Electric hand drill maintenance data collaborative modeling system based on federal learning

Abstract

The invention discloses an electric hand drill maintenance data collaborative modeling system based on federated learning, and belongs to the technical field of federated learning and distributed artificial intelligence. Maintaining the data nodes; the local update generation module is deployed at the maintenance data node and is used for training a local fault diagnosis model based on the locally stored original maintenance data and generating a local model parameter update vector; and the differential privacy injection module is deployed at the maintenance data node and is used for calculating a noise standard deviation according to a preset privacy budget parameter and injecting Gaussian noise into the local model parameter update vector based on the noise standard deviation so as to construct a noise addition update vector. And the compliance and the credibility of the system are greatly enhanced.

Description

Technical Field

[0001] This invention relates to the fields of federated learning and distributed artificial intelligence, specifically to a collaborative modeling system for electric drill maintenance data based on federated learning. Background Technology

[0002] In distributed application scenarios of equipment maintenance data, traditional fault diagnosis model building methods typically rely on centralizing raw data from various maintenance nodes, such as different maintenance centers or equipment sites, to a central server for unified analysis and training. However, this centralized approach faces significant challenges in practice.

[0003] Due to restrictions such as data sovereignty and trade secrets, maintenance nodes are usually unwilling or unable to directly share their original maintenance data containing sensitive information, which leads to a serious data silo problem. Moreover, traditional centralized systems are particularly vulnerable to malicious data attacks. Once a node submits deliberately constructed malicious data, i.e., a data poisoning attack, it may seriously pollute the entire dataset, thereby undermining the accuracy and reliability of the global diagnostic model.

[0004] The aforementioned situation and shortcomings mainly stem from the inherent flaws of traditional data processing architectures. This architecture requires the physical centralization of raw data, which fundamentally contradicts the core requirements of data privacy protection and security compliance in distributed scenarios. Without the ability to securely and reliably utilize data from multiple parties, it is difficult to construct a fault diagnosis model that reflects the overall situation accurately and reliably. This limits the depth and breadth of root cause analysis, impacting overall maintenance efficiency and the improvement of technical capabilities.

[0005] The information disclosed in the background section above is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a collaborative modeling system for electric drill repair data based on federated learning, in order to solve the problems mentioned in the background art. The technical solution of this invention includes:

[0007] Central server;

[0008] Maintenance data nodes;

[0009] The local update generation module is deployed on the maintenance data node and is used to train the local fault diagnosis model based on the raw maintenance data stored locally, and generate the local model parameter update vector.

[0010] The differential privacy injection module is deployed on the maintenance data node. It is used to calculate the noise standard deviation based on the preset privacy budget parameters, and inject Gaussian noise into the local model parameter update vector based on the noise standard deviation to construct a noisy update vector.

[0011] The exclusive consensus building module, deployed on the central server, is used to construct a corresponding exclusive mean vector for any maintenance data node by performing vector averaging on the noisy update vectors submitted by all other maintenance data nodes.

[0012] The semantic divergence quantification module, deployed on the central server, is used to quantify the directional deviation between the noisy update vector of each maintenance data node and the corresponding exclusive mean vector based on vector space cosine similarity calculation, so as to generate semantic divergence scores.

[0013] The adaptive weight allocation module, deployed on the central server, is used to input the semantic divergence score into a smooth decay function model based on the hyperbolic tangent function, and generate corresponding aggregate weights for each noisy update vector.

[0014] The normalized weighted aggregation module, deployed on the central server, is used to calculate the normalized global model update amount based on the aggregation weights of each node and the corresponding noisy update vector.

[0015] Preferably, the specific implementation of the differential privacy injection module includes:

[0016] The noise standard deviation is calculated based on preset privacy budget parameters;

[0017] Privacy budget parameters are used to quantify the strength of privacy protection;

[0018] The calculation of noise standard deviation follows the Gaussian mechanism of differential privacy theory;

[0019] Gaussian noise is generated by sampling from a Gaussian distribution with zero mean and standard deviation equal to the noise standard deviation.

[0020] The Gaussian noise is added to the local model parameter update vector to construct the noisy update vector.

