A method for processing police incident data based on semantically weighted federated learning

Abstract

This application relates to the field of police incident analysis technology and provides a police incident data processing method based on semantic weighted federated learning. The method includes: edge nodes enhancing local fire alarm voice data to obtain a local training sample set, acquiring and uploading the local semantic abundance coefficients and contribution coefficients of the local training sample set; the cloud performing global aggregation based on the local semantic abundance coefficients and contribution coefficients to obtain a student model; the cloud selecting teacher nodes based on the local semantic abundance coefficients and all contribution coefficients, and determining a teacher model based on all teacher nodes; the cloud performing knowledge distillation on the student model based on the teacher model to obtain the final global model parameters, and distributing the final global model parameters to the edge nodes until the global aggregation termination condition is met; and the edge nodes using the final global model parameters to perform police incident analysis. This application can achieve police incident analysis while balancing security, accuracy, and analysis efficiency.

Description

Technical Field

[0001] This application belongs to the field of police incident analysis technology, and in particular relates to a police incident data processing method based on semantic weighted federated learning. Background Technology

[0002] Currently, intelligent alarm receiving systems for fire alarm calls mainly adopt centralized processing architecture or simple edge computing mode. Both types of architecture have exposed significant shortcomings in practical applications, specifically in the systemic deficiencies in three core dimensions: data privacy protection, model generalization ability, and information integrity. Existing technologies are unable to meet the practical needs of modern fire rescue.

[0003] Centralized cloud processing mode: The core logic of this mode is to upload all alarm voice data completely to a remote central server, where the cloud centrally handles the entire process, including voice transcription, feature extraction, and alarm identification. This traditional architecture suffers from three irreconcilable flaws in both its technical principles and practical deployment:

[0004] The risk of privacy breaches is significant: alarm voice messages contain sensitive personal and public safety information, including the caller's identity, precise location coordinates, and details of the disaster. This constitutes core sensitive data as explicitly defined by relevant laws. In a centralized architecture, data undergoes multiple stages of transmission, cloud storage, and cloud processing. Each stage carries the risk of data theft, alteration, or unauthorized access. Existing cases demonstrate that leaks of such centrally stored voice data could lead to serious security risks, such as malicious actors accurately locating target areas and predicting rescue routes, directly violating relevant data security regulations in the fire protection field.

[0005] The system is highly dependent on network connectivity and lacks real-time capability: the uploading of voice data and the return of processing results depend entirely on network bandwidth and transmission latency. According to fire command requirements, key alarm elements must be extracted within 5 seconds of receiving a fire alarm. However, in centralized architectures with network latency exceeding 100ms or bandwidth below 1Mbps, the voice data upload process alone takes 3-8 seconds, resulting in end-to-end processing times generally exceeding 10 seconds, far exceeding the response threshold required in real-world scenarios. In remote areas, disaster sites, and other scenarios with unstable networks, data transmission interruptions may even occur, directly impacting the efficiency of issuing rescue commands.

[0006] The model's generalization ability is severely limited: Significant differences exist in dialects, accents, and noise environments in alarm scenarios across different regions, making it difficult for a single cloud-based model to cover all scenarios. Related research data indicates that centralized automatic speech recognition (ASR) models trained on general corpora experience an accuracy drop of over 30% when faced with dialectal scenarios, and their recognition rate for rare fire types such as forest fires and chemical industrial park fires is less than 40%. This is because the training data for centralized models cannot encompass the specific scenarios of different regions, resulting in weak adaptability to marginal cases and an inability to accurately identify low-frequency, high-risk special alarms.

[0007] To address data privacy issues in centralized architectures, some research has attempted to apply federated learning techniques to fire alarm voice processing scenarios. However, existing solutions directly employ classic federated learning algorithms such as FedAvg, failing to adapt them to the data characteristics of fire alarm scenarios. This leads to severe scenario mismatch issues, with the core shortcomings manifesting in two aspects:

[0008] The lack of value perception leads to model bias: The core assumption of traditional federated learning algorithms is that the data of each node satisfies the independent and identically distributed property, and the contribution of a node is determined only by the amount of data. However, in fire alarm scenarios, the local data of fire brigades in different areas have vastly different semantic values. Fire brigades in chemical industrial parks and areas with dense high-rise buildings have less local data, but they contain the core features of key alarms such as hazardous chemical fires and high-rise building fires. The rescue difficulty coefficient of these types of alarms is more than three times that of ordinary residential fires, and their contribution to improving model performance is extremely high. On the other hand, fire brigades in ordinary residential areas have a large amount of local data, but it is mostly common fire types with low value density. The equal aggregation logic of classic algorithms can lead to the global model being overly biased towards ordinary scenarios with larger data volumes. Related experiments have shown that the federated model using the FedAvg algorithm has a 25% lower accuracy in identifying rare key alarms than the local model, and cannot achieve effective aggregation of key knowledge.

[0009] Unclear evolutionary guidance leads to low training efficiency: Traditional federated learning lacks an evaluation mechanism for the performance shortcomings of each node's model, failing to identify which nodes' models have potential for improvement in specific alarm categories, and also failing to guide the global model to evolve in the direction of these shortcomings. Research shows that in federated training for fire alarm scenarios, traditional algorithms require an additional 30% of training epochs to achieve stable performance, and even after training is completed, the model's adaptability to the shortcomings is still less than 10%. This blind training mode not only consumes a large amount of computing resources and communication bandwidth, but also causes the evolutionary direction of the global model to be out of touch with actual combat needs, making it difficult to specifically improve the ability to handle key alarms.

[0010] Voice enhancement is a crucial preliminary step in fire alarm receiving systems, aiming to improve voice clarity to ensure the effectiveness of subsequent ASR (Automatic Speech Recognition) identification. Existing voice enhancement technologies are mainly divided into traditional methods (such as spectral subtraction and Wiener filtering) and deep learning-based methods (such as Generative Adversarial Network (GAN) models). Both types of methods focus solely on "eliminating all background noise and obtaining a clean human voice," completely neglecting the specific characteristics of fire alarm scenarios and exhibiting two major drawbacks:

[0011] Key background information is completely lost: Background sounds in fire alarm scenarios can be divided into meaningless interference noise (such as wind and ambient human voices) and high-value information sounds (sounds of burning material characteristics, abnormal building structure, and special environmental sounds). These information sounds are crucial for commanders to judge the nature, scale, and danger level of a disaster. Analysis of relevant rescue cases shows that such non-voice information can improve the accuracy of disaster assessment by 28%. However, traditional spectral subtraction directly eliminates all non-human voice frequency band signals through frequency domain threshold filtering, while GAN-type models completely suppress background sounds through generative adversarial training. This results in the enhanced speech retaining only human voices, making it impossible for commanders to obtain key decision-making clues such as the type of burning material and whether the building is about to collapse, thus reducing the comprehensiveness of the alarm assessment.

