy and continuity of belt scale weighing data measurement, the security of transmission and storage, and the convenience of management.

[0262] 2. By adopting a multi-modal combination of graph neural network (structural deformation encoding), 1D CNN + BiLSTM (dynamic load encoding), and adversarial noise robust network (environmental interference encoding), joint modeling of multi-dimensional interference factors such as mechanical deformation, dynamic load, and environmental interference is achieved, breaking through the limitations of traditional single-sensor error compensation, and thus greatly improving the accuracy of belt scale weighing data measurement.

[0263] 3. By setting up a spatio-temporal fusion layer to innovatively combine dilated causal convolution (temporal coupling features), graph attention network (spatial propagation features), and cross-modal cross-attention (non-linear compensation features), joint modeling of spatio-temporal correlation and non-linear interference is realized in the weighing field for the first time, significantly improving the compensation accuracy under dynamic working conditions.

[0264] 4. By setting up a composite loss function that includes an error compensation precision loss sub-function, a first regularization loss sub-function, a second regularization loss sub-function, and a third regularization loss sub-function, while optimizing the compensation accuracy, regularization constraints are imposed on the multi-modal feature extraction layer, spatio-temporal fusion layer, and dynamic compensation output layer respectively, effectively preventing the overfitting problem.

[0265] 5. By adopting a combined scheme of three-level model compression (hybrid quantization + structured pruning + knowledge distillation) and online drift compensation training, not only can the compensation accuracy loss after compression be repaired, but also the model degradation problem caused by the aging of belt scales in industrial scenarios can be effectively solved, significantly extending the life cycle of the weighing error compensation model.

[0266] 6. By constructing a three-level encryption system (first encryption rule, second encryption rule, third encryption rule), the terminal side uses the first encryption rule (confused binary + RC6 + data transformation + SM4) for composite encryption, the transport layer uses the second encryption rule (dynamic key derivation (SHA-3 → XOR operation → HMAC-SM3) + AES-GCM + RSA-OAEP + EdDSA) for composite encryption, the server uses the third encryption rule for composite encryption, and combined with blockchain to deposit and prove the fingerprint data, forming an end-to-end anti-tampering system, effectively improving the security of data storage and transmission.

[0267] 7. By adopting a preprocessing pipeline of sliding window filtering + outlier rejection + wavelet denoising, combined with Kalman filter dynamic compensation, real-time error correction with millisecond-level delay is achieved, meeting the timeliness requirements of industrial-level continuous weighing.

[0268] 8. By dynamically adjusting the learning rate during the training process and synchronously optimizing the hyperparameters and model structure parameters during the verification phase, the weighing error compensation model can maintain the optimal convergence state under complex working conditions.

[0269] 9. Deploy the weighing error compensation model through containerization technology to achieve rapid deployment and elastic scaling of the weighing error compensation model, support the clustered management of multiple belt scales, and reduce the on-site implementation complexity.

[0270] 10. The server uses the Isolation Forest algorithm for anomaly detection, forming a closed-loop feedback with the mobile terminal management system, supporting remote fault diagnosis and model iteration update, and effectively improving the system maintainability.

[0271] 11. By first incorporating mechanical characteristic parameters such as idler radial runout and load cell creep into the input features, a quantitative relationship model between structural mechanics characteristics and measurement errors is established to solve the problem that traditional methods are insensitive to mechanical deformation.

[0272] 12. Convert the weighing log to UTF-8 encoding to generate a binary data stream, calculate the integrity check code of the binary data stream through the SHA-3-512 algorithm, intercept the first 128 bits from the integrity check code as the check segment, divide the check segment into four groups of 32-bit data blocks, and insert them into the 1 / 4, 1 / 2, 3 / 4, and end positions of the binary data stream respectively to obtain a confused binary data stream. Perform a byte reversal operation on the confused binary data stream to obtain a first-layer encrypted data, encrypt the first-layer encrypted data into a second-layer encrypted data through the RC6 algorithm, convert the second-layer encrypted data into a Base85 encoded string, add 3 to the ASCII value of the vowel letters in the Base85 encoded string, circularly shift the consonant letters 5 bits to the right, and replace the numeric characters with modulo 7 operations to obtain a third-layer encrypted data. Encrypt the third-layer decrypted data into a first encrypted log through the SM4 algorithm, and adopt at least 11 layers of security measures before and after (integrity check code, check segment, check segment division, data block insertion position, byte reversal, RC6 algorithm, Base85 encoding, adding 3 to the ASCII value of vowel letters, circularly shifting consonant letters 5 bits to the right, replacing numeric characters with modulo 7 operations, SM4 algorithm), greatly enhancing the security of local storage of weighing logs.