[0021] Preferably, the specific implementation of the semantic divergence quantification module includes:

[0022] Calculate the dot product between the noisy update vector and the exclusive mean vector;

[0023] Calculate the magnitude of the noise-added update vector;

[0024] Calculate the magnitude of the exclusivity mean vector;

[0025] Divide the dot product by the product of the magnitude of the noisy update vector and the magnitude of the exclusive mean vector to obtain the cosine similarity value.

[0026] Subtract the cosine similarity value from the numerical value 1 to generate the semantic divergence score.

[0027] Preferably, the specific implementation of the adaptive weight allocation module includes:

[0028] Calculate the difference between the semantic divergence score and the preset semantic divergence isolation threshold;

[0029] Multiply the difference by the preset sharpness coefficient to obtain the adjusted divergence value;

[0030] Calculate the hyperbolic tangent function value of the adjusted divergence value;

[0031] Subtract the hyperbolic tangent function value from the numerical value 1 to obtain the median weight value;

[0032] Multiply the median weight by 0.5 to generate the aggregate weight.

[0033] Preferably, the configuration mechanism for the semantic divergence isolation threshold and sharpness coefficient includes:

[0034] The historical semantic divergence score sequence generated by the statistical system during normal operation in a trusted environment;

[0035] Calculate the mean and standard deviation of the historical semantic divergence score sequence;

[0036] The mean is added to three times the standard deviation to determine the semantic divergence isolation threshold;

[0037] In a verification environment that includes simulated data poisoning attacks, the sharpness coefficient is determined through experimental optimization.

[0038] Preferably, the specific implementation of the normalized weighted aggregation module includes:

[0039] The aggregate weight of each maintenance data node is multiplied by a scalar and a vector to obtain the weighted update vector.

[0040] The weighted update vectors of all maintenance data nodes are summed to construct the total weighted update vector;

[0041] Calculate the sum of the aggregate weights of all maintenance data nodes to obtain the weight denominator;

[0042] Divide the total weighted update vector by the weight denominator to calculate the global model update amount.

[0043] Compared with the prior art, the present invention has the following advantages:

[0044] 1. This system achieves quantifiable data privacy protection and compliance. The differential privacy injection module in the system injects Gaussian noise into the local model parameter update vector based on a preset privacy budget parameter. This not only avoids direct access to the original maintenance data, but also provides a mathematically provable privacy guarantee. Compared with traditional data anonymization or desensitization techniques, this system strictly links the strength of privacy protection to a clear and auditable privacy budget parameter, enabling the system to flexibly adapt to different data security compliance requirements and greatly enhancing the system's compliance and credibility.

[0045] 2. This system improves the accuracy and robustness of identifying anomalous data contributions. The system establishes an objective consensus reference for each maintenance data node through an exclusive consensus building module, unaffected by the node itself. Subsequently, the semantic divergence quantification module quantifies the directional deviation between the noisy update vector of each node and this consensus reference based on the cosine similarity of the vector space. This method focuses on the direction of the update vector rather than its length, enabling it to more fundamentally capture semantic inconsistencies caused by data poisoning attacks or local model failures. It is more sensitive to the identification of directional anomalies, thus achieving more accurate and robust identification of anomalous data contributions.

[0046] 3. This system constructs a flexible dynamic trust assessment and suppression mechanism. The adaptive weight allocation module of the system inputs the semantic divergence score into a smooth decay function model based on the hyperbolic tangent function, and assigns dynamic aggregation weights to the contribution of each node. Compared with a simple hard threshold filtering mechanism, this flexible governance capability can reduce the contribution of abnormal nodes proportionally according to the degree of deviation, rather than discarding them completely. This smooth decay characteristic avoids drastic changes in system decisions caused by small fluctuations, and significantly improves the stability of the global model training process. At the same time, its semantic divergence isolation threshold and sharpness coefficient are adaptively configured through data-driven optimization based on statistical principles and experiments, making the system's defense mechanism more scientific, automated, and environmentally adaptable.