[0012] Poor scene adaptability and susceptibility to distortion: The noise and key information sounds in fire alarm scenarios exhibit significant specificity in the time and frequency domains. For example, the frequency domain of liquid spray combustion sounds is concentrated in the 500-2000Hz range, highly overlapping with the frequency domain of human voices, while the time domain characteristics of building collapse sounds are short-pulse, high-energy signals. General enhancement models are not optimized for these characteristics. Spectral subtraction can lead to human voice distortion when processing overlapping frequency domain signals, with a distortion rate exceeding 40%. GAN-type models may introduce artifact signals during enhancement, which can be misidentified as valid semantics by subsequent ASR models, resulting in incorrect extraction of alarm elements. Experimental data shows that when general enhancement models are applied to fire alarm scenarios, they reduce ASR recognition accuracy by 22%, negatively impacting the overall system processing performance.

[0013] In summary, existing technologies suffer from irreconcilable shortcomings in three dimensions: privacy protection, data value utilization, and information integrity. Furthermore, there is a lack of end-to-end technical solutions that comprehensively address these three issues. Centralized architectures cannot guarantee privacy and real-time performance; traditional federated learning cannot adapt to the heterogeneous nature of data value; and general-purpose speech enhancement technologies can lose crucial decision-making information. The isolated application of these three technologies means that crime analysis cannot simultaneously achieve security, accuracy, and analytical efficiency. Summary of the Invention

[0014] This application provides a method for processing police incident data based on semantically weighted federated learning, which can solve the problem that police incident analysis cannot simultaneously take into account security, accuracy and analysis efficiency.

[0015] This application provides a method for processing police incident data based on semantically weighted federated learning, including:

[0016] The edge nodes participating in federated learning perform augmentation processing on each fire alarm voice data locally on the edge node to obtain the local training sample set of the edge node, and obtain the local semantic abundance coefficient and contribution coefficient of the local training sample set; the local semantic abundance coefficient is used to characterize the value of the local training sample set, and the contribution coefficient is used to characterize the value of the local training sample set to the global model.

[0017] Edge nodes participating in federated learning will upload the obtained local semantic abundance coefficients and contribution coefficients to the cloud;

[0018] The cloud determines the aggregation weight of each edge node participating in federated learning based on the local semantic abundance coefficient and contribution coefficient uploaded by each edge node. Based on the aggregation weight of all edge nodes participating in federated learning, the cloud performs global aggregation on the local model parameters uploaded by all edge nodes to obtain the initial global model parameters, which are then used as student models. All models corresponding to the local model parameters are used to analyze and process fire alarm voice data and output alarm categories.

[0019] Based on all the obtained local semantic abundance coefficients and all contribution coefficients, the cloud selects teacher nodes from all edge nodes participating in federated learning, and determines the teacher model based on all teacher nodes.

[0020] The cloud performs knowledge distillation on the student model based on the teacher model to obtain the final global model parameters. The final global model parameters are then distributed to each edge node participating in federated learning so that each edge node can train locally on the received final global model parameters and return to execute the steps of obtaining the local semantic abundance coefficient and contribution coefficient of the local training sample set until the global aggregation termination condition is met.

[0021] Edge nodes participating in federated learning use the model corresponding to the final global model parameters when the global aggregation termination condition is met to analyze and process the fire alarm voice data to be analyzed, and obtain the alarm analysis results used to indicate the alarm category of the fire alarm voice data.

[0022] Optionally, the local fire alarm voice data of each edge node is augmented to obtain a local training sample set for the edge node, including:

[0023] For each fire alarm voice data local to the edge node, the fire alarm voice data is input into the background sound detector for processing to obtain the background sound description vector of the fire alarm voice data. The fire alarm voice data and the background sound description vector are then input into the diffusion model for enhancement processing to obtain enhanced voice data. The background sound description vector is used to describe the background sound of the fire scene corresponding to the fire alarm voice data.

[0024] Each enhanced speech data obtained is used as a local training sample, and the set of all local training samples is used as the local training sample set of the edge nodes.

[0025] Optionally, obtain the local semantic abundance coefficients and contribution coefficients of the local training sample set, including:

[0026] The local semantic abundance coefficient of the local training sample set is calculated using the following formula:

[0027] ;

[0028] in, A unique identifier representing the edge node participating in federated learning. Indicates the first Local semantic abundance coefficients of the local training sample set of each edge node. Indicates the first The urgency weight of different types of police incidents Indicates the first The local training sample set of the nth edge node belongs to the nth edge node. The number of local training samples for similar police incidents;

[0029] The contribution coefficient of the local training sample set is calculated using the following formula:

[0030] ;

[0031] in, Indicates the first The contribution coefficient of the local training sample set of each edge node. This represents the Sigmoid function. Indicates the first The target F1 score for each edge node Indicates the first The local model parameters of the edge nodes in the th... F1 score on the hard set of edge nodes;

[0032] The set of difficult examples is the first A subset of the local training sample set of the nth edge node, for the nth edge node For each local training sample in the local training sample set of the n edge nodes, if the nth edge node... If the confidence level of the local model parameters of the i-th edge node for that local training sample is less than the preset confidence level, then that local training sample is used as the i-th edge node. A sample from a set of difficult cases of edge nodes.

[0033] Optionally, based on the local semantic abundance coefficients and contribution coefficients uploaded by each edge node participating in the federated learning, the aggregation weight of each edge node participating in the federated learning is determined, including:

[0034] The number is calculated using the following formula. Aggregate weight of each edge node :

[0035] ;

[0036] in, This represents the preset data volume weighting balance coefficient. Indicates the first The number of local training samples contained in the local training sample set of each edge node. Indicates the first The number of local training samples contained in the local training sample set of each edge node. Indicates the number of edge nodes. This represents the preset semantic abundance adjustment factor. This represents the preset contribution adjustment factor. Indicates the first Local semantic abundance coefficients of the local training sample set of each edge node. Indicates the first The contribution coefficient of the local training sample set of each edge node.

[0037] Optionally, the local model parameters uploaded by all edge nodes are globally aggregated based on the aggregate weights of all edge nodes participating in federated learning to obtain initial global model parameters, including:

[0038] The local model parameters uploaded by all edge nodes are globally aggregated using the following formula:

[0039] ;

[0040] in, Represents the initial global model parameters. This represents the parameters of the previous global model. This represents the learning rate for global model updates. Indicates the first Local model parameters uploaded by each edge node.

[0041] Optionally, based on all obtained local semantic abundance coefficients and all contribution coefficients, teacher nodes are selected from all edge nodes participating in federated learning, including:

[0042] For each edge node participating in federated learning, if the local semantic abundance coefficient of the local training sample set of the edge node is greater than or equal to a preset semantic abundance threshold, and the contribution coefficient of the local training sample set of the edge node is greater than or equal to a preset contribution threshold, then the edge node is regarded as a teacher node.

[0043] Optionally, the teacher model can be determined based on all teacher nodes, including:

[0044] The teacher model is calculated using the following formula:

[0045] ;

[0046] in, Representing the teacher model, This represents the set consisting of all teacher nodes. Indicates the first The aggregate weight of each edge node.

[0047] Optionally, the loss function for knowledge distillation is:

[0048] ;

[0049] in, This represents the value of the loss function. Denotes the set of difficult examples for all teacher nodes. Indicates the temperature coefficient. Indicates the number of police incident categories. Represents the teacher model for the set medium sample Belonging to the The predictive probability of similar police incidents Represents the student model on the set medium sample Belonging to the The predictive probability of similar police incidents;

[0050] = ;

[0051] = ;

[0052] in, Represents the teacher model for the set medium sample In the The original output logic value for the type of alarm. Represents the teacher model for the set medium sample In the The original output logic value for the type of alarm. Represents a set of student types medium sample In the The original output logic value for the type of alarm. Represents the student model on the set medium sample In the The original output logic value for the type of alarm. .