[0273] 13. By setting a validity period, obtaining the current timestamp, calculating the effective expiration timestamp based on the validity period and the current timestamp, creating a master key based on the AES256 algorithm, and calculating the hash value of the weighing log through the SHA-3 algorithm; concatenating the master key with the date part of the current timestamp to obtain concatenated data, and calculating the first-level dynamic key by applying the SHA-3 algorithm to the concatenated data; performing an exclusive OR operation on the first-level dynamic key and the hour part of the current timestamp to obtain the first exclusive OR data, and calculating the second-level dynamic key by applying the SHA-256 algorithm to the first exclusive OR data; performing an exclusive OR operation on the second-level dynamic key and the effective expiration timestamp to obtain the second exclusive OR data, and calculating the third-level dynamic key by applying the HMAC-SM3 algorithm to the second exclusive OR data; encrypting the weighing log by invoking the third-level dynamic key through the AES-GCM algorithm to obtain the first-level encrypted data; combining the first-level encrypted data, the hash value, and the effective expiration timestamp into combined data, encrypting the combined data into second-level encrypted data through the RSA-OAEP algorithm, encrypting the second-level encrypted data into third-level encrypted data through the ECC algorithm, and signing the third-level encrypted data, the effective expiration timestamp, and the master key through the EdDSA algorithm to obtain the first signature value, and encapsulating the third-level encrypted data, the effective expiration timestamp, the master key, and the first signature value into the second encrypted log; adopting at least 13 layers of security measures (validity period, AES256 algorithm, hash value, concatenated data, first-level dynamic key, first exclusive OR data, second-level dynamic key, second exclusive OR data, third-level dynamic key, AES-GCM algorithm, RSA-OAEP algorithm, ECC algorithm, EdDSA algorithm) before and after, which greatly improves the security of uploading the weighing log.

[0274] 14. By setting the key generation chain (master key → first-level → second-level → third-level dynamic key) of the second encryption rule to be deeply bound to the timestamp, the key is automatically rotated every hour. Even if a single key is leaked, it only affects the data for 1 hour, and the cracking cost is increased by 10^6 times.

[0275] 15. By adopting a combined strategy of Z-score anomaly localization → pre-trained regression model repair → interpolation method filling in the preprocessing stage, compared with the traditional threshold filtering method, high-precision repair is achieved while retaining the data distribution characteristics.

[0276] 16. Through an innovative multi-modal deep learning model architecture (integrating graph neural network, adversarial noise robust network and cross-modal attention mechanism), accurate extraction of the spatio-temporal coupling characteristics of the mechanical deformation, dynamic load and environmental interference of the belt scale is achieved. Combining a three-level model compression and dynamic drift compensation mechanism, while ensuring industrial-level real-time processing efficiency (inference speed of 24 ms), the weighing error is effectively reduced; a multi-level security system covering obfuscation encryption, spatio-temporal dynamic key derivation and blockchain evidence storage is constructed, overcoming the problems of industrial data anti-tampering and trustworthy evidence storage; through stream computing optimization and containerized deployment, the operation and maintenance costs are significantly reduced (energy consumption reduced, model update bandwidth reduced), forming a breakthrough advantage in terms of accuracy, security and industry adaptability (supporting multi-domain transfer learning), providing an intelligent solution with high precision, high robustness and full-life-cycle controllability for the industrial weighing field.