[0047] 4. This system ensures the stability and convergence of global model training. When integrating the contributions of each node, the normalized weighted aggregation module normalizes the total weighted update vector by dividing it by the sum of the aggregated weights. This step effectively eliminates the risk of drastic fluctuations in the global update magnitude caused by changes in the number of nodes participating in aggregation or the weight distribution in each round. This makes the learning step size of the global model more consistent and controllable, ensuring smooth convergence of the model training process and obtaining a better-performing and more reliable global fault diagnosis model. Attached Figure Description

[0048] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0049] Figure 1This is a structural diagram of a collaborative modeling system for electric drill maintenance data based on federated learning, according to the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0051] Example 1:

[0052] Please see Figure 1 This invention provides a collaborative modeling system for electric drill repair data based on federated learning, comprising:

[0053] Central server;

[0054] Maintenance data nodes;

[0055] The local update generation module is deployed on the maintenance data node and is used to train the local fault diagnosis model based on the raw maintenance data stored locally, and generate the local model parameter update vector.

[0056] The differential privacy injection module is deployed on the maintenance data node. It is used to calculate the noise standard deviation based on the preset privacy budget parameters, and inject Gaussian noise into the local model parameter update vector based on the noise standard deviation to construct a noisy update vector.

[0057] The exclusive consensus building module, deployed on the central server, is used to construct a corresponding exclusive mean vector for any maintenance data node by performing vector averaging on the noisy update vectors submitted by all other maintenance data nodes.

[0058] The semantic divergence quantification module, deployed on the central server, is used to quantify the directional deviation between the noisy update vector of each maintenance data node and the corresponding exclusive mean vector based on vector space cosine similarity calculation, so as to generate semantic divergence scores.

[0059] The adaptive weight allocation module, deployed on the central server, is used to input the semantic divergence score into a smooth decay function model based on the hyperbolic tangent function, and generate corresponding aggregate weights for each noisy update vector.

[0060] The normalized weighted aggregation module, deployed on the central server, is used to calculate the normalized global model update amount based on the aggregation weights of each node and the corresponding noisy update vectors.

[0061] This embodiment discloses a collaborative modeling system for electric drill maintenance data based on federated learning. The purpose of this system is to build an accurate and reliable global fault diagnosis model that can resist malicious data attacks in a distributed, multi-source environment without the need for centralized storage of the original maintenance data of each maintenance node. This solves the problem of data silos caused by data sovereignty and trade secret restrictions. The system includes a central server and multiple maintenance data nodes.

[0062] In a typical business process, local update generation modules are deployed at each maintenance data node. These modules are independent computing units running on each node, tasked with training local fault diagnosis models using locally stored, non-externally shareable raw maintenance data. In one embodiment, the raw maintenance data is structured data, including real-time operating current, voltage, motor temperature, casing vibration frequency, and cumulative operating time of the electric drill. Its label field indicates the specific fault type, such as brush wear, bearing failure, poor switch contact, or normal operation. Correspondingly, the local fault diagnosis model can employ a feedforward neural network model, for example... The model consists of an input layer matching the data feature dimension (5 in this example), two hidden layers with 64 and 32 neurons respectively, using the ReLU activation function, and an output layer matching the number of fault types (4 in this example), using the Softmax activation function, for multi-class classification tasks. This output layer originates from a machine learning framework deployed on the nodes. The execution of this module produces a local model parameter update vector. This local model parameter update vector is a high-dimensional vector carrying the knowledge learned by the local model after one round of training. Its function is to contribute knowledge to the global model, and its origin is the output of the local model training process.

[0063] The differential privacy injection module deployed on the same maintenance data node is activated; the differential privacy injection module refers to the data privacy processing unit, which aims to add noise to the local model parameter update vector that is about to be uploaded, so as to provide mathematically provable privacy protection and prevent the original data information from being leaked; the module calculates the noise intensity according to the preset privacy budget parameter and injects Gaussian noise into the local model parameter update vector to construct a noisy update vector;