[0053] The above-mentioned solution in this application has the following beneficial effects:

[0054] In the embodiments of this application, by constructing an end-to-end pipeline of local processing at edge nodes, weighted aggregation in the cloud, and edge model iteration, it is ensured that the original local fire alarm voice data does not leave the local area, thereby achieving a balance between efficient end-to-end processing and privacy security. This ensures data security and analysis efficiency. At the same time, since the fire alarm voice data used for federated learning is enhanced, and the final global model parameters are based on the local semantic abundance coefficients and contribution coefficients of each local training sample set, federated learning can tilt the final global model parameters towards high-value, rare, and critical alarm data. This effectively ensures the accuracy of alarm analysis when using the final global model parameters for alarm analysis.

[0055] Other beneficial effects of this application will be described in detail in the following detailed description section. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 A flowchart of a police incident data processing method based on semantically weighted federated learning provided in an embodiment of this application. Detailed Implementation

[0058] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0059] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0060] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0061] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0062] To address the current challenge of balancing security, accuracy, and efficiency in police incident analysis, this application provides a police incident data processing method based on semantically weighted federated learning. This method constructs an end-to-end pipeline of local processing at edge nodes, weighted aggregation in the cloud, and edge model iteration. This ensures that the original local fire alarm voice data remains within the local machine, achieving a balance between efficient end-to-end processing and privacy security. Furthermore, since the fire alarm voice data used for federated learning is enhanced, and the final global model parameters are based on the local semantic abundance and contribution coefficients of each local training sample set, federated learning allows the final global model parameters to learn towards high-value, rare, and critical police incident data. This effectively ensures the accuracy of police incident analysis when using the final global model parameters.

[0063] The following specific embodiments illustrate the police incident data processing method based on semantic weighted federated learning provided in this application.

[0064] like Figure 1 As shown in the embodiments of this application, the police incident data processing method based on semantic weighted federated learning includes the following steps:

[0065] Step 11: The edge nodes participating in federated learning perform augmentation processing on each fire alarm voice data locally on the edge node to obtain the local training sample set of the edge node, and obtain the local semantic abundance coefficient and contribution coefficient of the local training sample set.

[0066] The aforementioned local semantic abundance coefficient is used to characterize the value of the local training sample set, while the contribution coefficient is used to characterize the value of the local training sample set to the global model.

[0067] It should be noted that each edge node participating in federated learning needs to perform step 11 above. These edge nodes can be edge computing nodes deployed in various fire brigades, specifically, a server.

[0068] In some embodiments of this application, the specific implementation of enhancing the local fire alarm voice data of each edge node in step 11 above to obtain the local training sample set of the edge node includes the following steps 11.1 to 11.2:

[0069] Step 11.1: For each fire alarm voice data local to the edge node, the fire alarm voice data is input into the background sound detector for processing to obtain the background sound description vector of the fire alarm voice data. The fire alarm voice data and the background sound description vector are then input into the diffusion model for enhancement processing to obtain enhanced voice data. The enhanced voice data is clean voice data that retains key background sounds.

[0070] The aforementioned background sound description vector is used to describe the background sound of the fire alarm scene corresponding to the fire alarm voice data, such as the sound of liquid spraying, the sound of glass breaking, etc.

[0071] The aforementioned fire alarm voice data consists of voice calls made by alarm callers to the fire department. The aforementioned background noise detector is a pre-trained multi-label classification model, which comprises a lightweight convolutional neural network (CNN) and a fully connected layer (FC) connected sequentially. It employs lightweight convolutional networks such as MobileNetV3-Small (a lightweight CNN) or ShuffleNetV2 (a lightweight CNN), and through operations such as depthwise separable convolution and channel shuffling, it efficiently extracts acoustic features (such as frequency, temporal impulses, and voiceprint features) from the speech while significantly reducing the number of parameters, adapting to the specific time-frequency features of fire alarm background noise. The fully connected layer contains 12 independent output heads customized for background noise in fire alarm scenarios (such as liquid spraying sounds and glass breaking sounds). After processing with the sigmoid activation function, when the probability value of a certain category exceeds a threshold (e.g., 0.7), it is marked as present, ultimately generating a 12-dimensional binary conditional vector. (That is, the background sound description vector, to identify the background sound present in the fire alarm scene). This vector will serve as a "navigator" for the subsequent denoising process.

[0072] The processing procedure for the diffusion model is as follows: steps 11.11 to 11.13:

[0073] Step 11.11, Forward Diffusion Noise Addition: Simulates the process of speech signal degradation during transmission. This involves adding noise to the original clear speech signal. (i.e., fire alarm voice data) Gaussian noise is gradually added at each time step. Generate noisy speech Its mathematical expression follows the Markov chain property:

[0074] ;

[0075] in, Standard Gaussian noise, The accumulated product of preset noise scheduling parameters controls the intensity of noise injection. The expression represents a mean of 0 and a covariance matrix of... The multidimensional (or isotropic) Gaussian distribution is a generalization of the standard normal distribution in higher-dimensional spaces. Represents the identity matrix. Indicates the first The noise scheduling parameter for each step is a preset scalar between 0 and 1, used to control the noise level at each step. The intensity of injected noise.

[0076] Step 11.12, Condition-Guided Reverse Denoising: Reverse denoising is the core process of the diffusion model in restoring noisy speech to clear speech. This application has made a customized modification for fire alarm scenarios, with condition guidance. The input is not just noisy speech, but also a background noise description vector. The two work together to constrain the behavior of the denoising network, preventing key background noise from being mistakenly removed.

[0077] Noisy speech After forward diffusion, time step The fire alarm voice signal generated below, which contains Gaussian noise, is the basic processing object of the denoising network.

[0078] Condition vector The generated 12-dimensional binary background sound label vector (e.g., 1 for "gas leak, explosion" and 0 for no relevant features) is the core basis for the model to identify and protect key background sounds.

[0079] Understandably, in traditional diffusion models, when the input is noisy speech... At that time, time step coding is also required. The time-step features of the diffusion process (encoded into a high-dimensional vector through embedding) allow the model to perceive the current injection intensity of noise and adapt to denoising strategies for different noise levels.

[0080] Noisy speech Conditional vector and time step encoding The signal is fed into the denoising network at the same time as the input signal. In some embodiments of this application, the diffusion model can be a U-Net network fused with cross-attention. U-Net is an end-to-end image segmentation model based on a convolutional neural network (CNN).

[0081] The network internally utilizes a cross-attention mechanism to compute conditional vectors. The association weights with speech features force the model to predict noise. At that time, consciously protect and The corresponding spectral components are not mistakenly removed.