[0277] Although the specific implementation manners of the present invention have been described above, those skilled in the art of this technology should understand that the specific embodiments we described are illustrative rather than used to limit the scope of the present invention. Equivalent modifications and changes made by those skilled in the art in accordance with the spirit of the present invention should be covered within the scope protected by the claims of the present invention.

Claims

1. A method for managing weighing data of a belt scale based on deep learning, characterized in that: The method includes the following steps: Step S1: Create a weighing error compensation model based on a sequentially connected multi-modal feature extraction layer, a spatio-temporal fusion layer, and a dynamic compensation output layer, and set the loss function of the weighing error compensation model; The multi-modal feature extraction layer is used to extract fusion features from the input monitoring data; The spatio-temporal fusion layer is used to perform spatio-temporal correlation modeling on the fusion features to obtain compensation decision features; the dynamic compensation output layer is used to map the compensation decision features into weighing error compensation values and output them; Step S2: Obtain a large amount of historical monitoring data and historical weighing data, preprocess each piece of the historical monitoring data and historical weighing data, and construct a data set; the historical monitoring data includes the structural deformation amount of the scale frame, the radial runout distance of the idler, the belt tension, the creep amount of the weighing sensor, the slip of the speed sensor, the material humidity, the ambient temperature, the electromagnetic interference intensity, and the off-loading degree; Step S3: Divide the data set into a training set, a validation set, and a test set, train the weighing error compensation model through the training set and the loss function, and then verify and test the trained weighing error compensation model through the validation set and the test set; Step S4: Compress, perform drift compensation training, and deploy the weighing error compensation model that passes the test; Step S5: Collect the real-time weighing data and real-time monitoring data of the belt scale, preprocess the real-time weighing data and real-time monitoring data, input the real-time monitoring data into the deployed weighing error compensation model to obtain a weighing error compensation value, and correct the real-time weighing data based on the weighing error compensation value to obtain corrected weighing data; Step S6: Real-time display the corrected weighing data through a pre-associated display screen, record in real time a weighing log including at least the corrected weighing data, the real-time weighing data, and the real-time monitoring data, and based on a first encryption rule, encrypt the weighing log into a first encrypted log for local storage, and encrypt the weighing log into a second encrypted log based on a second encryption rule and upload it to the server through the TLS protocol; Step S7: The server creates a weighing database, sets the authentication mechanism of the weighing database, decrypts and verifies the received second encrypted log to obtain the weighing log, calculates the data fingerprint of the weighing log, stores the data fingerprint in the blockchain, encrypts the weighing log into a third encrypted log based on a third encryption rule, and stores the third encrypted log in the weighing database; Step S8: The server performs anomaly monitoring on the corrected weighing data in the weighing log, and the mobile terminal accesses the weighing database through the authentication mechanism to perform online management of the weighing log; Step S9: Continuously optimize and iterate the weighing error compensation model based on the weighing log; In step S1, the multi-modal feature extraction layer is constructed based on a structural deformation encoding module, a dynamic load encoding module, an environmental interference encoding module, and a feature output module; The structure deformation encoding module is used to extract offset features and connection stiffness features from the scale frame structure deformation amount, the radial runout distance of the idler, and the degree of eccentric load through a graph neural network; The dynamic load encoding module is used to extract spatially local features from the belt tension and the creep amount of the weighing sensor through a one-dimensional convolutional neural network, and extract temporal dependence features from the slip of the speed sensor through a bidirectional long short-term memory network; the environmental interference encoding module is used to extract robust features from the material humidity, the environmental temperature, and the electromagnetic interference intensity through an adversarial noise robust network; the feature output module is used to fuse the offset features, the connection stiffness features, the spatially local features, the temporal dependence features, and the robust features through a cross-modal attention mechanism and feature channel splicing to obtain fused features; The spatio-temporal fusion layer is constructed based on a temporal feature correlation module, a spatial feature correlation module, a feature interaction module, and a fusion module; the temporal feature correlation module is used to mine temporal coupling features from the fused features through a dilated causal convolutional network; the spatial feature correlation module is used to mine spatial propagation features from the fused features through a graph attention network; the feature interaction module is used to mine non-linear compensation features from the fused features through a cross-modal cross-attention mechanism; the fusion module is used to perform weighted fusion on the mined temporal coupling features, spatial propagation features, and non-linear compensation features to obtain compensated decision features.