[0064] Each maintenance data node submits its generated noisy update vector to the central server; the exclusive consensus building module deployed on the central server begins to work; the exclusive consensus building module refers to the computing unit used to generate the benchmark comparison vector, with the purpose of establishing an objective global consensus reference for each maintenance data node that is not affected by itself; for any maintenance data node, this module constructs a corresponding exclusive mean vector for that node by performing vector averaging on the noisy update vectors submitted by all other maintenance data nodes;

[0065] The semantic divergence quantification module deployed on the central server is invoked; the semantic divergence quantification module refers to the computing unit used to evaluate the quality of data contribution, with the aim of quantifying the directional deviation between the update vector of a single node and the global consensus; based on the calculation principle of vector space cosine similarity, this module quantifies the directional deviation between the noisy update vector of each maintenance data node and the exclusive mean vector constructed for it, and generates a scalar form of semantic divergence score.

[0066] The score is then passed to the adaptive weight allocation module deployed on the central server. The adaptive weight allocation module refers to the dynamic trust evaluation and adjustment unit, which aims to assign dynamic aggregate weights to each node's contribution based on the semantic divergence score. This module inputs the semantic divergence score into a smooth decay function model based on the hyperbolic tangent function to generate corresponding aggregate weights for each noisy update vector. The higher the score, the greater the bias, and the lower the weight assigned.

[0067] The normalized weighted aggregation module deployed on the central server performs the final integration. The normalized weighted aggregation module refers to the computing unit used to generate global model updates safely and stably. Its purpose is to aggregate the contributions of all nodes in a weighted average manner. Based on the aggregation weights assigned to each node and the corresponding noisy update vector, the module calculates the normalized global model update amount, which is then used to update the global fault diagnosis model maintained by the central server.

[0068] This implementation method, through the collaborative work of the above modules, forms a closed-loop governance mechanism from privacy protection, unbiased anomaly detection, trust assessment to stable weighted aggregation; it can effectively utilize distributed data for collaborative modeling without accessing the original maintenance data of each node, while adaptively identifying and suppressing the negative impact of potential malicious or abnormal data sources, ensuring the accuracy and reliability of the generated global fault diagnosis model when used for root cause analysis.

[0069] Example 2:

[0070] The specific implementation of the differential privacy injection module includes:

[0071] The noise standard deviation is calculated based on preset privacy budget parameters;

[0072] Privacy budget parameters are used to quantify the strength of privacy protection;

[0073] The calculation of noise standard deviation follows the Gaussian mechanism of differential privacy theory;

[0074] Gaussian noise is generated by sampling from a Gaussian distribution with zero mean and standard deviation equal to the noise standard deviation.

[0075] The Gaussian noise is added to the local model parameter update vector to construct the noisy update vector.

[0076] To achieve the functionality of the differential privacy injection module described above, this embodiment provides a quantifiable and provable privacy protection mechanism;

[0077] The implementation of this module is contingent upon the setting of privacy budget parameters; privacy budget parameters It refers to a preset scalar value used to quantify the strength of privacy protection. Its function is to define a mathematical upper limit on the risk of privacy leakage under the differential privacy framework. It is set according to the data protection regulations or business compliance requirements that the system needs to meet. The smaller the value, the stronger the privacy protection.

[0078] The calculation of the noise standard deviation follows the Gaussian mechanism in differential privacy theory; in order to satisfy the parameters The defined privacy protection strength, the minimum noise standard deviation required. Calculated; noise standard deviation This refers to the dispersion of the Gaussian noise distribution. Its function is to control the amplitude of the injected noise to meet the privacy budget, and its source is based on privacy budget parameters. The sensitivity of the local model parameter update vector is calculated using a predetermined formula from differential privacy theory.

[0079] Generate local model parameter update vectors Then, L2 norm clipping is performed on it, that is, an upper norm limit is set. ; The value can be set based on statistical analysis of the L2 norm distribution of the historical local model parameter update vectors generated in a trusted environment, such as the initial few rounds. For example, the median or 90th percentile of all norms can be used as the standard. The value of is chosen to strike a balance between limiting the impact of extreme values ​​and retaining most of the effective information; if Then update the vector to Through this pruning step, the L2 sensitivity of the vector is strictly limited to [value missing]. Within a given privacy budget Noise standard deviation The following settings can be configured to meet differential privacy requirements:

[0080]

[0081] in It is a small constant close to zero, representing the probability that privacy protection fails; The letter belongs to the noise standard deviation setting formula, which is a preset upper norm limit used to perform L2 norm clipping on the local model parameter update vector. This symbol belongs to the noise standard deviation specification formula and represents the natural logarithm function; The letter refers to the noise standard deviation setting formula, which is a preset privacy budget parameter used to quantify the strength of privacy protection. The smaller the value, the stronger the protection.