[0082] To balance denoising effectiveness with key information preservation, a multi-task joint loss function is used during training. :

[0083] ;

[0084] in, The mean squared error reconstruction loss is used to quantify and constrain the thoroughness of the model's denoising, ensuring enhanced speech clarity. The core customized loss term in this application is to quantify the retention effect of key background noise in the constrained model and avoid the accidental removal of key background noise. Preserve loss weights for background noise; To compensate for ASR perceptual loss and ensure that the enhanced speech is compatible with the ASR model, Weights for ASR perception loss.

[0085] Step 11.13, Accelerate Sampling Output: Utilize the Denoising Diffusion Implicit Model (DDIM) strategy to accelerate inference, compressing the sampling steps to 50. Noise is progressively removed through iterative formulas to reconstruct clean, enhanced speech that retains key background noise. (i.e., enhanced voice data).

[0086] Step 11.2: Each enhanced speech data obtained is used as a local training sample, and the set of all local training samples is used as the local training sample set of the edge node.

[0087] For each edge node, it trains its local model (the model corresponding to the local model parameters, which is a model delivered from the cloud; in the first iteration of federated learning, the cloud delivers an initialized model, which can be any speech recognition model, such as a Transformer model) using the local training sample set. In practice, the edge node loads the base ASR model delivered from the cloud, freezes the Transformer backbone parameters, and fine-tunes the adapter layer only for local dialects or terms. During inference, key-value (KV) caching and reuse techniques are used to store historical states. To achieve incremental decoding, the core expression for calculating the hidden state at the current moment in the streaming incremental decoding stage is:

[0088] ;

[0089] in, Indicates the first The Transformer encoder at each time step outputs a hidden state vector. Indicates Transformer encoder, Indicates the first Feature vectors of micro-batch speech segments enhanced by conditional diffusion at each time step. Indicates as of Historical context key-value cache features at each time step. Ensure that the transcription latency is controlled within 0.8 seconds to meet real-time requirements.

[0090] It is understandable that the alarm categories of the local training samples in the local training sample set are known, and the alarm category is the label of the corresponding local training sample during training. After training, the local model can be used to analyze and process the local training samples in the local training sample set to determine the hard case set of the edge nodes. This hard case set consists of local training samples that the local model in the local training sample set cannot recognize. These are samples whose recognition results do not reach the preset quantization threshold after processing by the edge-side full-link model, and which have high practical value in fire fighting.

[0091] In some embodiments of this application, the specific implementation of obtaining the local semantic abundance coefficient and contribution coefficient of the local training sample set in step 11 above includes the following steps 11.3 to 11.4:

[0092] Step 11.3: Calculate the local semantic abundance coefficient of the local training sample set using the following formula:

[0093] ;

[0094] in, A unique identifier representing the edge node participating in federated learning. Indicates the first Local semantic abundance coefficients of the local training sample set of each edge node. Indicates the first The urgency weight of different types of police incidents Indicates the first The local training sample set of the nth edge node belongs to the nth edge node. The number of local training samples for similar police incidents. Among them, the urgency weight... The weighting can be preset based on experience, such as giving "explosion" a higher weight than "weed fire". The local semantic abundance coefficient is logarithmically transformed with respect to the sample size, weakening the absolute difference in sample size and highlighting the value of rare cases, thus preventing a dominant case from being overly represented by a large sample size. The calculation results are obtained through urgency weighting. By assigning higher value weight to high-risk incidents and using logarithmic mapping, we can prevent ordinary incidents from being overshadowed by excessively large sample sizes, thus ensuring that "even a small number of high-risk incidents can contribute significantly to the overall value assessment." value".

[0095] In some embodiments of this application, the number of alarm categories is 12. These 12 alarm categories may include explosions, weed fires, residential fires, shop fires, factory and warehouse fires, vehicle fires, electrical wiring fires, kitchen oil fires, garbage and debris fires, construction site fires, forest fires, and fires caused by gas leaks.

[0096] The aforementioned local semantic abundance coefficient is used to measure the density of key information contained in the data: it quantifies the key information density of 12 core fire alarms in the local data of edge nodes. The larger the value, the higher the semantic value of the local training sample set of the edge node, and the stronger the contribution of high-value data to the global model learning of rare and key alarm features.

[0097] Step 11.4: Calculate the contribution coefficient of the local training sample set using the following formula:

[0098] ;

[0099] in, Indicates the first The contribution coefficient of the local training sample set of each edge node. This represents the Sigmoid function. Indicates the first The target F1 score for each edge node It can be pre-set according to actual combat needs. Indicates the first The local model parameters of the edge nodes in the th... The F1 score on the set of hard examples of edge nodes, i.e. The local model parameters corresponding to the edge nodes (the local model trained in the current round for each edge node) in the th... The F1 score on the hard set of edge nodes, and Subtraction measures the global model in the first step. The performance improvement potential on the difficult example set of edge nodes is the core source of the data's potential value. The F1 score is a commonly used comprehensive evaluation metric for models, and it can be calculated using its conventional method. The calculation process of the F1 score will not be elaborated here.

[0100] The above set of difficult examples is the first. A subset of the local training sample set of the nth edge node. For the nth edge node... For each local training sample in the local training sample set of the n edge nodes, if the nth edge node... If the confidence level of the local model parameters (i.e., the local model after the current training round of the edge node) for the local training sample is less than the preset confidence level (e.g., 0.6), then the local training sample is used as the first edge node. A sample is a set of difficult examples from a set of edge nodes. In practical applications, samples may also be required to contain high-value fire safety features such as rare information, complex noise, key background sounds, and dialects / technical terms before being defined as local difficult examples. Specifically, based on the model's identification of difficult samples, automated judgment can be performed by matching them to a pre-set high-value fire alarm feature library to select only samples with core value for actual fire rescue operations.

[0101] The contribution coefficients mentioned above are used to measure the potential contribution of local data to improving the performance of the global model. They are used to quantify the... The local data of each edge node has the potential contribution to the improvement of the global model performance. The larger the value, the more effectively the edge node data can make up for the performance shortcomings of the global model. It is a quantitative indicator of the potential value of the data. The core logic is that the worse the model performs on the data of this node, the higher the potential learning value of the data.

[0102] The formula for calculating the contribution coefficient implies that the worse the model performs (i.e., encountering unknown or difficult samples), the higher the data value of that node, and the more it needs to be learned. The difficult example set refers to the set of... The local hard case set of an edge node refers to the set of fire alarm voice samples in the local security domain of that node that are difficult for the global model / local model to identify accurately, but have high value for fire rescue operations. These samples usually contain features such as rare situation features, complex noise environment, dense dialects or professional terms, and ambiguous situation features. They are the main shortcomings of the model performance and are also the samples that the global model needs to focus on learning.

[0103] The two-dimensional indicators (local semantic abundance coefficient and contribution coefficient) work together to form the semantic value tuple of edge node data. ),in Reflecting the value of the data itself (whether it contains high-risk or rare informational characteristics); This reflects the value of data to the model (whether it can compensate for the model's performance shortcomings). The two form a two-dimensional evaluation system of "existing value + potential value", which realizes the accurate quantification of the value of local data, provides accurate and comprehensive value basis for cloud semantic weighted aggregation, ensures that the federated learning process can selectively absorb high-value knowledge, and improve the global model's ability to handle key alarms.

[0104] Step 12: The edge nodes participating in federated learning upload the obtained local semantic abundance coefficients and contribution coefficients to the cloud.