2. The belt scale weighing data management method based on deep learning according to claim 1, wherein: In the step S1, the formula of the loss function is as follows: ; Among them, represents the loss value of the loss function; represents the error compensation precision loss sub-function, using the mean square error function; represents the first regularization loss sub-function of the multi-modal feature extraction layer, using the L2 regularization function; represents the second regularization loss sub-function of the spatio-temporal fusion layer, using the L2 regularization function; represents the third regularization loss sub-function of the dynamic compensation output layer, using the L2 regularization function; and as well as both represent the regularization strength coefficient; The specific content of step S2 is as follows: Obtain a large amount of historical monitoring data and historical weighing data of belt scales, including the scale frame structure deformation amount, the radial runout distance of the idler, the belt tension, the creep amount of the weighing sensor, the slip of the speed sensor, the material humidity, the environmental temperature, the electromagnetic interference intensity, and the degree of eccentric load; the historical weighing data includes the instantaneous weight, the instantaneous flow rate, and the total weight; Locate outliers from each of the historical monitoring data and historical weighing data through the Z-score method, repair each of the outliers through a pre-trained regression model, fill in the missing values in the historical monitoring data and historical weighing data through interpolation, and then label each of the historical monitoring data with the historical weighing data and the weighing error compensation value to complete preprocessing, and construct a data set based on each of the preprocessed historical monitoring data and historical weighing data; The weighing error compensation value includes an instantaneous weight compensation value, an instantaneous flow rate compensation value, and a total weight compensation value.

3. A method for managing weighing data of a belt scale based on deep learning according to claim 1, characterized in that: The specific content of step S3 is as follows: Divide the data set into a training set, a validation set, and a test set according to a ratio of 6:2:2 through the stratified random sampling method, and call the training set through the adaptive moment estimation optimizer to train the weighing error compensation model. During the training process, dynamically adjust the learning rate until the loss value of the loss function decreases by less than 1e-5 in three consecutive training cycles; Calculate the F1 score and the coefficient of determination through the validation set to verify the trained weighing error compensation model, and synchronously tune the hyperparameters and model structure parameters of the weighing error compensation model; Calculate the mean absolute percentage error and Pearson correlation coefficient through the test set to test the verified weighing error compensation model; The specific step S4 is as follows: Perform primary compression on the verified weighing error compensation model through hybrid quantization technology, remove redundant neurons through structured pruning method to perform secondary compression on the model after primary compression, and perform tertiary compression on the model after secondary compression through knowledge distillation technology; Obtain the actual monitoring data of the belt scale to be compensated, perform drift compensation training on the weighing error compensation model after tertiary compression through the actual monitoring data, and then deploy the weighing error compensation model after drift compensation training through containerization technology.