[0082] The process of constructing the noisy update vector aims to mask the specific contribution of individual data by injecting noise with specific statistical properties into the local model parameter update vector; this process is accomplished by the following formula:

[0083]

[0084] in, Represents a node The noise-adding update vector; This represents the local model parameter update vector output by the local update generation module; This represents a Gaussian noise vector with the same dimension as the local model parameter update vector, where each component has a mean of zero and a standard deviation of . Obtained by independent sampling from a Gaussian distribution; : Subscript, representing node i; : Superscript, used to distinguish the vector after noise addition from the original vector;

[0085] The implementation involves the following steps: starting from a value with a mean of zero and a standard deviation of... Sampling is performed within a Gaussian distribution to generate a Gaussian noise vector. ; This noise vector is compared with the local model parameter update vector Perform vector addition to construct a noisy update vector. ;

[0086] This implementation links the strength of privacy protection with explicit, auditable privacy budget parameters. Strict linkage not only provides privacy protection functions, but more importantly, it provides a mathematically provable and quantifiable privacy guarantee, enabling the system to flexibly adapt to different data security compliance requirements, thereby greatly enhancing the system's compliance and credibility while ensuring data utility.

[0087] Example 3:

[0088] The specific implementation of the semantic divergence quantization module includes:

[0089] Calculate the dot product between the noisy update vector and the exclusive mean vector;

[0090] Calculate the magnitude of the noise-added update vector;

[0091] Calculate the magnitude of the exclusivity mean vector;

[0092] Divide the dot product by the product of the magnitude of the noisy update vector and the magnitude of the exclusive mean vector to obtain the cosine similarity value.

[0093] Subtract the cosine similarity value from the numerical value 1 to generate the semantic divergence score;

[0094] This embodiment is an implementation of the semantic divergence quantification module described above, with the aim of accurately quantifying the directional deviation between the contribution of a single data source and the global consensus;

[0095] The core of this module is based on the concept of cosine similarity in vector space, used to evaluate the angle between two high-dimensional vectors; the semantic divergence score is calculated to obtain a standardized scalar value that can intuitively reflect the degree of deviation; this score is calculated by adding noise to the updated vector output from the previous steps. With the exclusive mean vector The input is defined by the following formula:

[0096]

[0097] in, Represents a node The semantic divergence score; : Represents semantic divergence score; : Subscript, representing node i; Represents a node The noise-adding update vector; Represented as nodes Construct an exclusive mean vector; This symbol belongs to the semantic divergence score calculation formula, representing the Euclidean magnitude of the vector and the L2 norm calculation;

[0098] The underlying computational logic is: calculate the noisy update vector. With the exclusive mean vector Calculate the dot product of the two vectors and their Euclidean magnitudes respectively. and Divide the dot product by the product of the two moduli to obtain a cosine similarity value in the range [-1, 1]. To make the metric more intuitive, subtract this cosine similarity value from the numerical value 1 to generate a semantic divergence score. The score being close to 0 indicates that the node... The updates are highly consistent with the global consensus, while larger values ​​indicate significant deviations.

[0099] This implementation focuses on the direction of vectors rather than their length, enabling it to capture semantic inconsistencies more fundamentally. This makes quantitative evaluation insensitive to differences in the magnitude of updates at each node, but more sensitive to directional anomalies that may be caused by data poisoning attacks or local model failures, thus achieving more accurate and robust identification of anomalous data contributions.