[0105] The aforementioned cloud can be a server deployed in the fire management department, serving as a cloud device for federated learning. After each edge node completes local training and calculates its local semantic abundance coefficient and contribution coefficient, it uploads the trained local model (i.e., local model parameters), local semantic abundance coefficient, and contribution coefficient to the cloud for aggregation.

[0106] Step 13: Based on the local semantic abundance coefficient and contribution coefficient uploaded by each edge node participating in federated learning, the cloud determines the aggregation weight of each edge node participating in federated learning, and performs global aggregation on the local model parameters uploaded by all edge nodes based on the aggregation weight of all edge nodes participating in federated learning to obtain the initial global model parameters, and uses the initial global model parameters as the student model.

[0107] All the models corresponding to the above local model parameters are used to analyze and process fire alarm voice data and output alarm categories.

[0108] In some embodiments of this application, the first step can be calculated using the following formula. Aggregate weight of each edge node :

[0109] ;

[0110] in, This represents the preset data volume weighting balance coefficient, used to adjust the proportion of data volume in the weighting distribution. The semantic value weighting, as the core and dominant component of weight allocation, highlights the crucial role of high-value data in the overall model evolution. Indicates the first The number of local training samples contained in the local training sample set of each edge node. Indicates the first The number of local training samples contained in the local training sample set of each edge node. Indicates the number of edge nodes. This represents the preset semantic abundance adjustment factor. This represents the preset contribution adjustment factor. and All settings can be pre-defined according to the actual situation. Indicates the first Local semantic abundance coefficients of the local training sample set of each edge node. Indicates the first The contribution coefficient of the local training sample set of each edge node.

[0111] Understandably, the cloud will iterate through all edge nodes and execute the above formula for calculating aggregate weights to obtain the aggregate weights of all edge nodes.

[0112] In some embodiments of this application, the local model parameters uploaded by all edge nodes can be globally aggregated using the following formula:

[0113] ;

[0114] in, Represents the initial global model parameters. This represents the parameters of the previous global model. This represents the learning rate for global model updates, used to control the step size for updating model parameters. Indicates the first The local model parameters uploaded by each edge node are the local model parameters trained on the edge nodes. The core advantage of this aggregation process is that it abandons the simple logic of "average aggregation" in traditional federated learning. Through differentiated weight allocation, it allows the global model to prioritize learning the key knowledge of high-value nodes, significantly improving the model's ability to handle rare and high-risk situations, while also taking into account the basic data contributions of ordinary nodes to ensure the model's generalization ability.

[0115] The above aggregation process, through differentiated weight allocation, enables the gradient update amount of high-value nodes to dominate in the global model, thereby achieving the initial directional optimization of the global model.

[0116] It should be noted that federated learning involves multiple rounds of training, each consisting of local training at edge nodes and global aggregation in the cloud. During global aggregation in the cloud, the process follows the aforementioned steps. Aggregate the formulas. This can be understood as the global model parameters obtained from the current round of global aggregation (the initial global model parameters are referred to here because subsequent knowledge distillation is required to obtain the final global model parameters in this application). These are the parameters of the previous round of global model. If the current aggregation is the first round of aggregation, then... Initialize the model parameters for the cloud model (i.e., initialize the model parameters for the Transformer model).

[0117] Step 14: Based on all the obtained local semantic abundance coefficients and all contribution coefficients, the cloud selects teacher nodes from all edge nodes participating in federated learning, and determines the teacher model based on all teacher nodes.

[0118] In some embodiments of this application, for each edge node participating in federated learning, if the local semantic abundance coefficient of the local training sample set of the edge node is greater than or equal to a preset semantic abundance threshold (e.g., 0.7) and the contribution coefficient of the local training sample set of the edge node is greater than or equal to a preset contribution threshold (e.g., 0.6), then the edge node is regarded as a teacher node.

[0119] In some embodiments of this application, the teacher model can be calculated using the following formula:

[0120] ;

[0121] in, Representing the teacher model, This represents the set consisting of all teacher nodes. Indicates the first The aggregated weights of the edge nodes. In the above formula, the denominator term is used to renormalize the weights of the selected teacher nodes to ensure that the aggregated teacher model... It can centrally reflect the judgment logic of high-value nodes.

[0122] Step 15: The cloud performs knowledge distillation on the student model based on the teacher model to obtain the final global model parameters. The final global model parameters are then distributed to each edge node participating in federated learning so that each edge node can train locally on the received final global model parameters and return to the step of obtaining the local semantic abundance coefficient and contribution coefficient of the local training sample set until the global aggregation termination condition is met.

[0123] In some embodiments of this application, to further solidify the representation of high-value, rare informational knowledge in the global model and solve the knowledge transfer problem between heterogeneous models (edge ​​personalized models and cloud general models), the cloud initiates a knowledge distillation fine-tuning process after completing gradient-weighted aggregation. Specifically, knowledge distillation can be performed using the hard example set of all teacher nodes. During the knowledge distillation process, the soft labels of the teacher models are used as the supervision target, and the Kohlbek-Leibler (KL) divergence is used as the core distillation loss function to calculate and minimize the distribution difference between the student model and the teacher model. Specifically, the loss function for knowledge distillation is:

[0124] ;

[0125] in, This represents the value of the loss function. Denotes the set of difficult examples for all teacher nodes. This represents the temperature coefficient (a pre-set parameter in knowledge distillation), used to soften the probability distribution and amplify the values ​​of non-maximum probability categories (i.e., "dark knowledge"). Indicates the number of police incident categories. , Represents the teacher model for the set medium sample Belonging to the The predictive probability of similar police incidents Represents the student model on the set medium sample Belonging to the The predictive probability of similar police incidents;

[0126] = ;

[0127] = ;

[0128] in, Represents the teacher model for the set medium sample In the The original output logical values ​​(Logits) for similar alarm situations. Represents the teacher model for the set medium sample In the The original output logic value for the type of alarm. Represents a set of student types medium sample In the The original output logic value for the type of alarm. Represents the student model on the set medium sample In the The original output logic value for the type of alarm. It should be noted that the collection will be iterated over. Each specific sample in The data is input into both the teacher model and the student model for inference. Although their internal structures may differ, their output spaces (12 types of alarm elements) are completely identical. This application employs a response-based distillation strategy, generating soft tags through temperature smoothing. and .

[0129] It should be further explained that the above logical value is the value directly output by the model after processing the data. In the embodiments of this application, the model parameters (such as local model parameters, global model parameters, etc.) can be regarded as the model corresponding to the model parameters.

[0130] It is worth mentioning that by minimizing This forces the output logic curve of the student model to approach the teacher model infinitely, thereby enabling the lossless transfer of implicit experience from high-value nodes to the global model and completing a deep evolution from "parameter averaging" to "logic assimilation".