4. A method for managing weighing data of a belt scale based on deep learning according to claim 1, characterized in that: The specific step S5 is as follows: Real-time collect the real-time weighing data and real-time monitoring data of the belt scale through a multi-source sensor array, and use a streaming computing engine to perform preprocessing on the real-time weighing data and real-time monitoring data in turn by using sliding window filtering, outlier removal and wavelet denoising. Input the preprocessed real-time monitoring data into the deployed weighing error compensation model to obtain a weighing error compensation value, and fuse the weighing error compensation value into the real-time weighing data through the Kalman filter algorithm for correction, and synchronously output the corrected weighing data carrying the weighing time; In step S6, the specific first encryption rule is: Perform UTF-8 encoding conversion on the weighing log to generate a binary data stream, calculate the integrity check code of the binary data stream through the SHA-3-512 algorithm, intercept the first 128 bits from the integrity check code as the check segment, divide the check segment into four groups of 32-bit data blocks, and insert them into the 1 / 4, 1 / 2, 3 / 4 and the end positions of the binary data stream respectively to obtain a confused binary data stream. Perform a byte reversal operation on the confused binary data stream to obtain a first-level encrypted data, encrypt the first-level encrypted data into a second-level encrypted data through the RC6 algorithm, convert the second-level encrypted data into a Base85 encoded string, add 3 to the ASCII value of the vowel letters in the Base85 encoded string, circularly shift the consonant letters 5 bits to the right, and replace the numeric characters through modulo 7 operation to obtain a third-level encrypted data, and encrypt the third-level decrypted data into the first encrypted log through the SM4 algorithm; The specific second encryption rule is: Set a validity period, obtain the current timestamp, calculate the effective expiration timestamp based on the validity period and the current timestamp, create a master key based on the AES256 algorithm, and calculate the hash value of the weighing log through the SHA-3 algorithm; Concatenate the master key with the date part of the current timestamp to obtain concatenated data, and calculate a first-level dynamic key through the SHA-3 algorithm for the concatenated data; XOR the first-level dynamic key with the hour part of the current timestamp to obtain a first XOR data, and calculate a second-level dynamic key through the SHA-256 algorithm for the first XOR data; The second-level dynamic key is XORed with the valid expiration timestamp to obtain second XOR data, and the third-level dynamic key is calculated from the second XOR data through the HMAC-SM3 algorithm; The third-level dynamic key is called through the AES-GCM algorithm to encrypt the weighing log to obtain first-level encrypted data; the first-level encrypted data, the hash value, and the valid expiration timestamp are combined into combined data, and the combined data is encrypted into second-level encrypted data through the RSA-OAEP algorithm, and the second-level encrypted data is encrypted into third-level encrypted data through the ECC algorithm. The third-level encrypted data, the valid expiration timestamp, and the master key are signed through the EdDSA algorithm to obtain a first signature value, and the third-level encrypted data, the valid expiration timestamp, the master key, and the first signature value are encapsulated into a second encrypted log.

5. The belt scale weighing data management method based on deep learning according to claim 1, characterized in that: The specific content of step S7 is as follows: The server creates a weighing database for storing the third encrypted log and sets the authentication mechanism of the weighing database; The server receives the second encrypted log in real time, and parses the second encrypted log to obtain the third-level encrypted data, the valid expiration timestamp, the master key, and the first signature value; The third-level encrypted data, the valid expiration timestamp, and the master key are signed through the EdDSA algorithm to obtain a second signature value, the first signature value is verified through the second signature value, and the timeliness is verified through the valid expiration timestamp; The third-level encrypted data is decrypted into second-level encrypted data through the ECC algorithm, the second-level encrypted data is decrypted into combined data through the RSA-OAEP algorithm, the combined data is parsed to obtain the first-level encrypted data, the hash value, and the valid expiration timestamp, and the current timestamp is calculated based on the valid expiration timestamp and the validity period; The master key is concatenated with the date part of the current timestamp to obtain concatenated data, and the first-level dynamic key is calculated from the concatenated data through the SHA-3 algorithm; The first-level dynamic key is XORed with the hour part of the current timestamp to obtain first XOR data, and the second-level dynamic key is calculated from the first XOR data through the SHA-256 algorithm; The second-level dynamic key is XORed with the valid expiration timestamp to obtain second XOR data, and the third-level dynamic key is calculated from the second XOR data through the HMAC-SM3 algorithm; The third-level dynamic key is called through the AES-GCM algorithm to decrypt the first-level encrypted data to obtain the weighing log, and the integrity of the weighing log is verified through the hash value; The server calculates the data fingerprint of the weighing log through the HMAC algorithm, stores the data fingerprint in the blockchain, encrypts the weighing log into the third encrypted log based on the third encryption rule, and stores the third encrypted log in a structured manner in the weighing database; The specific content of step S8 is as follows: The server performs anomaly monitoring and early warning on the corrected weighing data in the weighing log through the isolation forest algorithm. The mobile terminal accesses the weighing database through the authentication mechanism. The server decrypts the third encrypted log of the weighing database through the third encryption rule to obtain the weighing log, so as to perform online management including addition, deletion, modification, and query on the weighing log.