[0100] Example 4:

[0101] The specific implementation of the adaptive weight allocation module includes:

[0102] Calculate the difference between the semantic divergence score and the preset semantic divergence isolation threshold;

[0103] Multiply the difference by the preset sharpness coefficient to obtain the adjusted divergence value;

[0104] Calculate the hyperbolic tangent function value of the adjusted divergence value;

[0105] Subtract the hyperbolic tangent function value from the numerical value 1 to obtain the median weight value;

[0106] Multiply the median weight by 0.5 to generate the aggregate weight;

[0107] This embodiment is an implementation of the adaptive weight allocation module described above. Its core purpose is to generate smooth and reasonable aggregate weights for the contribution of each node based on semantic divergence scores.

[0108] This module employs a smooth decay function model based on the hyperbolic tangent function to achieve dynamic weight allocation. This function is designed to apply gentle suppression, rather than rigid removal, to nodes with high semantic divergence scores. The weights are calculated based on the semantic divergence scores generated in previous steps. The core input is completed by the following formula:

[0109]

[0110] in, Represented as nodes The generated aggregate weights; The score is calculated by the semantic divergence quantification module; It is a preset semantic divergence isolation threshold; It is the preset sharpness coefficient; semantic divergence isolation threshold. This refers to the numerical boundary used to distinguish between normal and abnormal semantic divergence; its function is to define the inflection point where the weights begin to decay significantly; sharpness coefficient. It refers to a dimensionless hyperparameter that controls the steepness of the weight decay function, and its function is to adjust the system's penalty for abnormal deviations. This symbol belongs to the aggregate weight calculation formula and represents the hyperbolic tangent function;

[0111] To generate aggregate weights, calculate semantic divergence scores. Isolation threshold from preset semantic divergence The difference is calculated and multiplied by a preset sharpness factor. Based on the above results, the hyperbolic tangent function value is calculated, then the function value is subtracted from the value 1, and the result is multiplied by the value 0.5 to generate the aggregate weight with a value range falling within the interval [0, 1]. When the semantic divergence score of a node is much lower than the threshold, the weight will approach 1; when the score is much higher than the threshold, the weight will approach 0; the properties of the hyperbolic tangent function ensure that the weight change near the threshold is a smooth transition.

[0112] Compared to a simple hard threshold filtering mechanism, this implementation provides a flexible governance capability; it can reduce the contribution of abnormal nodes proportionally according to the degree of deviation, rather than discarding them completely. This smooth decay characteristic avoids drastic changes in system decisions caused by small fluctuations crossing the threshold, significantly improving the stability and robustness of the global model training process, while also enabling more refined handling of abnormal data of various degrees.

[0113] Example 5:

[0114] The configuration mechanisms for semantic divergence isolation threshold and sharpness coefficient include:

[0115] The historical semantic divergence score sequence generated by the statistical system during normal operation in a trusted environment;

[0116] Calculate the mean and standard deviation of the historical semantic divergence score sequence;

[0117] The mean is added to three times the standard deviation to determine the semantic divergence isolation threshold;

[0118] In a verification environment that includes simulated data poisoning attacks, the sharpness coefficient is determined through experimental optimization.

[0119] This embodiment refers to the semantic divergence isolation threshold mentioned above. With sharpness coefficient One implementation of the configuration mechanism aims to provide a data-driven, adaptive parameter setting method for the adaptive weight allocation module.

[0120] To determine the semantic divergence isolation threshold This configuration mechanism is based on statistical analysis of the system's behavior under normal operating conditions. The system operates in a trusted environment for a period of time. A trusted environment, for example, can refer to the first N aggregation rounds after the initial system deployment (e.g., N=100). In these rounds, the maintenance data nodes participating in collaborative modeling are all pre-selected nodes with a long history of good cooperation, and their data quality is considered reliable. Historical semantic divergence scores generated by the semantic divergence quantification module are continuously calculated to form a time series. The statistical mean of this historical score series is then calculated. with standard deviation Isolation threshold It is set as the sum of the mean and three standard deviations, i.e.:

[0121]

[0122] in, This letter belongs to the semantic divergence isolation threshold configuration formula and represents the semantic divergence isolation threshold; This symbol belongs to the semantic divergence isolation threshold configuration formula and represents the statistical mean of the historical semantic divergence score sequence; : Represents the statistical mean; : Subscript, representing semantic divergence score; This symbol belongs to the semantic divergence isolation threshold configuration formula and represents the standard deviation of the historical semantic divergence score sequence; : Represents standard deviation;

[0123] The technical principle behind this configuration is that, according to the three sigma criterion in statistics, scores generated by normal system fluctuations have a 99.7% probability of falling within this range; therefore, scores exceeding this threshold can be considered statistically significant anomalies.