[0131] It's important to note that federated learning typically requires multiple rounds of iterative training. Each round of training involves: the model being distributed from the cloud; edge nodes using their local training sample sets to train the received model locally; and after training, calculating the local semantic abundance coefficient and contribution coefficient for the current round (it's understood that the local semantic abundance coefficient is the same across all rounds, so it only needs to be calculated once and used directly in subsequent rounds). The edge nodes then upload these coefficients, along with their local model parameters, to the cloud. The cloud calculates the aggregation weight for each edge node based on the uploaded local semantic abundance coefficient and contribution coefficient, and performs global aggregation and knowledge distillation based on these weights to obtain the final global model parameters. After obtaining the final global model parameters in each round, the cloud distributes them to each edge node to begin the next iteration, until the global aggregation termination condition is met. For example, this global aggregation termination condition could be that the number of iterations (i.e., the number of global aggregations) reaches a preset number (e.g., 200).

[0132] It is worth mentioning that the aforementioned federated learning process is the core hub of the entire system's global collaborative evolution, undertaking the key responsibilities of cross-regional knowledge aggregation, global model optimization, and knowledge base updates. Its core design principle is "knowledge sharing without data sharing." By processing encrypted information such as model parameters and value metadata uploaded by edge nodes, it achieves global intelligent collaborative evolution without touching any edge node's raw voice data, fundamentally ensuring data privacy and security.

[0133] Step 16: The edge nodes participating in federated learning use the model corresponding to the final global model parameters when the global aggregation termination condition is met to analyze and process the fire alarm voice data to be analyzed, and obtain the alarm analysis result used to indicate the alarm category of the fire alarm voice data.

[0134] The fire alarm voice data to be analyzed is the voice call data when the person making the alarm dials the fire brigade's emergency call number. In the actual analysis and processing, the fire alarm voice data can first be processed according to the data augmentation processing method in step 11. Then, the augmented fire alarm voice data is input into the model corresponding to the final global model parameters (the final global model parameters when the global aggregation termination condition is met) for analysis and processing, and the output is the alarm analysis result indicating the alarm category of the fire alarm voice data. This alarm category can be one or more of the aforementioned 12 alarm categories.

[0135] Understandably, after receiving the police incident analysis results, relevant personnel will output the results so that they can promptly arrange for the appropriate police response.

[0136] It is worth mentioning that the police data processing method based on semantic weighted federated learning in this application ensures data security and analysis efficiency by constructing an end-to-end pipeline of local processing at edge nodes, weighted aggregation in the cloud, and edge model iteration. At the same time, since the fire alarm voice data used for federated learning is enhanced, and the final global model parameters are based on the local semantic abundance coefficients and contribution coefficients of each local training sample set, federated learning can tilt the final global model parameters towards high-value, rare, and key police data, thereby effectively ensuring the accuracy of police analysis when using the final global model parameters for police analysis.

[0137] In some embodiments of this application, to address potential issues such as identification errors and missing information from a single data source, a multi-source logical verification process is first performed on the structured data reported by each edge node. This multi-level verification mechanism ensures data reliability and includes three core operations:

[0138] Precise timestamp alignment: For scenarios involving cross-regional collaborative alarm reception or parallel processing by multiple edge nodes, the timestamps of structured data packets reported by different nodes are calibrated based on the alarm call access time, with calibration accuracy down to the millisecond level, ensuring that all data corresponds to the same alarm event and avoiding cross-event data confusion.

[0139] Multidimensional logical consistency verification: Construct a three-dimensional logical verification system of "element-tag-scene" to rigorously verify the rationality of the relationship between data.

[0140] From the perspective of element association, the consistency between address information and geospatial topology is verified. For example, when labeling a chemical industrial park, the corresponding regional coordinate range must be matched. From the perspective of element-tag matching, the logical association between disaster type and key background sound tags is verified. For example, a chemical plant fire must have a reasonable matching relationship with tags such as gas leak sounds and equipment alarm sounds. From the perspective of scenario-level matching, the adaptability between disaster type and hazard level is verified. For example, explosion-type disasters must correspond to high-hazard levels. The logical association degree is quantitatively calculated using a preset algorithm. If the association degree is lower than 0.7, it is determined that there is a logical conflict in the data.

[0141] Dynamic confidence level calibration: Based on the logical verification results, the initial confidence level of each structured data item is dynamically adjusted. For logically consistent data items, the confidence level is maintained or slightly increased; for data items with logical conflicts or missing information, the confidence level is reduced according to the severity of the conflict. For example, for disaster types with completely unrelated labels, the confidence level is reduced by more than 0.3; for orphan data reported by a single node, the confidence level is reduced by 0.1-0.2 compared to data cross-validated by multiple nodes, ensuring that the confidence level can truly reflect the reliability of the data.

[0142] To balance the efficiency of automated processing with the accuracy of decision-making in complex scenarios, the module is designed with a human-machine collaboration mechanism that prioritizes automated machine decision-making and supplements it with human intervention. Differentiated processing flows are triggered by confidence thresholds.

[0143] Confidence threshold setting: Based on real-world test data and the fault tolerance requirements for rescue decisions, a confidence threshold is preset for core data items, and the confidence threshold is not lower than 0.8. This threshold can be dynamically adjusted according to the rescue needs and data quality of different regions. Core data items include key information that directly affects rescue dispatch, such as address, disaster type, hazard level, and key background sound tags.

[0144] Automated decision-making process: For structured data packets where the confidence level of all core data items is higher than the threshold and there are no conflicts after logical verification, the system automatically initiates a standardized processing procedure without manual intervention. Through preset field mapping rules, the data is transformed into a format that meets the interface requirements of the rescue dispatch system, quickly generating decision data packets.

[0145] Manual intervention process: When the confidence level of a core data item falls below a threshold, logical verification conflicts cannot be automatically resolved, or ambiguous information is detected (e.g., multiple conflicting addresses for the same event), the system automatically triggers the manual intervention interface. The problematic data segment, including key transcribed text fragments, conflict tags, and confidence scores, is pushed to the dispatcher's workbench in real time, along with the conflict type and suspected problem points, to help the dispatcher quickly locate key conflicts. The dispatcher listens back to the audio segment (stored locally without original data transmission) or confirms information through supplementary questioning, then manually corrects the data and feeds it back to the system. The system updates the data and confidence level based on the manual corrections, completing the closed-loop optimization.

[0146] Standardized decision data packages are generated according to unified industry data protocols and interface specifications to ensure that the data can be directly parsed and accessed by the rescue and dispatch system. The specific output content and characteristics are as follows:

[0147] Standardized data packet structure: Data packets adopt JSON format (JSON is an open standard file and data exchange format), containing four core field sets: First, the transcribed text field set, covering the complete transcribed text of the alarm voice and highlighting key information, such as address and disaster keywords; second, the structured core element field set, containing 12 core elements such as address, disaster type, danger level, and estimated number of trapped people, with field definitions conforming to relevant specifications; third, the background sound label field set, containing the presence status of 12 key background sounds, marked with 0-1 binary numbers, and corresponding detection probabilities; fourth, the confidence level description field set, containing the final confidence level score for each core data item, data source node identifier, and whether it has undergone manual correction, etc., to facilitate dispatchers in judging data reliability.

[0148] Data Interaction Adaptability: Data packets strictly adhere to the RESTful API interface protocol, supporting seamless integration with mainstream fire and rescue dispatch systems. Data transmission utilizes encrypted channels to ensure security during transmission. Furthermore, dynamic adaptation of data formats is supported; field mapping relationships can be adjusted via configuration files according to the interface requirements of different dispatch systems without modifying the core code.