6. A belt scale weighing data management system based on deep learning, characterized in that: It includes the following modules: A weighing error compensation model creation module, which is used to create a weighing error compensation model based on a multi-modal feature extraction layer, a spatio-temporal fusion layer, and a dynamic compensation output layer connected in sequence, and set the loss function of the weighing error compensation model; The multi-modal feature extraction layer is used to extract fusion features from the input monitoring data; The spatio-temporal fusion layer is used to perform spatio-temporal correlation modeling on the fusion features to obtain compensation decision features; the dynamic compensation output layer is used to map the compensation decision features into weighing error compensation values and output them; A data set construction module, which is used to obtain a large amount of historical monitoring data and historical weighing data, preprocess each piece of historical monitoring data and historical weighing data, and construct a data set; the historical monitoring data includes the structural deformation amount of the scale frame, the radial runout distance of the idler, the belt tension, the creep amount of the weighing sensor, the slip of the speed sensor, the material humidity, the ambient temperature, the electromagnetic interference intensity, and the offloading degree; A weighing error compensation model training module, which is used to divide the data set into a training set, a validation set, and a test set, train the weighing error compensation model through the training set and the loss function, and then verify and test the trained weighing error compensation model through the validation set and the test set; A weighing error compensation model deployment module, which is used to compress, perform drift compensation training, and deploy the weighing error compensation model that passes the test; A weighing data compensation module, which is used to collect the real-time weighing data and real-time monitoring data of the belt scale, preprocess the real-time weighing data and real-time monitoring data, input the real-time monitoring data into the deployed weighing error compensation model to obtain a weighing error compensation value, and correct the real-time weighing data based on the weighing error compensation value to obtain corrected weighing data; A weighing log storage and backup module, which is used to display the corrected weighing data in real time through a pre-associated display screen, record in real time the weighing log including at least the corrected weighing data, the real-time weighing data, and the real-time monitoring data, and encrypt the weighing log into a first encrypted log based on the first encryption rule for local storage, and encrypt the weighing log into a second encrypted log based on the second encryption rule and upload it to the server through the TLS protocol; A weighing log certification module, which is used for the server to create a weighing database, set the authentication mechanism of the weighing database, decrypt and verify the received second encrypted log to obtain the weighing log, calculate the data fingerprint of the weighing log, store the data fingerprint in the blockchain, encrypt the weighing log into a third encrypted log based on the third encryption rule, and store the third encrypted log in the weighing database; The weighing log management module is used for the server to perform anomaly monitoring on the corrected weighing data in the weighing log, and the mobile terminal accesses the weighing database through the authentication mechanism to perform online management of the weighing log; The weighing error compensation model optimization module is used to continuously optimize and iterate the weighing error compensation model based on the weighing log; In the weighing error compensation model creation module, the multi-modal feature extraction layer is constructed based on the structural deformation coding module, the dynamic load coding module, the environmental interference coding module, and the feature output module; The structural deformation coding module is used to extract offset features and connection stiffness features from the scale frame structural deformation amount, the idler radial runout distance, and the offloading degree through a graph neural network; The dynamic load coding module is used to extract spatial local features from the belt tension and the creep amount of the weighing sensor through a one-dimensional convolutional neural network, and extract temporal dependence features from the speed sensor slip through a bidirectional long short-term memory network; the environmental interference coding module is used to extract robust features from the material humidity, the environmental temperature, and the electromagnetic interference intensity through an adversarial noise robust network; the feature output module is used to fuse the offset features, the connection stiffness features, the spatial local features, the temporal dependence features, and the robust features through a cross-modal attention mechanism and feature channel splicing to obtain fused features; The spatio-temporal fusion layer is constructed based on the temporal feature correlation module, the spatial feature correlation module, the feature interaction module, and the fusion module; the temporal feature correlation module is used to mine temporal coupling features from the fused features through a dilated causal convolutional network; the spatial feature correlation module is used to mine spatial propagation features from the fused features through a graph attention network; the feature interaction module is used to mine non-linear compensation features from the fused features through a cross-modal cross-attention mechanism; the fusion module is used to perform weighted fusion on the mined temporal coupling features, spatial propagation features, and non-linear compensation features to obtain compensation decision features.