[0124] To determine the sharpness coefficient This configuration mechanism aims to optimize the system's response performance in the face of real attacks. In an offline verification environment that includes simulated data poisoning attacks, such as a label flipping attack, one or more nodes in the verification environment are designated as malicious nodes, and a certain proportion, such as 30%, of their local training data is randomly mapped to incorrect fault types. For example, the label for brush wear is changed to bearing failure, thereby simulating the interference of malicious data sources on the global model. This can be achieved by setting different... Multiple rounds of experiments were conducted to evaluate the system under different conditions. The effectiveness of the system in resisting attacks under certain conditions is assessed, such as the decrease in accuracy of the global model and the convergence speed. Through experimental optimization, the optimal balance between rapid attack response and maintaining system stability is determined. value;

[0125] This configuration mechanism ensures that the system's core judgment parameters are no longer fixed values ​​set based on experience, but rather can adaptively optimize based on the system's historical performance and response under specific threat scenarios; thresholds are set based on statistical principles. This makes the anomaly detection baseline more scientific and realistic, reducing false positives; the sharpness coefficient was determined through experimental optimization. This ensures the effectiveness of the suppression strategy; this data-driven configuration method significantly improves the automation level, environmental adaptability and security defense capabilities of the entire distributed traceability management system.

[0126] Example 6:

[0127] The specific implementation of the normalized weighted aggregation module includes:

[0128] The aggregate weight of each maintenance data node is multiplied by a scalar and a vector to obtain the weighted update vector.

[0129] The weighted update vectors of all maintenance data nodes are summed to construct the total weighted update vector;

[0130] Calculate the sum of the aggregate weights of all maintenance data nodes to obtain the weight denominator;

[0131] Divide the total weighted update vector by the weight denominator to calculate the global model update amount;

[0132] To prevent numerical instability, a robustness check is required: set a very small positive threshold, for example... If the weight denominator is less than the threshold, it means that the contribution of all nodes in this round of aggregation is judged to be highly unreliable, and the system should skip this round of update and not modify the global model; if the weight denominator is greater than the threshold, the total weighted update vector is divided by the weight denominator to calculate the global model update amount.

[0133] This embodiment is a specific implementation of the normalized weighted aggregation module described above. Its purpose is to integrate the contributions from all trusted nodes into the final global model update in a stable and unbiased manner.

[0134] This module performs a normalized weighted average calculation to ensure that the magnitude of the global update is not affected by changes in the number of nodes participating in the aggregation in the current round; the calculation of this update amount is based on the noisy update vectors of each node generated in the previous steps. and the aggregate weights assigned to them For input, defined by the following formula:

[0135]

[0136] in, This represents the final calculated global model update amount; This represents the total number of maintenance data nodes participating in this round of aggregation; It is a node Aggregate weights; It is a node The noisy update vector; : Superscript indicates that the summation ends at the last node M; This letter belongs to the global model update calculation formula and serves as an index, representing a maintenance data node;

[0137] The implementation steps are as follows: for each maintenance data node Aggregate its weights With the corresponding noisy update vector Perform scalar-vector multiplication to obtain the weighted update vector of the node; sum the weighted update vectors of all nodes to construct the total weighted update vector; simultaneously, calculate the sum of the aggregated weights of all nodes to obtain the weight denominator; divide the total weighted update vector by the weight denominator to calculate the final global model update amount. ;

[0138] The normalization step in this implementation, namely dividing by the sum of weights, is key to ensuring the stability of model training. It eliminates the risk of drastic fluctuations in the global update magnitude that may be caused by different numbers of participating nodes in each round or changes in node weight distribution. This makes the learning step size of the global model more consistent and controllable, thereby ensuring the smooth convergence of the federated learning process and ultimately obtaining a better-performing and more reliable global fault diagnosis model.