[0149] Decision support across all dimensions: The output standardized data package not only provides text and structured element information, but also supplements non-voice decision-making clues through background sound tags, forming a multi-dimensional intelligence support of "text + elements + tags", which helps dispatchers quickly and accurately assess the disaster situation, formulate scientific and reasonable rescue plans, and achieve efficient transformation from technical data to practical decision-making.

[0150] The following examples illustrate the alarm data processing method of this application.

[0151] In this example, a typical alarm handling process for "gas leak in a chemical industrial park causing deflagration" is simulated. Edge nodes are deployed at various fire brigades and forward duty points in key chemical industrial parks, while cloud deployment is carried out at relevant management offices.

[0152] Phase 1, Model Initialization and Warm-up:

[0153] Global initialization: The cloud uses publicly available datasets and historical de-identified data to pre-train the basic diffusion model and ASR model.

[0154] Edge deployment: The basic model parameters are compressed and deployed to the edge nodes of the chemical industrial park branch. The edge nodes perform lightweight fine-tuning of the ASR's Adapter layer based on local historical data (specific chemical terminology, factory environmental noise).

[0155] Phase Two: Real-time Alarm Reception and Edge Processing

[0156] Scenario input: An emergency call is received, the caller is agitated, and the background is accompanied by the hissing sound of high-pressure gas jets and the muffled sound of an explosion in the distance.

[0157] Background noise detection and condition generation: Edge nodes divide the audio stream into 500ms micro-batches.

[0158] The background noise detector performs real-time analysis and outputs a probability vector: [liquid jet: 0.1, gas leak: 0.92, explosion sound: 0.85, ...]. Among them, a judgment probability > 0.7 is considered valid. A binary condition vector is generated, and "gas leak" and "explosion" are marked as retained features.

[0159] Conditional diffusion speech enhancement: Input noisy speech and conditional vectors into the conditional diffusion model.

[0160] During the reverse denoising process, the model uses a cross-attention mechanism to remove the rapid breathing sound of the alarm caller and the background noise of the current, while forcibly retaining the hissing sound and explosion echo in the background.

[0161] The output speech is enhanced, with the signal-to-noise ratio increased from 8dB to 26dB.

[0162] Streaming transcription and feature extraction: Enhanced speech input into the ASR model, using KV caching and reuse technology, transcribed text in 0.8 seconds: "The pipeline in the No. 2 tank area of ​​XX Chemical Plant exploded, and there is a gas leak."

[0163] Local value assessment: Calculating local semantic abundance coefficients Calculate the contribution coefficient Generate semantic value tuples .

[0164] Phase Three: Federated Collaboration and Model Evolution

[0165] Encrypted upload: During idle periods (or when the accumulation threshold is reached), edge nodes homomorphically encrypt the local model parameters generated by this inference and upload them to the cloud along with the semantic value tuple.

[0166] Semantic weighted aggregation: The cloud receives updates from a node in the chemical industrial park (small data volume, but high value) and a node in a residential area (large data volume, low value).

[0167] Calculate weights according to the formula and perform global aggregation: Chemical Industrial Park node due to and The weighting coefficient is extremely high, with a weighting coefficient of 0.6, while the weight of the residential area node is 0.4.

[0168] Knowledge distillation and distribution: The cloud uses soft tags from high-value samples to fine-tune the global model. After updating the knowledge base, the more powerful new model is distributed to all teams. The models of other teams (such as neighboring areas) are thus able to identify "chemical pipeline leaks".

[0169] Phase Four: Command Center Integrated Decision-Making

[0170] Data reception and verification: The command center receives the data packets uploaded from the edge.

[0171] Text: The pipeline in the No. 2 tank area of ​​XX Chemical Plant exploded.

[0172] Background sound tags: [Gas leak (0.92), Explosion (0.85)];

[0173] System verification: The text mentions "gas leak," and the tag contains "gas leak sound." Logical match score = 1.0.

[0174] Address verification: The address "XX Chemical Plant" exists in the map database and is located within the jurisdiction of this detachment.

[0175] Overall confidence level: 0.95 (> threshold 0.8).

[0176] Standardized output: Within 4.2 seconds of receiving an alarm, the system automatically generates a JSON command and sends it to the dispatch system.

[0177] Instructions: Deploy a chemical convoy, equipped with chemical protective suits and leak-stopping equipment (based on the "leak" label), and prepare foam trucks (based on the "explosion" text).

[0178] The key parameter settings involved in the above examples are shown in Table 1.

[0179] Table 1 Key Parameter Settings

[0180]

[0181] Based on the above example, compared with the traditional FedAvg, the recognition accuracy of this application is improved by more than 15%. Furthermore, in this example, the method of this application receives edge-de-identified data in 0 seconds, completes timestamp alignment in 2.0 seconds, performs "text-tag-scene" logical verification in 2.5 seconds, dynamically calibrates confidence (threshold ≥ 0.8) in 3.0 seconds, automatically processes data that meets the criteria and triggers manual intervention if the criteria are not met in 3.5 seconds, generates standardized JSON containing 12 elements and 12 background sound tags in 4.2 seconds, and pushes it to the rescue dispatch system 5 seconds in advance, ensuring accurate and efficient decision-making.

[0182] In summary, the police data processing method provided in this application takes the federated learning paradigm of "data remains stationary while the model moves" as its core, achieving a deep unification of data privacy protection and continuous model evolution.

[0183] On the edge side, a customized conditional denoising diffusion probability model is introduced. A lightweight detector captures twelve key background sounds in real time, such as liquid jet combustion sound, building collapse precursor sound, and high-pressure gas leakage sound, and generates multi-label guiding conditions. The speech enhancement task is reconstructed into a dual-objective optimization problem of suppressing interference noise and retaining decision cues. While reconstructing clear human voices to ensure ASR recognition effect, it accurately retains environmental acoustic cues that are crucial to judging the nature, scale and danger level of the disaster.

[0184] In the cloud, breaking through the limitations of traditional federated learning's equal-weighted aggregation based on data volume, a semantic value-oriented dynamic aggregation algorithm is designed. By quantifying the semantic abundance of data at each edge node, i.e., the coverage of key disaster categories, and the model's evolutionary potential, i.e., the potential for improvement in weak category, high-value nodes are given higher aggregation weights. This ensures that rare but high-risk disaster characteristics such as chemical fires and high-rise building fires can be quickly captured and absorbed by the global model, achieving targeted optimization and efficient evolution of the global model.

[0185] At the command center, the de-identified structured data uploaded from edge nodes, including transcribed text, core alarm elements, key background sound tags, and confidence levels, is aligned with timestamps, checked for logical consistency, and processed for multi-source ambiguity. This process integrates the data to form high-confidence, multi-dimensional decision support information, providing rescue dispatch with a multi-dimensional basis for judgment that goes beyond a single text. This completely solves the practical pain points of traditional systems, such as privacy insecurity, incomplete information, inaccurate models, and untimely response.

[0186] Overall, the alarm data processing method provided in this application achieves the following four core objectives:

[0187] First, it breaks through the limitations of traditional federated learning: It innovatively defines and quantifies the semantic value contribution of node data, and through dynamic weighted aggregation, it allows the global model to learn towards high-value, rare, and critical police data, thus solving the scenario adaptation problem of "data volume ≠ value volume".