7. The belt scale weighing data management system based on deep learning according to claim 6, characterized in that: In the weighing error compensation model creation module, the formula of the loss function is as follows: ; Among them, represents the loss value of the loss function; represents the error compensation precision loss sub-function, and the mean square error function is adopted; represents the first regularization loss sub-function of the multi-modal feature extraction layer, and the L2 regularization function is adopted; represents the second regularization loss sub-function of the spatio-temporal fusion layer, and the L2 regularization function is adopted; represents the third regularization loss sub-function of the dynamic compensation output layer, and the L2 regularization function is adopted; and as well as both represent the regularization strength coefficient; The dataset construction module is specifically used for: Obtaining a large amount of historical monitoring data including the scale frame structural deformation amount, the idler radial runout distance, the belt tension, the creep amount of the weighing sensor, the speed sensor slip, the material humidity, the environmental temperature, the electromagnetic interference intensity, and the offloading degree, as well as historical weighing data; the historical weighing data includes the instantaneous weight, the instantaneous flow rate, and the total weight; Locating outliers from each of the historical monitoring data and historical weighing data through the Z-score method, repairing each of the outliers through a pre-trained regression model, filling in the missing values in the historical monitoring data and historical weighing data through interpolation, and then annotating each of the historical monitoring data with the historical weighing data and the weighing error compensation value to complete preprocessing, and constructing a dataset based on each of the preprocessed historical monitoring data and historical weighing data; The weighing error compensation value includes an instantaneous weight compensation value, an instantaneous flow rate compensation value, and a total weight compensation value.

8. The belt scale weighing data management system based on deep learning according to claim 6, characterized in that: The weighing error compensation model training module is specifically used for: The dataset is divided into a training set, a validation set, and a test set in a ratio of 6:2:2 by the stratified random sampling method. The training set is used to train the weighing error compensation model by calling the adaptive moment estimation optimizer. During the training process, the learning rate is dynamically adjusted until the loss value of the loss function decreases by less than 1e-5 in three consecutive training cycles. The F1 score and the coefficient of determination are calculated through the validation set to verify the trained weighing error compensation model, and the hyperparameters and model structure parameters of the weighing error compensation model are tuned synchronously. The mean absolute percentage error and the Pearson correlation coefficient are calculated through the test set to test the verified weighing error compensation model. The weighing error compensation model deployment module is specifically used for: The weighing error compensation model that passes the test is compressed at the first level through the hybrid quantization technique, the redundant neurons are removed by the structured pruning method to perform the second-level compression on the first-level compressed weighing error compensation model, and the second-level compressed weighing error compensation model is compressed at the third level through the knowledge distillation technique. The actual monitoring data of the belt scale to be compensated is obtained, and the weighing error compensation model after the third-level compression is trained for drift compensation through the actual monitoring data, and then the weighing error compensation model after the drift compensation training is deployed through the containerization technique.