[0139] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A collaborative modeling system for electric drill repair data based on federated learning, characterized in that, include: Central server; Maintenance data nodes; The local update generation module is deployed on the maintenance data node and is used to train the local fault diagnosis model based on the raw maintenance data stored locally, and generate the local model parameter update vector. The differential privacy injection module is deployed on the maintenance data node. It is used to calculate the noise standard deviation based on the preset privacy budget parameters, and inject Gaussian noise into the local model parameter update vector based on the noise standard deviation to construct a noisy update vector. The exclusive consensus building module, deployed on the central server, is used to construct a corresponding exclusive mean vector for any maintenance data node by performing vector averaging on the noisy update vectors submitted by all other maintenance data nodes. The semantic divergence quantification module, deployed on the central server, is used to quantify the directional deviation between the noisy update vector of each maintenance data node and the corresponding exclusive mean vector based on vector space cosine similarity calculation, so as to generate semantic divergence scores. The adaptive weight allocation module, deployed on the central server, is used to input the semantic divergence score into a smooth decay function model based on the hyperbolic tangent function, and generate corresponding aggregate weights for each noisy update vector. The normalized weighted aggregation module, deployed on the central server, is used to calculate the normalized global model update amount based on the aggregation weights of each node and the corresponding noisy update vector.

2. The collaborative modeling system for electric drill repair data based on federated learning according to claim 1, characterized in that, The specific implementation of the differential privacy injection module includes: The noise standard deviation is calculated based on preset privacy budget parameters; Privacy budget parameters are used to quantify the strength of privacy protection; The calculation of noise standard deviation follows the Gaussian mechanism of differential privacy theory; Gaussian noise is generated by sampling from a Gaussian distribution with zero mean and standard deviation equal to the noise standard deviation. The Gaussian noise is added to the local model parameter update vector to construct the noisy update vector.

3. The collaborative modeling system for electric drill repair data based on federated learning according to claim 1, characterized in that, The specific implementation of the semantic divergence quantification module includes: Calculate the dot product between the noisy update vector and the exclusive mean vector; Calculate the magnitude of the noise-added update vector; Calculate the magnitude of the exclusivity mean vector; Divide the dot product by the product of the magnitude of the noisy update vector and the magnitude of the exclusive mean vector to obtain the cosine similarity value. Subtract the cosine similarity value from the numerical value 1 to generate the semantic divergence score.

4. The collaborative modeling system for electric drill repair data based on federated learning according to claim 1, characterized in that, The specific implementation of the adaptive weight allocation module includes: Calculate the difference between the semantic divergence score and the preset semantic divergence isolation threshold; Multiply the difference by the preset sharpness coefficient to obtain the adjusted divergence value; Calculate the hyperbolic tangent function value of the adjusted divergence value; Subtract the hyperbolic tangent function value from the numerical value 1 to obtain the median weight value; Multiply the median weight by 0.5 to generate the aggregate weight.

5. A collaborative modeling system for electric drill repair data based on federated learning according to claim 4, characterized in that, The configuration mechanism for the semantic divergence isolation threshold and sharpness coefficient includes: The historical semantic divergence score sequence generated by the statistical system during normal operation in a trusted environment; Calculate the mean and standard deviation of the historical semantic divergence score sequence; The mean is added to three times the standard deviation to determine the semantic divergence isolation threshold; In a verification environment that includes simulated data poisoning attacks, the sharpness coefficient is determined through experimental optimization.

6. The collaborative modeling system for electric drill repair data based on federated learning according to claim 1, characterized in that, The specific implementation of the normalized weighted aggregation module includes: The aggregate weight of each maintenance data node is multiplied by a scalar and a vector to obtain the weighted update vector. The weighted update vectors of all maintenance data nodes are summed to construct the total weighted update vector; Calculate the sum of the aggregate weights of all maintenance data nodes to obtain the weight denominator; Divide the total weighted update vector by the weight denominator to calculate the global model update amount.

Images