[0188] Second, it reconstructs the paradigm of speech enhancement technology: abandoning the single goal of "pure human voice", it proposes the concept of "selective reconstruction" and, through the conditional diffusion model, improves the clarity of human voice while controllably preserving key background sounds of disaster, making up for the information loss shortcomings of traditional technologies.

[0189] Third, create a closed loop for privacy protection: construct an end-to-end pipeline of "edge local processing - cloud weighted aggregation - edge model iteration" to ensure that the original voice data does not leave the local area and achieve the unity of cross-regional knowledge sharing and privacy security.

[0190] IV. Meeting core practical needs: Under complex noise and privacy constraints, it achieves end-to-end processing within 5 seconds, significantly improves the accuracy of key alarm identification, and outputs multi-dimensional decision-making information with background sound tags, adapting to the "fast, accurate, and safe" requirements of fire rescue.

[0191] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principles described in this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for processing police incident data based on semantically weighted federated learning, characterized in that, include: The edge nodes participating in federated learning perform enhancement processing on each fire alarm voice data locally on the edge node to obtain the local training sample set of the edge node, and obtain the local semantic abundance coefficient and contribution coefficient of the local training sample set; the local semantic abundance coefficient is used to characterize the value of the local training sample set, and the contribution coefficient is used to characterize the value of the local training sample set to the global model. Edge nodes participating in federated learning will upload the obtained local semantic abundance coefficients and contribution coefficients to the cloud; The cloud platform determines the aggregation weight of each edge node participating in federated learning based on the local semantic abundance coefficient and contribution coefficient uploaded by each edge node. It then performs global aggregation on the local model parameters uploaded by all edge nodes based on the aggregation weight of all edge nodes participating in federated learning to obtain initial global model parameters, which are used as student models. All models corresponding to the local model parameters are used to analyze and process fire alarm voice data and output alarm categories. The cloud-based system selects teacher nodes from all edge nodes participating in federated learning based on all obtained local semantic abundance coefficients and all contribution coefficients, and determines the teacher model based on all teacher nodes. The cloud performs knowledge distillation on the student model based on the teacher model to obtain the final global model parameters, and then distributes the final global model parameters to each edge node participating in federated learning, so that each edge node can perform local training on the received final global model parameters and return to the step of obtaining the local semantic abundance coefficient and contribution coefficient of the local training sample set until the global aggregation termination condition is met. Edge nodes participating in federated learning use the model corresponding to the final global model parameters when the global aggregation termination condition is met to analyze and process the fire alarm voice data to be analyzed, and obtain the alarm analysis results used to indicate the alarm category of the fire alarm voice data.

2. The alarm data processing method according to claim 1, characterized in that, Enhancement processing is performed on each fire alarm voice data locally at the edge node to obtain the local training sample set of the edge node, including: For each fire alarm voice data locally at the edge node, the fire alarm voice data is input into a background sound detector for processing to obtain a background sound description vector of the fire alarm voice data. The fire alarm voice data and the background sound description vector are then input into a diffusion model for enhancement processing to obtain enhanced voice data. The background sound description vector is used to describe the background sound of the fire scene corresponding to the fire alarm voice data. Each enhanced speech data obtained is used as a local training sample, and the set of all local training samples is used as the local training sample set of the edge node.

3. The alarm data processing method according to claim 2, characterized in that, Obtain the local semantic abundance coefficients and contribution coefficients of the local training sample set, including: The local semantic abundance coefficient of the local training sample set is calculated using the following formula: ； in, A unique identifier representing the edge node participating in federated learning. Indicates the first Local semantic abundance coefficients of the local training sample set of each edge node. Indicates the first The urgency weight of different types of police incidents Indicates the first The local training sample set of the nth edge node belongs to the nth edge node. The number of local training samples for similar police incidents; The contribution coefficient of the local training sample set is calculated using the following formula: ； in, Indicates the first The contribution coefficient of the local training sample set of each edge node. This represents the Sigmoid function. Indicates the first The target F1 score for each edge node Indicates the first The local model parameters of the edge nodes in the th... F1 score on the hard set of edge nodes; The set of difficult examples is the first... A subset of the local training sample set of the nth edge node, for the nth edge node For each local training sample in the local training sample set of the n edge nodes, if the nth edge node... If the confidence level of the local model parameters of the i-th edge node for that local training sample is less than the preset confidence level, then that local training sample is used as the i-th edge node. A sample from a set of difficult cases of edge nodes.

4. The alarm data processing method according to claim 3, characterized in that, Based on the local semantic abundance coefficient and contribution coefficient uploaded by each edge node participating in the federated learning, the aggregation weight of each edge node participating in the federated learning is determined, including: The number is calculated using the following formula. Aggregate weights of edge nodes : ； in, This represents the preset data volume weighting balance coefficient. Indicates the first The number of local training samples contained in the local training sample set of each edge node. Indicates the first The number of local training samples contained in the local training sample set of each edge node. Indicates the number of edge nodes. This represents the preset semantic abundance adjustment factor. This represents the preset contribution adjustment factor. Indicates the first Local semantic abundance coefficients of the local training sample set of each edge node. Indicates the first The contribution coefficient of the local training sample set of each edge node.

5. The alarm data processing method according to claim 4, characterized in that, Based on the aggregated weights of all edge nodes participating in federated learning, the local model parameters uploaded by all edge nodes are globally aggregated to obtain the initial global model parameters, including: The local model parameters uploaded by all edge nodes are globally aggregated using the following formula: ； in, Represents the initial global model parameters. This represents the parameters of the previous global model. This represents the learning rate for global model updates. Indicates the first Local model parameters uploaded by each edge node.

6. The alarm data processing method according to claim 5, characterized in that, Based on all obtained local semantic abundance coefficients and all contribution coefficients, teacher nodes are selected from all edge nodes participating in federated learning, including: For each edge node participating in federated learning, if the local semantic abundance coefficient of the local training sample set of the edge node is greater than or equal to a preset semantic abundance threshold, and the contribution coefficient of the local training sample set of the edge node is greater than or equal to a preset contribution threshold, then the edge node is regarded as a teacher node.

7. The alarm data processing method according to claim 6, characterized in that, The teacher model is determined based on all teacher nodes, including: The teacher model is calculated using the following formula: ； in, Representing the teacher model, This represents the set consisting of all teacher nodes. Indicates the first The aggregate weight of each edge node.

8. The alarm data processing method according to claim 7, characterized in that, The loss function for knowledge distillation is: ； in, This represents the value of the loss function. Denotes the set of difficult examples for all teacher nodes. Indicates the temperature coefficient. Indicates the number of police incident categories. Represents the teacher model for the set medium sample Belonging to the The predictive probability of similar police incidents Represents the student model on the set medium sample Belonging to the The predictive probability of similar police incidents; = ； = ； in, Represents the teacher model for the set medium sample In the The original output logic value for the type of alarm. Represents the teacher model for the set medium sample In the The original output logic value for the type of alarm. Represents a set of student types medium sample In the The original output logic value for the type of alarm. Represents the student model on the set medium sample In the The original output logic value for the type of alarm. .

Images