9. The belt scale weighing data management system based on deep learning according to claim 6, characterized in that: The weighing data compensation module is specifically used for: The real-time weighing data and real-time monitoring data of the belt scale are collected in real time through the multi-source sensor array. The streaming computing engine uses sliding window filtering, outlier removal, and wavelet denoising to preprocess the real-time weighing data and real-time monitoring data in sequence. The preprocessed real-time monitoring data is input into the deployed weighing error compensation model to obtain the weighing error compensation value. The weighing error compensation value is fused into the real-time weighing data through the Kalman filter algorithm for correction, and the corrected weighing data with the weighing time is output synchronously. In the weighing log storage and backup module, the first encryption rule is specifically: The weighing log is encoded and converted to a binary data stream in UTF-8. The integrity checksum of the binary data stream is calculated through the SHA-3-512 algorithm. The first 128 bits are intercepted from the integrity checksum as the check segment. The check segment is divided into four groups of 32-bit data blocks, which are inserted into the 1 / 4, 1 / 2, 3 / 4, and end positions of the binary data stream respectively to obtain the confused binary data stream. The byte reversal operation is performed on the confused binary data stream to obtain the first-level encrypted data. The first-level encrypted data is encrypted into the second-level encrypted data through the RC6 algorithm. The second-level encrypted data is converted into a Base85 encoded string. The ASCII value of the vowel letters in the Base ​ Set a validity period, obtain the current timestamp, calculate the effective expiration timestamp based on the validity period and the current timestamp, create a master key based on the AES256 algorithm, and calculate the hash value of the weighing log through the SHA-3 algorithm; Concatenate the master key with the date part of the current timestamp to obtain concatenated data, and calculate a first-level dynamic key by calculating the concatenated data through the SHA-3 algorithm; Exclusive OR the first-level dynamic key with the hour part of the current timestamp to obtain a first exclusive OR data, and calculate a second-level dynamic key by calculating the first exclusive OR data through the SHA-256 algorithm; Exclusive OR the second-level dynamic key with the effective expiration timestamp to obtain a second exclusive OR data, and calculate a third-level dynamic key by calculating the second exclusive OR data through the HMAC-SM3 algorithm; Call the third-level dynamic key through the AES-GCM algorithm to encrypt the weighing log to obtain first-level encrypted data; combine the first-level encrypted data, the hash value, and the effective expiration timestamp into combined data, encrypt the combined data into second-level encrypted data through the RSA-OAEP algorithm, encrypt the second-level encrypted data into third-level encrypted data through the ECC algorithm, sign the third-level encrypted data, the effective expiration timestamp, and the master key through the EdDSA algorithm to obtain a first signature value, and encapsulate the third-level encrypted data, the effective expiration timestamp, the master key, and the first signature value into a second encrypted log.

10. A belt scale weighing data management system based on deep learning according to claim 6, characterized in that: The weighing log evidence storage module is specifically used for: The server creates a weighing database for storing the third encrypted log and sets the authentication mechanism of the weighing database; The server receives the second encrypted log in real time, and parses the second encrypted log to obtain the third-level encrypted data, the effective expiration timestamp, the master key, and the first signature value; Sign the third-level encrypted data, the effective expiration timestamp, and the master key through the EdDSA algorithm to obtain a second signature value, perform a signature verification operation on the first signature value through the second signature value, and then perform a timeliness verification through the effective expiration timestamp; Decrypt the third-level encrypted data into second-level encrypted data through the ECC algorithm, decrypt the second-level encrypted data into combined data through the RSA-OAEP algorithm, parse the combined data to obtain the first-level encrypted data, the hash value, and the effective expiration timestamp, and calculate the current timestamp based on the effective expiration timestamp and the validity period; Concatenate the master key with the date part of the current timestamp to obtain concatenated data, and calculate a first-level dynamic key by calculating the concatenated data through the SHA-3 algorithm; Exclusive OR the first-level dynamic key with the hour part of the current timestamp to obtain a first exclusive OR data, and calculate a second-level dynamic key by calculating the first exclusive OR data through the SHA-256 algorithm; Exclusive OR the second-level dynamic key with the effective expiration timestamp to obtain a second exclusive OR data, and calculate a third-level dynamic key by calculating the second exclusive OR data through the HMAC-SM3 algorithm; The three - level dynamic key is called through the AES - GCM algorithm to decrypt the first - level encrypted data to obtain the weighing log, and the integrity of the weighing log is verified through the hash value; The server calculates the data fingerprint of the weighing log through the HMAC algorithm, stores the data fingerprint in the blockchain, encrypts the weighing log into a third - encrypted log based on the third encryption rule, and stores the third - encrypted log in a structured manner in the weighing database; The weighing log management module is specifically used for: The server performs anomaly monitoring and early warning on the corrected weighing data in the weighing log through the isolation forest algorithm. The mobile terminal accesses the weighing database through the authentication mechanism. The server decrypts the third - encrypted log in the weighing database through the third encryption rule to obtain the weighing log for online management including addition, deletion, modification, and query of the weighing log.

Images